Snowflake Summit 2023 Keynote Full Breakdown | Snowflake Summit

0:00Hey it's Tim here, in today's video I'm

0:01going to be breaking down the Snowflake

0:03Summit keynote.

0:04We're just going to watch the whole thing,

0:06this is going to be like an honest

0:07reflection, it's quite

0:08long, this video is long as a result of

0:10that. If you want the breakdown of this

0:12video, if you want

0:13something shorter than 20 minutes, go ahead

0:15and check the description and the first

0:16comment on

0:17this video, you'll find a link to my

0:19breakdown of that as soon as it's ready.

0:21Now I have to say I'm

0:22coming at this in lots of different hats, I

0:25've actually been a Snowflake customer

0:27myself

0:27for about three years, I actually use Snow

0:30flake to power the back-end analytics for my

0:32channel,

0:33I pipe the data out of YouTube using the

0:35API, I use 5tran to do that, so YouTube to

0:395tran to

0:40Snowflake is basically the simple pipeline,

0:42then I visualize it in Tableau, sort of all

0:44these

0:44technologies doing amazing things together.

0:47And so what I wanted to do for some time is

0:49start covering

0:49new technologies on this channel, this is

0:51going to be the next thing, I've got so

0:53many interesting

0:54announcements related to Snowflake coming

0:56up over the next few months, so stay tuned

0:58for that,

0:59not just here on YouTube but also on other

1:01places if you know where I mean. So let's

1:04get stuck into

1:04this particular keynote, let's break it

1:06down, I'm going to try and give my honest

1:08perspective on all

1:09of these things, but as ever I'm new to

1:10this, I'm learning it just like you, so if

1:12you've got some

1:13feedback or some thoughts, put them in the

1:15comments below and we'll get stuck into

1:17that. Anyway as

1:17ever, let's get started. Okay. Good morning

1:27everybody. Good morning. Do you guys catch

1:32us

1:32on stage last night with Jensen? Now right

1:35out of the gate that's already a bit of a

1:37flex,

1:38you know Jensen is the NVIDIA CEO and for

1:42the record NVIDIA have been absolutely

1:45knocking it

1:45out of the park for the last let's say

1:47three years. Jensen made this big bet on AI

1:50and he

1:50thought it was going to be the big

1:52revolutionary thing that would sort of

1:54catapult NVIDIA into the

1:55future and as a result pretty much every

1:58single AI breakthrough that has happened in

2:00the last few

2:01years has pretty much happened using NVIDIA

2:03hardware. If you don't know this, when you

2:06do

2:06a query into chatgpt or when you do any

2:08sort of use of large language models, the

2:11computational

2:12power to run those models actually runs on

2:14graphical processes. Graphical processes

2:16have

2:16these specific mathematical computations

2:19that allow them to do that sort of

2:20operation very

2:21efficiently and so that's why GPUs are

2:23typically used to train large language

2:25models but then also

2:26when you run these queries and when you

2:29push all this sort of information into

2:30these models, GPUs

2:32tend to be able to process these things

2:34much faster and so NVIDIA is pretty much

2:36the only company

2:36doing it at any sort of scale, any sort of

2:39power as well and so there are other

2:40companies doing

2:41this stuff. You'll see companies like Apple

2:44and Google making chips for their own

2:46hardware

2:46but fundamentally NVIDIA is doing it at

2:48scale, it's doing it in the enterprise and

2:51it's also

2:51doing it on the consumer market as well so

2:53pretty much all the technologies are

2:55covering this

2:56capability and if you look at the NVIDIA

2:58stock price you can see they made a good

3:00bet at the

3:01right time. Anyway let's carry on and what

3:03Frank is mentioning here is a session

3:05previous to this

3:05keynote where he had the NVIDIA CEO talking

3:12to this sort of particular aim but anyway

3:15let's

3:15carry on and see what we say. Isn't that

3:17man incredible? What a force of nature and

3:20I love

3:21his vision you know you just put the AI

3:23factory on the data and flip it on and poof

3:25the insights

3:26starts flowing and we're just going to show

3:29you a little bit later on. It's a demo of

3:31course so

3:31don't take it too seriously but you know

3:36the two letters AI are being mentioned so

3:39it's almost like

3:40there's the only two letters left in the

3:42alphabet right and I hear there's a

3:45drinking game going on

3:46in the back that every time I say AI they

3:49have to do a shot okay so I'm going to be a

3:52little

3:52bit sparing with the AI references just to

3:55kind of keep order here but what we want to

3:58talk about

3:59what I'm going to talk about in the next 15

4:00minutes so it'll be relatively quick

4:02because

4:02I'm going to vacate the stage and we're

4:04going to really show you some incredible

4:06things. But

4:08we always say in order to have an AI

4:10strategy we have to have a data strategy

4:14you're going to hear

4:14that you know more often in the next couple

4:17of days because we can't just turn it on

4:19and hope

4:20for a miracle to appear. Data strategy is

4:22something that you know I can't help but

4:25three times a day

4:27in a normal week you know I end up having

4:29the data strategy conversation you know

4:31with with CIOs and

4:33all kinds of C-level people because it's a

4:37fundamental choice that you have to make on

4:41your

4:41journey because if you're not having your

4:44data strategy wired at hand you cannot just

4:46sort of

4:46continue on with the past you know we got

4:49to make very specific choices so that we

4:51enable these AI

4:53factories you know. And I couldn't agree

4:55with this more it's super interesting how

4:59lots of companies

4:59jumping on the AI bandwagon but as I've

5:01said in previous videos when we talk about

5:03Tableau

5:03and actually something I'm going to be

5:05covering on another channel where we're

5:07going to focus on AI.

5:08AI is going to be the technology or the

5:11innovation that really exposes companies

5:15and shows you how

5:18much or how little they've invested in a

5:20data strategy and or data cultures because

5:22ultimately

5:23as Frank is saying you can't just turn it

5:26on. Secondly this is sort of an underapp

5:28reciated

5:29aspect of this is that you really do have

5:32to have really good understanding of what

5:35your data can do

5:36what are the opportunities and you have to

5:38have been sort of down this level of

5:40maturity in terms

5:41of your data strategy to actually even

5:43start thinking about AI. It's a very

5:45similar sort of

5:46comparison to well it's actually not that

5:48different predictive analytics as a concept

5:51you know once you get to the point where

5:53you're able to actually deploy predictive

5:55analytics in

5:55a way that can actually help your business

5:57you're actually quite far down that sort of

5:59maturity

6:00scale and so companies who are that far

6:02down companies who've thought about that

6:04and actually

6:05achieving results with those sort of

6:06capabilities I think also stand to benefit

6:08here when they come

6:09with AI. But if you're just starting out if

6:11you're still pushing out excel files and a

6:13lot of your

6:13analysts are still using basic tools to do

6:15all of this stuff AI is just not going to

6:18help you it's

6:19going to be a really difficult sort of

6:21innovation to bring in because the largest

6:24the most you know

6:25computationally intensive aspect the most

6:27time intensive aspect of this is training

6:30systems AI

6:32and systems on your data and all the large

6:34language models that you've seen at the

6:36moment

6:36they're general large language models they

6:38're not specifically catered to your

6:40organization and so

6:41if you want to capitalize on this stuff you

6:43want to capitalize on the data that you

6:45already have

6:46you're going to need to have to a have the

6:49expertise to work with AI and large

6:51language

6:51models to have people who truly genuinely

6:54understand your data and can help curate

6:57what

6:58goes into training these models and then

7:00lastly once you've trained the models once

7:02you've

7:02actually got sort of the insight out of

7:04them there's going to be an element of sort

7:05of assistance

7:06all these good AI models have actually had

7:09a human element essentially human

7:10intervention to make them

7:12better and for that you're not going to

7:14just need skilled people and skilled data

7:16analysts you're

7:17also going to need your business engaging

7:19in that data conversation so all three

7:21components of a

7:22good data strategy are going to be tested

7:24to the max anyway let's carry on all the

7:26benefits that

7:26we're we're seeking from AI learning so

7:29there's real choices to be made as opposed

7:31to you know

7:32we'll just perpetuate what we've been doing

7:34in the past we're just modernizing we're

7:35going to the

7:36cloud we're using new technologies now

7:37there are fundamentally different choices

7:39to be made along

7:40the way so our data strategy no surprise is

7:44the data cloud you can't shut me up about

7:47data clouds

7:48they're my favorite two words in the

7:50dictionary these days so for that occasion

7:53you know we have

7:55i'm descending the data cloud on las vegas

7:58you know beautiful yeah so it'll be here

8:03all week

8:03and i'll move somewhere else next week but

8:06the fundamental thing you know about the

8:08data cloud

8:09is look look at this thing we call it the

8:11snow globe affectionately you know within

8:13snowflake

8:13but ever and this is not a visual uh

8:16interpretation this is an actual rendering

8:20every single

8:21white spec is a unique snowflake account

8:24and every line that you see in here is a

8:27data networking

8:28connection with another snowflake account

8:30and it's very dynamic because the data is

8:33flowing

8:34so that's fundamentally what it is now this

8:35is all like okay 10 000 feet 50 000 feet

8:38very

8:38conceptual so we'll drill down here just

8:42just for the for the for sake of argument

8:45to see what's

8:45what's interesting about this little

8:47visualization is um obviously you've got

8:50nodes where you have

8:51things connecting and super interested to

8:53see what those connections are and how they

8:55work

8:56you can see there's an element to which

8:57some of them are coalescing around each

8:59other some of them

8:59seem to be you know quite pivotal nodes

9:01maybe there's a data providers the things

9:04that are

9:04coming from the marketplace but i also love

9:06this really interesting trend where you've

9:09got lots of

9:10customers just sitting on the periphery and

9:12i wonder where they are in terms of their

9:14journey

9:14are they new customers to snowflake who

9:16just you know literally creating accounts

9:18doing their own

9:19thing not really sort of engaged in the

9:21marketplace and connect to other people and

9:23therefore is this

9:24visualization a good representation of snow

9:26flake customers or is this a representation

9:29of what

9:29snowflake would like you to do with your

9:32data i.e they want you networking with

9:34other cloud providers

9:36they want you sort of sharing your data and

9:38working with other setups it's a really

9:40interesting

9:40thing of course when you create a rendering

9:41when you create a painting when you create

9:43a picture

9:43the perspective you take it from matters it

9:46changes sort of your understanding of it so

9:48and seeing as this is snowflake's own

9:50curated image of this and it's interesting

9:53sort of what

9:53they set out to achieve by visualizing it

9:55this exact way i always think you can't

9:57take these

9:58things for granted you can't just look at

9:59this and go wow i want to be in the middle

10:01like you

10:01you've got to think like this was created

10:03in such a way to sort of communicate a

10:06message

10:06and to communicate where snowflake wants to

10:09head with its platform and i think what's

10:11really being

10:12shown here is look they're showing the

10:14power and capability of having all of this

10:17stuff in the

10:17cloud making it easier for all these

10:19connections to happen and opportunities to

10:21come out of that

10:22anyway that's just my take on that what's

10:25going on below the covers this is kind of a

10:27google earth

10:28like kind of thing we tried to make it look

10:30like that and we drilled in to the data

10:33clouds of one

10:34of our great customers fiserv thank you fis

10:36erv by the way you know for letting us talk

10:38because for

10:39a lot of people this feels very proprietary

10:42but fiserv is more than okay we're running

10:45snowflake

10:45they have data networking relationships

10:48with other entities i just want to use this

10:50by example

10:50to see look there's hundreds and thousands

10:53of these living in the data cloud but for

10:56example

10:56they have a connection with a company

10:59called clover that's actually a wholly

11:02owned subsidiary

11:03you know of fiserv they in turn have a

11:06relationship with heap analytics they

11:08acquire

11:09that data to create custom 360 type of

11:11views right i think what's it like it's

11:13exactly what i just

11:14said it's now playing out they're showing

11:16you look here's an ink customer and thank

11:17you for letting

11:18us talk about this and by the way look look

11:20at what they're doing they're connecting

11:21with some

11:22of these other businesses and this is how

11:23they're doing this and they're sort of

11:26really trying to

11:26showcase this capability why this is a usb

11:29of snowflake this is a core capability that

11:32it can do

11:33that is broadly quite hard for other

11:35databases to do for other databases to do

11:37this what they have

11:38to typically do is stand up a new

11:40environment and in that environment

11:42companies exchange information

11:44but typically that environment is either co

11:46-located somewhere or shared there's a lot

11:48of replication

11:49that goes on and that replication itself

11:51has costs in terms of bandwidth traffic

11:54storage compute all

11:55of that stuff almost has to get duplicated

11:57just for the purpose of sharing and a usb

12:00of snowflake

12:00is that actually um a lot of that work is

12:02taken care of for you and you can just get

12:05on with the

12:05ideation of setting that kind of connection

12:07with another business rather than having to

12:09you know

12:10worry about all the issues that come out of

12:12that and because of snowflake's sort of dec

12:15oupled

12:15setup with compute and storage being decou

12:18pled it actually makes that integration

12:21makes that sort of

12:22collaboration a lot easier i serve as a

12:24very large financial services concern you

12:27probably remember

12:28first union credit card processing things

12:31like things of that sort so creating data

12:34is is very

12:35very very central to to how they operate

12:38and what they do for a living but the data

12:40connections

12:41they're like synapses not that i like to

12:43use that word but you know they they are

12:45creating synapses

12:46that are literally i'm sure he loves using

12:48that right here's the relationship with

12:50their banking

12:51clients right there is there is a fintech

12:53in there as well that either uses their

12:55data or provides

12:56data back to visor and it builds and builds

12:59and builds we're just showing a few of them

13:02here

13:02carrot is actually a data as a services

13:05company that five surf you know has built

13:07itself and

13:08that's how they provide data to other

13:09entities in this case exxon mobile

13:11everything is snowflake to

13:12snowflake to snowflake to snowflake so the

13:15point here of a data cloud is is not

13:17defined by

13:18enterprise boundaries anymore it's defined

13:21by your business relationships defined by

13:23your ecosystem

13:24and it's dynamic right you can build these

13:26connections you can take them away you can

13:28reconfigure them and that's really what's

13:31going on every single day of the week all

13:34over the data

13:34cloud now when i zoom back out i'm just z

13:37ooming out not completely i'm just zooming

13:41out to the

13:42financial services data cloud this is our

13:44industry data cloud this is our only

13:45entities that are in

13:47the financial services business come

13:49together you know we now have seven

13:51industry data clouds there's

13:53all these virtual different views either as

13:55a as an institution as an enterprise i have

13:58my data cloud

13:59industries have their data clouds and then

14:02we have the big big snowflake thing across

14:0540 different regions all over the world so

14:09i want to emphasize a couple of really

14:12key underpinning concepts here so this that

14:16really shows that rendering is a marketing

14:20tool

14:20that pretty much sums up that sort of trail

14:22of thought it's a marketing tool designed

14:25to show

14:25you what you should be doing with snowflake

14:27and the opportunities that come out of that

14:29i think

14:29the the financial services uh snow you know

14:32zoom out that he did whatever you want to

14:34call that

14:35that's again another big sort of call out

14:37hey you know everyone in the sector look

14:39this is what's

14:40going on and by the way you all play the

14:42same game and it's quite handy to be able

14:43to um share data

14:45now what what is always interesting about

14:48this is that you know these sort of node

14:50diagrams make it

14:51look so easy and sort of flexible to do at

14:53the heart of that is customer data right

14:56these

14:57companies are all doing businesses with

14:59people and those customers that data the

15:01the activities

15:02is generating lots of data points and that

15:04date those data points are what's actually

15:07being

15:07exchanged here and so um he didn't touch on

15:09it here but i think probably touching it

15:11later like

15:12you need to also bear in mind privacy and

15:14security concerns with moving data from

15:16people and you

15:18especially in europe and especially in some

15:20countries where there's quite high

15:21implications

15:22of moving customer data around you need to

15:24also be treating the data in a very secure

15:26way and you know

15:26snowflake has features to handle that but

15:28it's interesting that's not touched on here

15:30especially

15:30when you call out to finance finance to me

15:32is one of those mostly heavily regulated

15:34industries and so

15:35him not to mention that so slightly strange

15:37maybe it's coming later but i just wanted

15:40to sort of

15:40highlight that just so that you're not

15:42sitting there going oh you can't do these

15:43like yes it's

15:44possible to manage all of those um

15:46capabilities inside of the platform so what

15:48you see here is

15:49the data silo 70 years ago or according to

15:52jensen 60 years ago when the first database

15:55you know was

15:56created there was not a siloing problem

15:58because there was only one data base that

16:00was great when

16:01there were two databases now we have a sil

16:03oing problem why is this table over here not

16:05over there

16:07ever since you know we've been mushrooming

16:10like crazy right this is an incredibly hard

16:13problem because silos are created just by

16:15running workloads this happens on my own

16:18you have to literally keep creating

16:21databases very strong strategic posture

16:24towards a single

16:25data universe without boundaries obviously

16:29this is hugely important to enable the ai

16:32revolution

16:33take another shot but any other form of

16:36data science analytics it's incredibly

16:40inhibiting

16:40and impeding you know when these kind of

16:42boundaries say okay now we got to start

16:44pumping

16:44data across again ftp kpis all the three-

16:48letter acronyms so unsiling is really the

16:52core theme

16:53of what we try to do this is like a matrix

16:57moment okay we're evaporating the silos isn

17:01't that

17:02beautiful but keeping it going like that is

17:05the hard part now secondly the data cloud

17:10doesn't

17:10just consist of structured data obviously

17:13in the world data warehousing it was

17:16hopefully concepts

17:17of being able to support lots of different

17:18types of data structured data semi-struct

17:20ured data and

17:21structured a couple years ago and we really

17:24started we can we will do it beyond what

17:26those

17:26three means at some point in the future but

17:28it's pretty straightforward if you can just

17:30have a look

17:30for ai all these applications are very much

17:33training on textual data but the problem

17:36with

17:36unstructured data is how does it become a

17:38full participant in analytical processing

17:41if we know

17:42very little about it right it may have a

17:44name on a timestamp and what's inside of it

17:46very human

17:46readable but not readable for software so

17:50last year we acquired a company called appl

17:54ika later

17:55on this session we're going to refer to

17:57this as document ai this is the ai that we

18:00already

18:01acquired into the talent and the technology

18:04more than a year ago to help us derive

18:07structure from

18:07unstructured documents using language

18:10models and using ai techniques i really

18:12love this rendering

18:13you get this formless mass turning in so at

18:16this point it's probably you know worth

18:18highlighting

18:19i'm gonna have to explain this unstructured

18:21data in this case he's talking about

18:22documents

18:23another form of unstructured data could be

18:25something like a video or a photo where

18:27there is

18:27it's information it's stored in a specific

18:29way but it's not in a typical structure

18:31that a database

18:32would look at columns rows and metadata

18:34that come come from that so um applika i

18:37think have a

18:37capability i don't know too much about appl

18:39ika i'm just going off what he's just said

18:41but applika

18:42have a capability of being able to process

18:44documents in a specific way using ai but i

18:47'm

18:47sure using other technologies like ocr

18:49optical character recognition all that

18:52stuff being bundled

18:53into a beautiful application called applika

18:56and i think um yeah stokebook acquiring

18:58that kind of

18:58company means that they've now got an

19:01ability to extract data from these often

19:03difficult to

19:04extract data from i think a classic example

19:07is invoices in pdfs so many companies trade

19:10invoices

19:11just by pdfs that's all you get you don't

19:13get an excel or csv file that you can do

19:14something with

19:16and so the pdf becomes a data source and so

19:18if you then have to go and scrape

19:20information out of a

19:21pdf the current structure is to have

19:22someone you know manually do it there are

19:24some technologies

19:25if you look at altrix altrix has a

19:26capability that kind of allows you to

19:28scrape pdfs it's a little

19:30bit clunky but it kind of works if you know

19:31what you're doing and you're doing it in a

19:33structured

19:33way um and you know technologies like appl

19:35ika hopefully allow you to do this at scale

19:38inside

19:38of the database meaning you can just put

19:40all your documents inside of the database

19:43and then

19:43have this technology sort of go and do this

19:45analysis for you so that's a bit of context

19:48for that true i know the set of structures

19:50but this really helps unstructured objects

19:53to become

19:53full participants in analytical processing

19:56right so that really amps up that that

19:58opportunity we're

19:59very excited about it you're going to see

20:00more about it later this morning so it

20:02doesn't stop

20:05there one of the things that we have

20:08invested most of our engineering resources

20:10in in terms

20:11of enabling the data cloud are workloads

20:14right i mean the whole point of the data

20:16cloud is the work

20:18needs to come to the data we want to stop

20:20the data from going to the work because

20:22that endlessly

20:24silos and re-silos the world yeah so if we

20:27can bring the workload capabilities full

20:30spectrum

20:30from analytical to transactional to search

20:33and everything in between then the data can

20:36stay put

20:36because the work can be executed

20:38effortlessly effortlessly you know on that

20:40data now there's

20:41a lot here okay so just to explain that in

20:44more detail what is essentially talking

20:47about is let's

20:47say you have a database with information in

20:50it and you have your analytics team that

20:52wants to do

20:53analytical processing to that data they

20:55maybe want to do some transformation you

20:57might think of terms

20:58like etl you might want to then do some

21:01statistical analysis and then a data

21:04science on top of that

21:05some predictive analytics some workloads on

21:08top of that what has typically happened in

21:10the past is

21:11your database is the source but then you're

21:13constantly sort of extracting the data and

21:15taking

21:15it somewhere else to go and do the work

21:17right so because databases typically haven

21:19't had these tools

21:20inside of them your database is just a

21:21store think of it just as like a cupboard

21:23or like a library

21:25if you want to go read a book you could

21:26read it in the library it's possible to do

21:28that but actually

21:29it takes quite a long time to read a book

21:31so what you end up doing is taking the book

21:33home to do the

21:34work and so this is essentially what he's

21:36explaining here he's explaining that these

21:39workloads have typically had to move out of

21:41the context of the database but actually

21:43they've been

21:43investing money and time into making sure

21:45that these workloads can happen in the

21:47database now

21:48this has two benefits the first one is your

21:51data doesn't have to move and actually

21:53moving data is

21:54quite expensive why well if you imagine

21:56data at the kind of scale and volume that

21:59enterprises have

22:00and the simple cost of moving data from one

22:03place in the next is actually quite hard

22:05and if you

22:06think of the scale they're doing it it

22:08doesn't even make sense to move it over the

22:10internet and

22:10in some cases because the data is just so

22:12large and the bandwidth just doesn't exist

22:14to move that

22:15kind of scale of data and so you you maybe

22:17you've heard of something called aws

22:19snowball it's

22:20literally a hard drive that's shipped to

22:23your company because it's faster to plug

22:25the hard

22:26drive into your server copy the data across

22:29ship it to amazon and then have amazon

22:32import it into

22:33s3 for you it's the same concept now what

22:35if you don't have to do that to move the

22:37data around what

22:38if you could just leave the data where it

22:39is and process it and that's exactly what's

22:41going on here

22:42there is also this other concept known as e

22:44gress cost essentially the cost to move data

22:47around can

22:47be more expensive if you do it using um

22:50certain routes and paths rather than you

22:53know the route

22:54and path that's intended by the platform

22:57provider a good example again is aws if you

22:59're moving data

23:00from one data center to another absolutely

23:03fine if you move it from one data to

23:05another via the

23:06internet it's going to cost you a lot more

23:08because you're not using amazon own sort of

23:11cost effective

23:12let's call it tunnel road for the data you

23:14're using the public internet and that

23:17obviously

23:17costs amazon a lot more money to do so um

23:19things like that start to come into play as

23:21well starting

23:24off i know most of you in this room are

23:26doing data warehousing data warehousing

23:28once upon a time it

23:29was a market it was an industry there were

23:31whole companies that's all what they did so

23:33hey is going

23:33through the different sort of data

23:35warehouse which makes sense type of

23:37workloads we can kind of argue

23:38about exactly what it is and what it is and

23:40we sort of kind of know what it is so

23:43beyond that

23:44you know we obviously we have data lakes

23:47now data lakes are maybe not the same the

23:50data warehouse

23:51what's a data warehouse what's a date this

23:55is another video easily i'm not going to

23:56get into

23:56that now because i think snowflake has the

23:59daily to be honest i don't think i'm 100

24:01sure what it

24:01is because it's very very broad and i won't

24:03pretend to know this morning you know how

24:05we're even more

24:07expanding our data lake scope in terms of

24:09the type of data that we can address but

24:12make no mistake

24:12about it that is very much a workload that

24:15we execute on unit store we talked about

24:18that last

24:19year these are our transactional capability

24:22this is about as different as you can get

24:24you know from

24:24large scale analytical processing you know

24:27we now have to update data objects in place

24:29it really

24:30implies a very different underlying stack

24:32we introduced hybrid tables hybrid table is

24:35an

24:35object that is both transactional and

24:37analytical right we now have a handful of

24:39customers already

24:41in production even though the capability so

24:43um transactional data sources um if i give

24:47you an

24:47example let's say your supermarket when you

24:50're taking data from your point of sales

24:52your till

24:53as it were you're scanning transactions and

24:56you do a purchase the customer pays for it

24:58and everything

24:59then goes into a system right that is

25:01almost known as a transactional data store

25:03and that has to be

25:05running all the time if it doesn't if it

25:07doesn't run you can't store the information

25:09that records

25:10the cells and you'll find that whenever

25:12that kind of system goes down everything

25:15stops because

25:15um even if you allow that to sort of let's

25:17just let's just assume you keep going for

25:20an hour

25:20without that working you'll just create an

25:22absolute nightmare so um it's a very

25:24different

25:24type of database and different type of

25:26workload compared to analytical processing

25:29where you have

25:30things set up in a way that allow you to do

25:32essentially analytical tasks the two don't

25:35work

25:36well for each other's use cases and so what

25:38he's talking about here is look we can have

25:40something

25:41called a hybrid table which can both be

25:43analytical and transactional at the same

25:45time so again um

25:46i'm just ticking off topics we're going to

25:48cover in the future related to snowflake

25:50but um i think

25:51it's super useful context to be aware of

25:53and i just want to try and add some color

25:55to these

25:55examples because i see most people watching

25:57this maybe don't know too much about

25:59databases

26:00or snowflake and if you do and i'm wrong

26:01let me know in the comments there but let's

26:03keep going

26:04abilities themselves are still in preview

26:07so there's tremendous demand for this and

26:09we'll be

26:09pushing very hard you know to get bigger

26:11and stronger and more functional in this

26:13area so

26:14stay tuned for that very important for our

26:17isv customers as well and so it goes these

26:21are all

26:21workloads are all orbiting around data

26:24cloud collaboration data sharing makes

26:26sense 70 percent

26:28of our large we define large consumers as

26:30snowflake customers that consume them a

26:33million

26:33dollars a year are using data collaboration

26:36on average they have six to seven of these

26:39data

26:39networking relationships we call them edges

26:41um you know we have all different kinds of

26:44edges

26:4470 percent of them are using it and it's

26:46literally growing in leaps and bounds this

26:49is a very

26:49important part again data cloud is defined

26:52by its networking relationships not by any

26:55arbitrary

26:56workload perimeter enterprise perimeter or

26:59anything of that sort interestingly

27:03cyber security became a workload on snow

27:05flake initially it was really an economic

27:07play because

27:08people said hey i can save 75 cents on the

27:11dollar just by keeping most of that data in

27:13snowflake instead of in a in a sim but then

27:15they figured out you know cyber security is

27:18also a big

27:18data integration problem right if i'm a

27:20security analyst you know it's much easier

27:22if i can query

27:23a series of ip arranges across all these

27:26different data sets you can just imagine if

27:28you let loose

27:30one of jensen's ai factories you know on

27:32top of this data you know what kind of

27:34questions

27:34you'll be able to answer it's very exciting

27:36cyber security we move on data engineering

27:42yeah we'll

27:43we're all data engineers here in this room

27:46you know we sometimes estimate it's very

27:48hard to

27:49know exactly that roughly 40 percent now

27:54data engineering i i hate this term not

27:57because i

27:58don't like the activity not because i think

28:00it's a bad thing and not because i don't

28:01believe in it

28:02data engineering is absolutely something

28:04that's here to stay it's a good activity

28:06what i found

28:08interesting is that i think data

28:10engineering has been going on for a long

28:12time but very recently

28:14it got branded as such data engineering it

28:16got this nice sexy name tools like dbt came

28:19up and

28:20really sort of helped amplify this branding

28:22and when in fact in my view it's been

28:24happening all

28:25along it's just been happening in different

28:28places and now it's come to the point where

28:30everything is

28:30happening at such a scale that it's become

28:32its own distinct activity you actually have

28:35data engineer

28:36professionals who are not doing analytical

28:38work in the you know data visualization

28:40tools like they

28:40used to they're not doing any data ware

28:42housing or data lake architecture they're

28:44not doing any of

28:45the other workloads that have been talked

28:46about before their sole purpose is just to

28:49make sure

28:49data goes from a to b in a way that's

28:51structured easy to use and consistent so

28:53that as you start

28:55to grow your business on these data

28:56platforms you can answer more and more

28:58questions as they come up

29:00rather than what people have been finding

29:01in the past that you know as soon as a

29:03question or

29:03business problem comes up you have to stand

29:05up whole analytical workflows just to

29:07answer that

29:07one question so data engineers have pretty

29:09much sort of solved that problem but that

29:11term sounds

29:12like a new term sounds like something

29:13really new and sexy and it is it's been

29:15marketed as such but

29:16maybe i'm wrong maybe maybe i've completely

29:19misunderstood uh the origin of the term

29:22data

29:22engineering uh let me know but um yeah it's

29:25it's one of these sort of i would call it

29:27hypes in

29:28analytics and data at the moment and i

29:30think it will it become its own thing and

29:32there'll be a

29:32new one something else will sort of come

29:34out of data engineering and it'll be it'll

29:36be called

29:36something else like metadata engineer or

29:38something like that where you know you're

29:40not working with

29:41the pipelines you're just working with the

29:43metadata and stuff like that will come up

29:45so yeah um let's

29:46keep going of the compute that we consume

29:49in the snowflake data club actually is data

29:51engineering

29:52function so it is enormous um all our

29:54efforts i wouldn't say all our efforts a

29:57lot of our efforts

29:58in snow park over the last six months ever

30:01since python went ga have been focused on

30:03really retiring

30:04you know legacy spark jobs this is the

30:06easiest pickup if you're looking to save

30:08some money i

30:09mean you can you can you can save two to

30:12four x on your spend on data engineering

30:14and that is because

30:16you know our interpretation of spark snow

30:19park i mean it will uh it will run faster

30:22cheaper

30:22operationally simplified and much safer in

30:26terms of governance so we've seen enormous

30:28uptake on

30:29data engineering in snow park and if you

30:31haven't looked at it encouraged you to do

30:33it there's just

30:33free money to be had there it's a great

30:35source but it's free money to be had all

30:38these plans oh

30:39ai another shot we're going to talk about

30:42this a great deal you know 70 of our large

30:45consumers in

30:46snowflake obviously you know are running

30:48these type of jobs i mean ml has become

30:50very mainstream

30:51over the last couple of years we see people

30:53running models all over the place and this

30:55is

30:55just going to go into overdrive in a huge

30:58way uh that's because of all the new

31:00technology that's

31:01arriving one of the big things we're going

31:03to talk about this morning is how we are

31:05going to enable

31:06that inside of snowflake because that's

31:08really important for you to know exactly

31:10how we're going

31:10to enact all of this applications we've

31:15made a huge effort for snowflake to become

31:20a place where

31:22people want to build applications because

31:24essentially you know it's it's a layered

31:26cake right

31:27with the infrastructure because of the

31:29public cloud you know we have the live data

31:31we have the

31:32workload enablements we have the market

31:34places we have the transactional capability

31:36for monetization

31:37that's a very attractive place for software

31:39companies and software development to come

31:42to

31:42build something to sell something to

31:44transact something and we now have hundreds

31:47and hundreds

31:47of isvs something some of them quite large

31:50actually that are now building on snowflake

31:53we're really looking to to set up a

31:55renaissance in software development on snow

31:56flake we really

31:57think that applications should be built on

32:00a data cloud you know not on the database

32:02it's a really interesting take uh at this i

32:06don't know so applications are

32:10you know what is an app that is that's

32:13probably where this sort of debate starts

32:17what is an

32:17application and um you know he's typically

32:20calling out here on stage he's saying

32:22software developers

32:23can come and build applications on our

32:26platform but i don't think it's just

32:28software developers

32:29i think i think there are people who work

32:31in analytics who you know work in as simple

32:34tool

32:34as excel who go away and build really

32:36complex sort of setups in the excel um work

32:40books and they will

32:41call those applications and so i think the

32:43term application it means a lot of things

32:46and you have

32:46to be super clear about what type of

32:48applications are we talking about uh things

32:51that should be uh

32:53dashboards or things sorry things dash

32:55boards that should be applications are we

32:57talking about that

32:58sort of level of of interaction where it's

32:59it's actually no software development

33:01required

33:02whatsoever it's a really smart data-driven

33:05dashboard with a couple of um you know

33:08actions

33:08that come out of that that let you do

33:10various things um in the tableau world

33:12tableau's tried

33:13to do this with snowflake and sorry with

33:14the salesforce and a couple of other

33:16capabilities

33:17and they've also you know introduced things

33:20like um extensions to allow you to do

33:22actions out to

33:23other systems and platforms and but when we

33:26talk about applications in my mind um it's

33:29interesting

33:30i think i think snowflake the bigger the

33:33bigger piece to me is actually enabling

33:36your data atlas

33:37you know people who are doing analytics on

33:39these uh platforms to be able to write a

33:42software and

33:43applications without having to worry about

33:45software development and everything else

33:47that

33:48comes with that that's sort of the beauty

33:49of that and there is there is actually a

33:51couple of a

33:52couple of companies in the world doing this

33:53uh retool is a really good example and ret

33:55ool

33:56has sort of taken the approach of giving

33:58you a an interface to be able to design

33:59applications

34:00and then essentially take the software

34:03development out of your hands and you still

34:05need to know a bit

34:05about your data you still need to know a

34:07bit about what you can do the metadata and

34:09all of that but

34:10fundamentally your job is not to do your

34:12software developer your job is to

34:14understand the business

34:15process and they give you an interface to

34:18build an app that solves that business

34:20problem in my mind i

34:21think that's a slightly better solution and

34:23i think snowflake could could could do from

34:26learning

34:26a lot from that um there is also this sort

34:28of tendency when you're working with

34:30databases to be

34:31you know proper nerdy and say oh here are

34:33all the tools here are all the things but

34:35the the fact of the matter is those tools

34:37in themselves are not always all accessible

34:40they

34:40come with their own baggage they come with

34:42their own learning curves and so if you're

34:43trying to you

34:44know go out to the data cloud i think apps

34:46to me have to take on a new meaning it has

34:49to be more

34:50than just you know apps that software

34:52developers build to me it has to be apps

34:54that we see in

34:55businesses that people are building in

34:56excel that people are building in tableau

34:58that they really

34:59shouldn't be building there because you

35:01want all of that activity to be happening

35:03on your data so

35:04you can a see the metadata that comes from

35:06that b and you can potentially monetize it

35:09for specific

35:10uses and gains within an organization or

35:12publicly through collaboration and other

35:15capabilities but

35:15you know at the end of the day that that's

35:17just my take on that but you know it's an

35:19interesting play

35:19and i think why we believe it's easy to get

35:21lost in that play now there's lots of them

35:23out there

35:24some of them are built by large enterprises

35:26that are building customer facing

35:28applications um you

35:30know dtcc a great example that's actually a

35:32regulatory function in the financial

35:35services

35:35market they're really owned by their member

35:37banks they manage the liquidity levels that

35:40banks have

35:40to maintain it's very very cool and because

35:42so many financial institutions are on snow

35:45flake

35:45it's all snowflake snowflake type of

35:47functionality um she's simon data bank of

35:52new york melon

35:53um octa and it goes on and on and on one of

35:56them called out specifically i was on a

36:00stage

36:00actually here in vegas a couple of months

36:03ago with blue yonder blue yonder is the

36:05former jda

36:06they are the largest supply chain

36:08management software company in our industry

36:10i've always had

36:11a personal affinity with supply chain

36:13management because it's one of the largest

36:16one of the last

36:17remaining areas in the enterprise that has

36:20never been platformed there's actually

36:22reasons why it's

36:23never been platformed and it's very

36:25frustrating because it has been a

36:27spreadsheet email business

36:28very very inefficient all right and there's

36:31really an enormous opportunity for for

36:33automate for

36:33digital transformation and really you know

36:36bring this into the modern age but what are

36:38the problems

36:38in supply chain management supply chains

36:41consists of many different entities right

36:43so getting

36:44visibility across the data on different

36:47entities guess what you know we need a

36:49single data universe

36:50you know we can't have these act we can't

36:52have these these uh these silos and all

36:54these boundaries

36:55stand in the way of that it's fundamentally

36:57a data integration problem we can solve

36:59that with

36:59snowflake and the second thing about blue y

37:02onder is a supply chain management is you

37:05know when they

37:05have events happening in the supply chain

37:08they run these incredibly compute intensive

37:10processes

37:11very very fast uh to be able to figure out

37:13what do we do what are our options what are

37:15the scenarios

37:16that we that we can enact here and of

37:17course you know snowflake lends itself

37:19really well you know

37:20because we can fire up instantaneously

37:22these workloads massively provision and

37:24then get the

37:25results um so we finally have an

37:26opportunity to go after supply chain

37:28management and of course

37:29the network effect for retailers and

37:31consumer packaged goods companies is going

37:34to be enormous

37:35so blue yonder is a 1.5 billion dollar

37:37software company so it's quite large for

37:39people to re-platform

37:40onto snowflake so we're excited about that

37:44but they are now one of the things that's

37:48i remember five years ago in san francisco

37:502019 was my first uh summit so what what

37:54else are

37:54they're stopping this here because it's

37:56quite evident what snowflake is doing here

37:57in this

37:58initial part of the keynote they are they

38:01are framing this is a sales pitch

38:02everything going

38:04on here is a sales pitch they start off

38:05with this network diagram then they show

38:07you this universe

38:08with sort of planets orbit it's orbiting

38:10this sort of data cloud as the sun there's

38:13a lot of good

38:13analogy used there and then outcomes

38:16governance and that's sort of this you know

38:19invisible web

38:19that goes around this whole thing to give

38:21it structure then will come security then

38:24will come

38:24a bunch of other things um so this this

38:27sort of opening keynote like now that i'm

38:29sort of seeing

38:30it play out is this opening of the keynote

38:32is actually more about framing what snow

38:35flake is what

38:36it can do today and where it's heading that

38:39's ultimately it conference i remember doing

38:42a

38:42presentation i don't remember what i said

38:44but i walked off the stage and somebody

38:46said you didn't

38:47talk about governance and i'm like what i

38:50didn't know much i was only at a job for

38:52six weeks at a

38:53time and we we said yeah we didn't talk

38:56about governance and we should have because

38:59governance

39:00is so central to the data cloud right i

39:03mean what this visual is trying to show you

39:06is that there's

39:07there is a shield there's a protective

39:08shield around everything that we're doing

39:10and you may

39:10like well that's nice but the problem is

39:12you know when you come from an on-premise

39:14world you know

39:15you have your own security perimeter yeah

39:17when you get to the cloud that becomes a

39:20whole different

39:20dynamic so governance has sort of been been

39:23infiltrating all our thinking every time we

39:26do

39:26something we need to fully think through

39:28all the risks you know that are represented

39:30by new

39:30functionality the risk of exfiltration the

39:33risk of you know compliance violations all

39:36these types of

39:36things you know when we brought python out

39:39i mean you can do that in a week if it's

39:40just python right

39:42but the way we had to do it to harden it

39:43make it non-porous and really eliminate all

39:46the risk

39:47it took two years to do that to really make

39:49you know python a enterprise-grade you know

39:52high trust

39:53capability so every time we do something

39:56this is this is the first step this is also

39:58why it takes

39:59time for us right to take things out of

40:01preview because these things are in preview

40:04guess what

40:04it's always security compliance type of

40:06issues that are slowing down the release of

40:08these features

40:09so everything we've done in terms of

40:12workload enablement it runs inside that

40:16governance

40:17perimeter and that's really the the big

40:19change from data cloud from before where it

40:22's just a data

40:23universe a big one and something that runs

40:25across enterprise boundaries and so on but

40:28we brought

40:29the workload platform into the governance

40:31perimeter as well we call that snow park

40:34snow park is the

40:35programmability platform for snowflake and

40:38it is completely governed in terms of its

40:40implementation

40:41very important aspect of what we do now the

40:45user types you know historically you know

40:48when you come

40:49from where we come from as database people

40:52of course you know sql people you know

40:54obviously

40:54the general population of very important I

40:57mean this really speaks to what I was just

41:00saying about

41:01the kinds of people like when when snowfl

41:06akes talking about app app builders I think

41:11every

41:12single one of these users has the potential

41:16of building an app and to just only assume

41:19that only

41:20app developers can build apps is sort of an

41:23interesting thing it's very similar to you

41:27know

41:27let's take video editing right like video

41:29editing used to be something that required

41:31a very specific

41:32skill to do and then apps like tiktok and

41:35instagram reels and youtube shorts came out

41:38and now

41:38virtually everyone can edit a video on

41:41their phone without any prior experience of

41:44knowing

41:44how to edit a video and more so you can

41:47create videos that create more engagement

41:49than someone

41:50who's who's made a you know a blockbuster

41:52film or someone who's been editing all

41:55their life

41:55in traditional tools you can get a 30

41:58second 40 second video that gets far more

42:00interesting

42:01engagement even amateurs can pick up an ip

42:04ad or an iphone record the video on their um

42:07you know

42:07devices go into imovie and create a simple

42:09edit put it on youtube and more people will

42:12see that

42:12video than again a reward-winning film

42:16might do at the movies and so I think you

42:19have to sort of

42:21critique the fact that is snowflake maybe

42:24thinking too narrow about its potential for

42:27apps and maybe

42:27that's essentially driven how they've built

42:30the solution and why can't a data scientist

42:33build an

42:33app why can't a data analyst build an app

42:36why can't business users build apps and why

42:39can't

42:40you enable them to do that on your platform

42:41that that would be my question to snowflake

42:43over the last couple years we have

42:46massively focused on the left side of the

42:48spectrum the

42:49much more technical user you know to give

42:51them you know the type of experiences and

42:53the type

42:54of interactions you know that they want to

42:56have so we really address the entire

42:58spectrum and we view

42:59all these roles as equally prominent and

43:01important and we will continue down that

43:04path

43:04so we'll make three announcements really

43:08quickly and then in the next section we're

43:11going to take

43:12you 15 feet on the ground show you in

43:14excruciating detail what's going on they're

43:16very very excited

43:17this morning to share these things with you

43:19because some of these things have been in

43:21the

43:22works for the last year some of them have

43:24been in the works for much longer than that

43:26so the first

43:27one is you can probably interpret this

43:31picture icebergs we have talked about our

43:35support for

43:36iceberg open table file formats for some

43:40time but we now have greatly expanded that

43:44there is

43:45a whole discussion here you know when do i

43:47use you know snowflake internal objects

43:49when do i use

43:50iceberg open table format do i want to

43:52manage my own storage do i not want to what

43:54functionality

43:55if i'm given up not given up is it a

43:57managed versus an unmanaged object we're

44:00going to talk

44:00about that but this is a huge step forward

44:03we really obviously the goal here is we

44:05want you

44:05to be able to reference i know nothing

44:07about icebergs other than you know the

44:10icebergs that

44:11live out at sea so if you know what this is

44:13about if you've got some prior contacts let

44:15me know this

44:16is a completely new concept to me so i'm

44:18sitting here going okay um please explain

44:20to me what you're

44:21talking about it's kind of like an insider

44:24sort of feature here like there's been lots

44:26of discussion

44:28okay yeah tell me tell me what it is any

44:30and all data so we're putting the

44:32infrastructure in the

44:33data cloud to make sure there are no limits

44:35to your ability to reference and address

44:38data you

44:38know especially in the face of you know all

44:40the new developments in in data science

44:43super important

44:44no limits icebergs

44:46what we're also announcing number two today

44:51is what we call a native applications

44:55framework

44:56we've been working on this for some time we

44:58have by the way when we go to the to our

45:01partner

45:01pavilion all the exhibition areas you'll be

45:03able to see these applications these are

45:06built in our

45:06native applications framework nice the

45:09highest level sort of analogy that i can

45:11give you for

45:12what this is sort of the the apple iphone

45:16their app store the the the apple iphone

45:18apps there is

45:19now such a thing as a snowflake application

45:24and we're now now comparing yourself to

45:27apple is a

45:28bold move but actually i think it's it's it

45:34it shows and speaks to um maybe the

45:39aspiration

45:40the interesting thing about that comparison

45:45though is again like who are you catering

45:49to and

45:54i think companies with an existing

45:57application sort of mindset will have

46:00people that can already

46:01go and do this you see here on screen you

46:02've got capital one you've got mapbox over

46:04there on the

46:05top right i don't know if that's medium um

46:08there's a carto or cap toe and all the

46:11others are a bit

46:13sort of um sort of foreign to me but

46:14nonetheless these are companies that have

46:16been building apps

46:17or know how to work with data platforms and

46:19build these integrations they just have

46:21these people on

46:22staff and they can take them off one

46:24project redeploy them somewhere else but

46:26you know to

46:26everyday businesses who don't have those

46:29people um the question is still um like why

46:32why should

46:33i go and do this and what's the benefit

46:35what benefit is there in in doing this um

46:38on the

46:38platform and again i just still think if

46:41you just take data as a very simple sort of

46:45concept there's

46:46just so many people building app-like

46:49behaviors into very simple things to me

46:52that is a bigger

46:54opportunity um far bigger than um you know

46:56these very sort of complicated

46:58infrastructure projects

46:59which which really only serve business sort

47:02of use cases what about serving sort of

47:04much smaller

47:05simpler to solve problems building

47:08applications as snowflake applications we

47:10're super excited

47:11about that because they're using all the

47:13common services they're using all the

47:15common governance

47:16frameworks obviously our database engine

47:18which is really the centerpiece to the

47:20entire platform

47:21that's all here now

47:25if you're on the iphone and you want to you

47:32want to run on android different

47:35development environment

47:36different deployment environments you

47:38pretty much have to maintain separate

47:39platforms to address

47:40that the great thing here in cloud you

47:43built one application that can run un

47:45modified completely

47:47agnostic to the underlying cloud so if you

47:49're a software developer you can adjust the

47:51entire

47:51market with a simple with a simple

47:53implementation definitely and we think that

47:56's incredibly cool

47:58so these apps there's going to be hundreds

48:03of them if we're successful

48:07there are going to be thousands of them

48:09this is really a core part of our strategy

48:11an app

48:11interesting line there if we're successful

48:13there's going to be thousands of them i

48:15mean

48:15back yourself a little bit i don't i think

48:18that's sort of an unnecessary line maybe it

48:20's just frank

48:21being honest right like he strikes me as a

48:23very honest presenter you know says it how

48:25it is but

48:27yeah i think i'm going to definitely need

48:29to watch a session on this and to try and

48:31better

48:32understand sort of where they're going for

48:34this and you know are these app developers

48:36here or are

48:36they busy building apps for the iphone if

48:39it to carry on with the analogy right like

48:41um snowflake

48:43to me hasn't felt like a destination for

48:45app development when i open as a customer

48:48when i

48:48open snowflake and i go look at my data

48:51nothing screams to me app development

48:52nothing screams to

48:54me build an app to solve your business

48:56nothing screams to me unload your videos

48:58and let's

48:58analyze them as unstructured data and so

49:02you know if that is the case if this is

49:05sort of a really

49:06big thing and there's definitely a huge

49:09marketing piece to help existing customers

49:12like me and

49:13existing users of the platform know that

49:15this is a functionality and these are the

49:17kind of people

49:17to start getting into organizations to help

49:19you achieve these goals which are not just

49:22limited to

49:22application developers software companies

49:25everybody these days is a software

49:26developer everybody

49:27is building customer facing applications

49:29internal applications and so on and finally

49:33we're

49:34announcing snowpark container services you

49:37just saw an app disappear in that container

49:39this is a huge expansion of snowpark and we

49:42're going to show that in dramatic fashion

49:45in the next

49:46segment when we first introduced snowpark

49:50and we started to ga python a lot of people

49:54saying yeah

49:54this is great for a spark job and the dupe

49:56jobs and you know we recompile them you

49:58know run them

49:58on snowpark but a lot of people say wait a

50:01second you know i have engines i have

50:03applications i don't

50:04even know how that you know what to do with

50:06these things you know i cannot just sort of

50:08you know

50:08rewrite them and recompile them and all of

50:11that that's why we developed container

50:14ization right

50:14we can take whole applications whole sets

50:17of services and we can now run them inside

50:20the

50:20snowflake governance parameter and if you

50:23've been on the uh the pavilion for partners

50:25there's quite

50:26a few of them that are already running

50:28inside the container it's actually quite

50:29easy to do even

50:30though this is a very very new thing a lot

50:32of people are already running it very very

50:34quickly

50:35so the concept of container is um uh our

50:38user our user computer analogy so let's say

50:42you have

50:42your laptop or your computer right and in

50:44order to install software you have to get

50:47the software

50:47for the version of whatever you're running

50:49so if you're windows you have to get the

50:51windows

50:51installer if you're a mac user you have to

50:54get a mac user and the concept of container

50:56ing allows

50:57you to essentially take an application or a

51:00system and decouple it from the operating

51:03system so

51:04what this essentially means is you can take

51:07let's say a let's let's take um adobe um

51:11photoshop as

51:12an example and you install adobe photoshop

51:14into a container now that container simply

51:17knows that

51:18there's an operating system called windows

51:20and the windows can do a bunch of other

51:21things

51:22when you do it that way what are you then

51:25able to do is uh use your computer in a

51:27much more flexible

51:28way because all the applications don't know

51:31other applications exist they only really

51:33ask for the

51:33resources they need when they need them and

51:35they sort of talk to the computer to get

51:37those resources

51:38and those those containers those boxes just

51:41have a simple understanding that they are

51:43sitting in a

51:43windows environment now the reality is it

51:46might not be the reality is is that

51:48actually in that

51:49box it has no clue of what's outside of

51:50that box but that box could actually be

51:52living in another

51:53put in a world another world in another

51:54place and so it makes these applications

51:56more portable it

51:57makes them flexible you can move these

51:59things around it's a bit more like a

52:00commodity if that

52:01makes sense and um and that i'm i'm hearing

52:04myself talking i know that's just a bad

52:07allergy for this

52:08but nonetheless what they seem to have done

52:10here is taken snow park now remember snow

52:12park is

52:13data engineering applications and ai and ml

52:16essentially all packaged into a product

52:19called

52:19snow park and they're taking that sort of

52:21capability and putting it inside of a

52:23container

52:23and so there's a lot of benefits that come

52:25with that but to be honest i'm going to

52:27need to sort of

52:28see use cases and examples to better

52:29understand the value of that and why that's

52:31important

52:32and i think you should also go and check

52:34out the session it's coming up next in this

52:36in this

52:36keynote so we'll obviously take a look at

52:38it but um i just wanted to sort of briefly

52:40sort of touch

52:41on that concept if it's new to you at all i

52:43mentioned blue yonder earlier they have a

52:46lot

52:46of legacy engines that they need to still

52:48need to be made available they can't

52:50rewrite them

52:51right they just want to host them inside

52:54snowflake use the database engine and just

52:56make the services

52:57available like they did before so this

52:59really helps us swoop up any function any

53:02application

53:03that already exists and say hey i want to

53:05run up in snowflake this is how you do it

53:07now it's not a

53:09coincidence that this is also going to be

53:11the strategy this is a large language model

53:14that's

53:15now disappearing you know into the

53:17container this is how we're going to host

53:19large language models

53:20and we're going to show you that in the

53:22segment coming up it's a great strategy

53:24because we don't

53:25have to port them we don't have to hack

53:26them you know they're going to go wholesale

53:28you know into

53:29the container and then they can be

53:31addressed you know by the by the

53:33applications themselves

53:35very very good way because we are

53:37anticipating very very rapid development in

53:40these areas

53:40and as a result you know we need to have a

53:43strategy to adopt that and make that

53:45available

53:46to the world yeah that makes sense so with

53:49that the messages you know of the

53:51announcement this

53:52morning's no limits no limits means we're

53:54pushing the limits back on applications

53:56pushing the limits

53:57back on the use of data pushing the limits

53:59back on snowflake i mean the whole strategy

54:01here is

54:02for the data cloud not to have any

54:04limitations and you have our word for it

54:06and we will continue

54:07to do that in any and all areas no more

54:09that we yield the stage and we'll thank you

54:11very much

54:12you've been a great audience thank you so

54:16that was pretty good that was i really

54:22enjoyed that um

54:24it was a framing of what snowflake is what

54:27it can do and where it's heading it's as

54:29simple as that

54:29and uh lots of lots of keynotes in this

54:32space i've just did a you know tableau

54:35keynote breakdown and

54:37in that keynote the message was really like

54:38this is where we've been this is what we've

54:40done and

54:41now this is where we're heading and so for

54:44this this feels to me like a keynote sort

54:46of intro

54:47that's presented a product that's really

54:49confident about what it's doing where it's

54:52heading and you

54:53know what it wants to take on in the future

54:54that's not to say it's right about all of

54:56its views and

54:57standpoints but they're at least internally

54:59confident about what they want to do with

55:01these

55:01things and how they're going to do it and

55:03if you disagree or don't agree then you

55:05know take take

55:06it for what it is you've always got choice

55:08you can go and use the competitor data

55:10breaks as a word i

55:11know everyone at snowflake will shudder

55:13when i mention it but absolutely all these

55:15other sort of

55:16takes on the same thing exists and you can

55:18absolutely go and go and check that out so

55:21i think that was really good and the other

55:24thing is i think as a as a as a

55:26presentation it was also

55:28a very good framing of how snowflake wants

55:30you to use the platform sometimes brands

55:32are a little

55:33scared of this but sometimes brands will

55:35just say i'll just use our use our platform

55:36to do whatever

55:37you want you know here you've got this

55:39problem yeah just do that yeah just do this

55:40just do this

55:41no here's snowflake we're very deliberate

55:43like that they're literally saying to you

55:44look yeah

55:45bring it bring your traditional warehouse

55:47and lake into our thing but don't just

55:49bring it in to do

55:49that bring it in to do apps bring it in to

55:52do ai machine learning it bring it in to do

55:54data

55:55engineering bring it in to do collaboration

55:57and by the way we're very opinionated about

55:59how we do

56:00these things it's kind of an extension of

56:02my previous point but they are they are

56:04very sort

56:04of confident about that they're telling you

56:06how to use the platform and that is in some

56:08cases

56:09actually quite rare you've got lots of

56:11incumbent companies today that um you know

56:13will be very

56:13scared of telling you how to use their

56:15platform because they have a vested

56:17interest and a customer

56:18huge customers that use their platforms in

56:20hundreds of different ways they don't want

56:22to tell any one

56:23of them that they're wrong and because as

56:24soon as that's the case you know you know

56:26it's time to

56:27switch so um yeah that's that's what i sort

56:29of took out of that but nonetheless i

56:32thought this

56:32was a very interesting session um it was

56:35pretty cool okay here we go snowflake co-

56:38founder and

56:39president of product the noir dash ville

56:41snowflake senior vice president and founder

56:44of niva

56:45stridar ramaswamy and snowflake director of

56:49ai ml engineering mona atarian so um if you

56:54're not

56:54familiar niva was the company that uh snow

56:57flake recently acquired so that's what this

56:59sort of

57:00just discussions about hello everyone it's

57:04so great to be here my name is mona atarian

57:09i'm the snowflakes director of ai machine

57:11learning this year has been absolutely

57:14amazing

57:15i feel like the world is now as excited as

57:17i've always been should be going for Paris

57:20next so

57:20before we dive into all of our product

57:22announcements we wanted to have a short

57:23conversation

57:24about our vision in this space we believe

57:27that data cloud is the foundation that

57:30organizations

57:31need for ai but what are we providing on

57:34top of that so that all of our customers

57:36can truly seize

57:37this opportunity what better way to find

57:40out than to hear from our founder benoit

57:43and our new spp

57:45shridhar benoit kick us off please yeah

57:48thank you mona thank you and hello everyone

57:51it's really

57:52great to be here at summit so just like you

57:55mona i am very excited about this topic and

57:58i want to

57:58talk about ai with really you all and i'm

58:02especially thrilled to have a shridhar here

58:05with us

58:06and shridhar you have been with us for how

58:09long already week four

58:11yeah so you're new to snowflake but shrid

58:15har is a veteran in ai and machine learning

58:19so

58:19tell us how you got here today thank you

58:23benoit mona i'm thrilled to be here i

58:26worked as the svp

58:27of ads and commerce at google and built the

58:29ads business there for over 15 years

58:32we built some of the so this is a you know

58:36this is a conversation now what what this

58:38really is is

58:40so i think it's just a quieter company um

58:42people will be asking you know what are

58:44they going to do

58:45with this company and you've got always got

58:47the headlines but what this really feels

58:49like and

58:49what's going on on stage here is really it

58:52's a credential ization piece right so shrid

58:55har is here

58:55explaining his background he came from

58:57google this is his path that should fill

58:59the market with a bit

59:00of confidence so don't forget that these

59:02keynotes go out to things like investors

59:04and so they'll be

59:05in the room they'll be watching online they

59:07'll be sort of thinking about these

59:08discussions they'll

59:09be writing articles and blogs about this

59:11whole thing to go out to other investors

59:12who might be

59:14looking at snowflake snowflake is on the

59:15stock market as well so it's a super

59:17important thing to

59:18do the other thing is is also contextual

59:20ization you know you buy a company it's very

59:23new he's

59:24only been there for four weeks so this is

59:26really the first opportunity for him to

59:27meet everyone in

59:29the snowflake community not just you know

59:31people who are customers but also partners

59:33a whole

59:34ecosystem of people who support snowflake

59:36work alongside snowflakes this is again

59:38another

59:38important opportunity to do that's why it's

59:41come right after the ceo that can be a more

59:43important

59:43placement for anything else that sort of

59:45hierarchy should tell you everything you

59:47need to know about

59:48how snowflake values this acquisition and

59:51they have the existing sort of uh you know

59:53person who

59:54looks after ai and ml on stage and then

59:56they've got this new vice president who's

59:58looking after

59:59this product neva that they acquired and it

60:02feels like a partnership that's going to

60:04sort of come

60:04into play and we should expect to see more

60:06of shrida in the future as well but anyway

60:08let's carry on

60:09and largest machine learning systems on the

60:13planet more recently i founded neva the

60:17world's first

60:18ai powered consumer search engine its snow

60:20flake acquired last month now we are all

60:23excited to be

60:24at snowflake and to bring the power of ai

60:26and data to all of our customers good stuff

60:30we are

60:30super excited to have you all right let's

60:33dive in but what is our vision in ai so yes

60:38as you know

60:39mona so over and again you're asking the

60:41one of the founders what is our vision for

60:44ai so like

60:44that's again another really important

60:46placement it's coming from the founder this

60:48this vision is

60:49coming from the founder like you know i'm

60:51sure they all know him i'm sure they've

60:53discussed it

60:53quite clearly internally and so any one of

60:55these three people could probably

60:56communicate it quite

60:57effectively but this is you know benoit in

61:00the middle flanked by his two ai uh you

61:03know master

61:04master people and uh they are going to uh

61:07you know sort of help push this company

61:10into the future and

61:11maybe this is something that benoit himself

61:13is super passionate about normally there's

61:14always

61:15one person the executive leadership team

61:17that sort of takes this thing on board or

61:19pushes these

61:19innovations because they passionately

61:22believe in them and so um it's good to kind

61:24of you just you

61:24just need this sort of opportunity this

61:27this photo here in itself um you know is is

61:29quite an important

61:30important piece of the whole thing two

61:33years we have made many many investments in

61:35ai and ml

61:36but whether it's really excited about this

61:40moment is that generative ai for the first

61:43time is really

61:43going to democratize access to data and

61:46that's critical and before for example you

61:49had to know

61:50how to program in sql or python you know to

61:53create data and only few people at this

61:56level of expertise

61:58so now all the sudden you know so before

62:01you know this expert had had to program you

62:04know dashboards

62:05or any other things for for for business

62:08users to be able to access the data so so

62:11there was always

62:12this friction now all the sudden you don't

62:15need to know how to program in sql or

62:18python now anyone can

62:20directly ask a question you can talk in

62:23natural language with your data and the

62:26generative ai

62:27layer can translate this natural language

62:30question into appropriate queries and even

62:33visualization

62:34i think benoit is reaching into the future

62:38a tiny bit so yes this is possible yes you

62:41can you know

62:42they're they're good examples of this and

62:44what is super interesting is that all the

62:46demos you've seen

62:47on social media all the demos you've seen

62:49at companies market those have all been the

62:51ones

62:52that have worked really really well and

62:53this is the social media effect this is

62:55sort of the

62:56the nature of how we communicate things if

62:57something goes bad you're not going to go

62:59out in

62:59the world and show look how badly gener

63:01ative ai handle this particular situation

63:03that just doesn't

63:04take off no one's interested in it no one's

63:06going to like or really like and reshare it

63:08however when

63:08it nails it when it when you do do that and

63:10it goes and gets the right response and

63:12gives you

63:12visualization and it does all this stuff

63:15that's just going to get amplified so much

63:16more even

63:17though in in real terms there are just as

63:19many failures as there are successes and so

63:22where i

63:22think benoit is reaching here a bit is that

63:25yes the technology can do that how far away

63:27from uh

63:28being able to reliably do that are we i

63:30think we're a little bit further away and

63:32so there's

63:33a lot that needs to happen between now and

63:35then if you look at the adoption curve of a

63:37i we are we

63:38are i think we're right before the the sort

63:40of the we're right before the peak at the

63:42first peak

63:43and what will happen is a second peak comes

63:45further down the line when things are more

63:47mature

63:48and if you look at companies that are

63:49leading this stuff you look at open ai if

63:51you look at google

63:52if you look at microsoft in all their

63:54sessions in all the way they talk about

63:56this capability

63:57and they frame in two very specific ways

63:59well number one they say that the the

64:01version of ai

64:02and generative capabilities we have today

64:05is not ready for prime time use they all

64:07say that

64:08unanimously number two they do say that it

64:10's still valuable putting it out there

64:12because there's a lot

64:14to learn from actually putting it in front

64:15of people to find out how they're being

64:17used so

64:17with every single thing i think you still

64:19have to go at this from the approach that

64:21you still need a

64:22skilled person to be able to assess how

64:24well the ai is doing and that is in itself

64:27the process of

64:28training in a very sort of long way right

64:30and as the person is using these queries

64:32and saying oh

64:33this one was good this one was bad i've had

64:35to redo this all again and that in itself

64:37is going

64:37to make these systems better so in the

64:39future we might get to this world that ben

64:41oit is painting

64:41but you know if i go into snowflake right

64:44now yes i could probably go to uh you know

64:46github

64:46copilot and get it to write me some nice

64:49sql that could do my very basic task but as

64:51soon as i'm

64:52working in an enterprise context and i'm

64:53trying to make sure that this code is

64:54actually going to

64:55do what it's going to do um yeah i better

64:57have the skills to be able to validate that

64:59for myself

65:00and when i do the documentation i better be

65:02able to get in front of someone and explain

65:04to them how

65:05it's working and be able to sort of go

65:07through all the levels of qa and any sort

65:09of issues that come

65:10up in in a very sort of intelligible way so

65:12i i don't think ai is quite there yet but

65:15we are at

65:15a keynote we're at a conference i get the

65:17sort of general gist of what benoit is

65:19saying is like today

65:21that is now a possibility right that's that

65:23's how i think it should be framed and it

65:25could be a

65:25language or cultural thing but nonetheless

65:27i think i think it's just important to just

65:29step back a

65:30little bit from the hype train and just

65:32make sure that we we contextualize the

65:33progress that's being

65:34made correctly and um we realize that yes

65:3780 of the time it's right but 20 of the

65:40time it's very

65:41very wrong and you still need people who

65:43know how to critique this work to be able

65:45to figure that

65:46out the other aspect is frank you know just

65:50explain how you know the data cloud brings

65:54all workloads to your data giving you the

65:57security governance performance and ease of

66:01use which is

66:02you know really critical and this applies

66:05also to ai you really want to run all your

66:08ai workloads

66:10in the data cloud so so really our vision

66:14is is three things first you know

66:16that statement is super interesting yes

66:20today you do want to do that because the

66:25computing requirements and the resources

66:27required needed to be able to do those

66:30things just don't

66:31exist you can't have a desktop that can run

66:33a large language model but i also think

66:36this is

66:36technology we're talking about i think in

66:38the future there will be use cases where

66:40actually

66:40you don't want to run that capability in

66:42the cloud because um you might you might

66:44have you might have

66:46a private cloud that does this you might

66:48have a specific um compute resources within

66:51an organization

66:52that's actually able to do this training

66:54and largely because you might be in a

66:55regulatory

66:56framework where you can't do everything in

66:58the cloud you might have to work in an air-

66:59gapped

67:00environment for example if you work in

67:02defense of security of course you run this

67:04stuff in the

67:05cloud you're going to have to run it in a

67:07much much more secure framework um and so

67:09um you you

67:11i think you also have to sort of take some

67:14of that with um you have to critique some

67:16of this stuff

67:17because yes of course you know your iphone

67:20does certain ai and ml operations on device

67:24and uh

67:24google will do certain ai operations on

67:26device and it will do the rest of them in

67:28the cloud there is

67:29always a like a framework for deciding what

67:31gets done where and i don't think every

67:33workload has to

67:33be done in the cloud some can be done on

67:36device some can be done maybe in

67:38partnership so a little

67:39bit of work on device a little bit of work

67:41in the cloud some might be done in this way

67:44where the

67:44model is trained in the cloud but once the

67:46model is small enough to fit on your device

67:48your device

67:48can actually run the model a good example

67:50is transcription when i transcribe this

67:52video the

67:54computer then trained the model to be able

67:56to take what i'm saying just literally into

67:58the microphone

67:59and translate it into a transcript was

68:01trained on a much more powerful computer

68:03that i'm going to use

68:04to actually run the model to generate the

68:06transcript and by the way it will do it

68:08faster

68:08than the video uh takes if that makes sense

68:11so an hour um transcript will take 40

68:13minutes to to

68:14create when i run it so um you don't have

68:17to run everything in the cloud but it's it

68:19's desirable

68:20for the reasons we explained before if your

68:22data is there you're not having to move it

68:24around there's

68:24no sort of big cost in doing that and i

68:26think that's a slightly better setup no

68:29user should

68:30have direct access to data through natural

68:33language second you will be able to run any

68:37model inside the snowflake cloud and

68:40optionally embed these models inside that

68:44application that

68:45then can be distributed through the entire

68:48data cloud and third all of this will run

68:52on your data

68:53without any security or government trade-

68:56offs so mona your team is really at the

68:59center of

69:00this development default right yes yes we

69:03are a bit of a bit of a like uh an

69:06interesting dynamic

69:07there so like found other company ask you

69:09hey so your team's on top of this right

69:11right yes

69:11of course like no it's not going to be the

69:13answer there is a broad effort across many

69:16teams to make

69:17this vision a reality from giving our

69:20customers a platform to develop and deploy

69:23models to easily

69:24accessing models automatically built on

69:27your data from sql to our own llms that

69:30understand documents

69:32and of course uh llm-powered products that

69:34make all of our customers more productive

69:37there's a

69:37lot to be excited about all right shridhar

69:40you've built machine learning so this is

69:43also weird

69:43because i think i think it just needs a

69:45host someone who's not one of these three

69:46people asking

69:47them the questions because then i think it

69:49's sort of strange to have a question asked

69:51someone answers

69:52then that person asks the next person the

69:54question it kind of it's difficult to

69:56follow and um it

69:58almost feels like this is a scripted

70:00discussion and it is of course it's a

70:02scripted discussion

70:04you ask this this is the answer this is how

70:06it goes and it doesn't play out like a

70:08conversation

70:08unfortunately in my in my opinion this

70:10early on anyway but nonetheless the the

70:13talking points

70:13are still really really important learning

70:16systems at incredible scale what about the

70:19vision that

70:19benoit just laid out made you believe that

70:21snowflake is the place to build the future

70:23in

70:23this space absolutely let's double click on

70:26what benoit said just now and what jensen

70:29also talked

70:30about yesterday which is language models as

70:33the new human computer interface remember

70:36that for

70:37the past 50 years we have had to live by

70:39rules that computers and programmers set

70:41for us if you

70:42entered a number in the wrong format well

70:45it's your problem all of a sudden we can

70:47interact with

70:49computers with applications in natural

70:51language and actually have them understand

70:54what we're

70:54saying this is a huge unlock but there's a

70:58big but before it goes into the bar nailed

71:01it 100

71:02percent all this technology is just an

71:04interface to me a lot of people put this

71:07stuff as oh this is

71:08gonna get rid of my job no no no this is

71:11just a new interface you're going to work

71:13with to get

71:14your job done and it will allow you to do

71:16more and do it more effectively it needs to

71:18be paired with

71:19a bunch of other innovations at the same

71:21time ai and large angle models on their own

71:23aren't going

71:24to be the answer and there's going to need

71:26to be other sort of contextual technologies

71:29that need

71:30to come online at the same time but those

71:31technologies are being built trust me and

71:33now

71:34the ai is out in the open companies are

71:36scrambling to get those solutions you know

71:38snowflake today

71:39you know coming out with all these

71:41capabilities now these these platforms need

71:43a place to run

71:44a place to work and so everyone including

71:46snowflake is sort of pushing to make those

71:48available

71:49language models by themselves do not

71:51understand what is real what is

71:54authoritative what is

71:55believable what is real time you need to

71:58combine them with the power of retrieval

72:01with the power

72:01of search to set the right context and this

72:04is what we did at neva to launch an ai

72:06powered

72:07searching it's really the combination now

72:10of course snowflake is the trusted platform

72:13that is

72:14safe secure efficient for data and

72:17applications it's the combination of these

72:20that's going to be

72:22the magic unlock for all of us so whether

72:24it's much better catalog search or an

72:27assistant that

72:29can help you write sql faster or an

72:31assistant that you can just talk to and get

72:33insights about you

72:34know your data your customers you can

72:36expect a lot from us we're very excited and

72:39that is an

72:40interesting thing like you've got this

72:43capability of um you know neva with

72:47applications but actually

72:48there's probably a lot of work that neva

72:51have done to uh the snowflake can take just

72:53to make its own

72:54interface a little bit work work a little

72:57bit better can maybe surface um

72:59relationships between

73:00different databases different tables

73:02different columns all of that metadata

73:04sitting inside of

73:05snowflake there's probably a lot that neva

73:07could probably add in terms of their

73:08understanding of

73:09ai to that technology and so i always think

73:12when a company gets acquired the obvious

73:14innovation is

73:15obviously you know table stakes but to me

73:17there's also other hidden benefits like

73:18things that that

73:19company is really good at that you can port

73:22to your company to do new things yeah yeah

73:25i agree

73:26thank you shrida and and indeed you should

73:28expect a lot from us and we have already

73:31built a lot and

73:32you are going to be really really amazed by

73:35what creation is going to walk you through

73:38and

73:39especially my favorite part of summits the

73:42summit keynotes which are live demos and

73:44you are going to

73:45see many of them let's do it thank you so

73:48much thank you manuel shirdar thank you

73:50everybody

73:51that was a that was a that was a really

73:53interesting section um really short it's

73:56clear that the

73:57acquisition is so young that they've got

73:59nothing to show just yet the integration is

74:02happening the

74:02work is going on behind the scenes next

74:04year or summer you'll probably see stuff uh

74:07coming out

74:07and because of the way i think you heard

74:10frank say this um stuff has to go into

74:12preview and so on and

74:13so forth so it might be maybe a year or two

74:15before we really see the innovation coming

74:17from neva but

74:18nonetheless and that was a good way of just

74:20showcasing look here's what we're going to

74:22work

74:22on here's what's going to happen i think it

74:25could have been shorter but nonetheless um

74:28i think it

74:28was important to see these three people on

74:30stage talking about their vision for ai and

74:33where they

74:33want things to go

74:38cool change of stage welcome snowflake

74:45senior vice president of product christian

74:50kleinerman

74:52good morning good morning snuffling summit

74:57good morning

75:02you needed the energy to see you all here

75:04thank you for being here with us today

75:07it's amazing and i can assure you all of us

75:09at snowflake are committed to giving you an

75:12amazing experience at the conference

75:14hopefully all of you go back to your

75:16organizations

75:17excited and inspired about what is possible

75:21with our new innovations good you heard

75:24from frank

75:25the broad framing of the data cloud you

75:28just heard from steedar benoit and mona

75:31about ai and how we're

75:32thinking about it and now we're ready to go

75:35one level deeper on actual innovations

75:38actual new

75:39launches now i'm going to say something

75:43controversial here as a tableau as a you

75:47know

75:47someone who's been using tablet for a while

75:49this this this this would have been the

75:51dream keynote

75:52for tableau right here's what the platform

75:55is here's where we're heading and now here

75:57's the

75:58really good stuff here's the features

75:59features for the analyst features for the

76:01core audience

76:02and i'm kind of jealous i'm sitting here

76:04with envy because it's it's you know table

76:06au

76:07made it pretty abundantly clear they're

76:09focusing on the user base they've not

76:12historically focused

76:13on before and in order to do that they have

76:14to make some sacrifices that's not to say

76:16that

76:16you know data analysts today aren't going

76:18to get features for use in desktop and in

76:20altering it's

76:21just to say the focus of the company is

76:23heading in a different way and so what i

76:24get the sense of this

76:25is only my first summit what i get the

76:27sense of from this at least structure of

76:29this keynote is

76:31that if you went to this keynote as someone

76:33who's been using snowflake for a while i

76:36think this

76:36structure absolutely speaks to you and it

76:38speaks to sort of what you're trying to

76:39achieve so i

76:40think it's a really really nice sort of

76:43framing you want to see some demos yeah not

76:47not really

76:49how many demos how many you want ten you're

76:54crazy

76:55yeah we have some some announcements and

76:59demos and we we love showcasing the actual

77:02technology

77:03that's good so with no further ado let's

77:06get into it absolutely today's

77:08announcements we've included

77:11three different chapters on the talk first

77:16one is around single platform second one is

77:20around

77:20distributed deploying and monetizing

77:23applications and the last one is how do we

77:26help you program

77:27data get value out of your data without

77:30trade-off so with that let's see let's get

77:33started on the

77:33single platform single platform and if you

77:36have been with us for a long time in the

77:39very beginning

77:40are found i think that diagram is actually

77:41a diagram for the single platform they want

77:45to

77:45create internal tables here they are on

77:47stage where you ingest data into snowflake

77:49and you are

77:50able to get value out of it but over the

77:53years many of you gave us feedback and you

77:57said i want

77:58to be able to interact with data that is in

78:01external storage here at snowflake summit

78:05five

78:05years ago we introduced external tables

78:07later on we introduced external tables for

78:10iceberg tables

78:11and last year we introduced iceberg tables

78:16and many of you tried it loved it but you

78:20also said

78:20i want sometimes to have snowflake control

78:23the reads and writes and the transactional

78:26consistent of the tables sometimes i do not

78:29want that and today there are trade-offs in

78:32performance

78:33if it's an external table you don't get the

78:35best performance if it's a iceberg table

78:37you do so

78:38today we have an exciting announcement for

78:41for all of you which is we do not like to

78:44make you

78:45have these trade-offs so we're announcing

78:49unified iceberg tables as a single mode to

78:56interact with

78:56external data interesting and the good news

78:59is it's going to have two modes one of it

79:02is in an

79:02unmanaged mode where coordination of

79:05changes and rights happens by a different

79:08system your choice

79:09you can have a managed mode where we take

79:12control of the data and we coordinate

79:14rights

79:15but the most important thing is we do not

79:18want to give you any trade-offs in terms of

79:22performance

79:22interesting and what you see here first

79:26blue bar it's an unmanaged table i'll let

79:30you figure out

79:30who wrote those parquet files they're not

79:32very optimized but it's still more than two

79:36times faster

79:37than external tables and if you use a

79:40managed tables where snowflake is doing the

79:42rights

79:42the performance is on par with internal

79:45lighting interesting interesting i'm not

79:51hearing i'm not

79:52hearing applause in the crowd so it's one

79:55of these things where i can't tell if that

79:57is because people

79:58don't understand it or if it's or if it's

80:01just it's just the crowd right like

80:03sometimes you get

80:04a tough crowd and you know it's good but

80:06they're not like the kind of people that go

80:08whoo you know

80:09all that jazz so i have no idea how to read

80:11the room and the reaction um i also have no

80:14idea how

80:14to read that sort of announcement to be

80:16brutally honest i don't think i understand

80:18enough about it

80:18to be you know be like oh my word this is

80:20going to be amazing so um i think i want to

80:23sort of see

80:24a demo see the benefits um but it's

80:25interesting they have the manager and unman

80:28aged setup

80:29i think whenever a company does that that

80:31really shows you they understand their

80:33customer audience

80:34because um you know it's very easy for a

80:36company just to say oh we're always going

80:38to manage it for

80:38you this is how it works come along with

80:40the journey and you know tabular cloud is a

80:42bit like

80:42that actually a tabular cloud is like this

80:44is the way it's working this is how you're

80:45going to do it

80:46this is this is this is you get what you

80:48get right and it feels like snowflake has i

80:51think maybe

80:52because it's slightly smaller than tableau

80:54um in terms of like customer base it has to

80:57has to bring

80:57customers with them so new features have to

80:59have this sort of world where there's no

81:01trade-offs for

81:02them to really get any sort of adoption so

81:05um interesting to see that of our platform

81:07we added

81:08support two years ago and one of the things

81:11that we heard very consistently for many of

81:13you is

81:14how do i more easily get value out of my un

81:17structured data and frank already alluded to

81:21before any of us were talking about llms i

81:23assume you heard about a limbs yeah me too

81:27before they were in vogue we acquired appl

81:33ica and what we're very very excited to

81:35announce today

81:36is the private preview of document ai which

81:41what it lets you do is ask questions in

81:47natural language

81:48from documents that you have stored in snow

81:50flake and most important i don't know that

81:54anyone has

81:54operationalized an end-to-end pipeline

81:57where you can give feedback to the answers

82:00you're getting

82:01from the ai and be able to fine-tune and

82:04retrain the model okay what you'll see is

82:07the opportunity

82:08to take fields and structure out of

82:11documents and then you can put in a table

82:13use it for another

82:15pipeline use it for ai for a ml anything

82:18else you want the model that powers this is

82:21a text and

82:22image model and we want to show you how

82:24cool this works yeah so with this we're

82:26going to jump into

82:27the first demo of the morning let's have a

82:29look let's have a look i'm not going to go

82:30wild until

82:31i've seen it first introduced so dash he's

82:33like our demo master come here show us

82:36wearing snow

82:38glasses glasses you might as well be

82:41wearing an apple vision pro when i tell you

82:44that we are

82:45look at the shoes and now i want to

82:50introduce polita palos engineering leader

82:53at snowflake

82:53and she's gonna do random merch insert

83:04thanks christian and good morning at snow

83:08flake we believe it should be really easy to

83:11use ai

83:12to get more out of your data right and so

83:15that's why i am really excited to be with

83:18you here today

83:19to talk to you about how snowflake is using

83:22large language models to put your unstruct

83:25ured data

83:26to work let's check it out with a demo for

83:30this demo dash and i dash you're looking

83:34great

83:35are going to play the role of a data

83:37engineer at ski gear co a company that

83:40produces ski goggles

83:41we want to make sure that we're able to

83:44manufacture those goggles on time with our

83:48expected

83:49volumes but recently we've been having some

83:51problems with our injection molding machine

83:54let's check it out interesting to

83:57understand these issues i'm going to use

84:00snowflake to analyze

84:01inspection forms we're going to be able to

84:04see a full history i'm already happy

84:05problem exactly

84:07what i said earlier on this this is a this

84:09is a typical business today inspection

84:11management

84:12happening in a form manually signed maybe

84:14typed up and then signed off and the text

84:19is written now

84:20if this company has gone through digital

84:22transformation they might go over to ipads

84:25they might go over to a whole bunch of

84:26different things of putting this stuff in

84:28but in earnest

84:29they're still always going to be parts of a

84:31company that work in this way because that

84:33is just the

84:33fastest way to work if you think about it

84:36and the way factories are set up any small

84:38change can you

84:39know throw a whole production lines a ski

84:42so um this is an interesting innovation

84:45because what i

84:46think we're about to see in the demo and i

84:48'll shut up in a second we'll actually watch

84:49it

84:49is that it's essentially saying to you look

84:52if you have this workflow if you have this

84:54information in

84:55a document we're going to be able to pull

84:57it out so what i think they're about to

84:59show us is um

85:00they're going to be able to pull out this

85:02pertinent information about this inspection

85:04from this document which has been scanned

85:06in and we're going to be able to query it

85:08uh with

85:09snowflake using a standardized interface it

85:11's going to be absolutely beautiful let's

85:13have a

85:14look to the injection molder when they

85:17occurred and why to start with i've created

85:20a project

85:21called machine inspections and uploaded

85:24about a year of our pdf inspection

85:27documents this is

85:29built directly into the native snowflake ui

85:33and i can upload some more if i need to but

85:37i think

85:37i've got enough to get started so these

85:41documents contain a mix of fields and free

85:45text and analyzing

85:46them is either going to be error prone and

85:50time consuming or it's going to require ml

85:54expertise

85:54which i don't have but with snowflake's new

85:58document ai currently in preview i can do

86:02this

86:02with no ml expertise required by default

86:06snowflake's document ai uses a zero shot

86:10model which

86:11just means i can get good results without

86:14having to fine tune or train the model but

86:16if i need to

86:18i can always fine tune it to improve my

86:21results okay so on the right you can see a

86:23preview of one

86:24of the documents that i've uploaded to

86:27start extracting information i can just

86:29type questions

86:31in plain english so you can see i've

86:33already created a few of them like when did

86:35the inspection

86:36happen who performed the inspection and you

86:40can notice here that document ai is

86:43actually reading

86:43that signature it's reading the signature

86:46and what was the inspection grade dash let

86:48's add one more

86:49question i'd like to know what part was

86:52defective and when i ask this question snow

86:56flake's document

86:57ai is going to do an analysis of this

87:00document and give me back an answer along

87:03with a confidence

87:04score in this case it told me that the

87:06injection molder has the problem okay let's

87:10flip through a

87:11few more of these documents dash to see how

87:14it's performing okay what you see right

87:19here in real

87:20time is what what they did there is you you

87:22're kind of training it on the left and then

87:25as you

87:25switch through documents it's like you're

87:27going through different rows in a in a

87:29database so each

87:29row is a record and each record is

87:31essentially just a document and in this

87:33particular table

87:34you can see this is a pass everything is

87:37good but what is great here is you're

87:39building a question

87:40it's taking your questions understanding

87:43what you want to get it's then analyzing

87:46this document

87:47reading through all the text passing out

87:49the information that things it needs to

87:51pass

87:51finding what it thinks you're talking about

87:54extracting that information using obstacle

87:59character recognition and then giving you

88:01an answer and then on top of that giving

88:03you a

88:04confidence score saying hey look i think

88:06this is the i'm 75 sure this is this is

88:09what it is

88:09i think it's pretty good that's pretty

88:11powerful i mean like i've seen tools try

88:14and do this and

88:15fail epically i'd love to see this on a

88:17although this document is quite structured

88:19i'd love to see

88:20this on like a handwriting only document

88:22right like maybe you've taken notes from a

88:24meeting and

88:25is it able to you know scrape the same

88:27information out of something like that does

88:29it fall over if

88:30you don't have like a neat structure like

88:31this format form does it'd be really

88:33interesting to

88:34see all those details but of course it's in

88:36preview for a reason um so yeah let's let's

88:38keep going

88:39snowflake analyzing this document with its

88:42large language model and we got an answer

88:45the mold

88:45clamping unit but this is interesting

88:47because you can see that the inspection

88:49actually passed

88:51with no issues found so this is no problem

88:54we're just going to update this and tell it

88:56none

88:57and what this does is it provides feedback

89:01to snowflake about how we're going to fine-

89:04tune

89:04the model and now i can click start

89:07training and publish the model to my

89:10account and in the

89:12interest of time i've already published

89:14this to my account and now that it's in my

89:16account

89:17i can actually share this with other teams

89:20that can use my fine-tuned model as well

89:22and then with a simple sql query i can run

89:27this new model on all of my inspection

89:30documents in one go

89:32so um for those of you who are not exactly

89:36sql experts um essentially it's basically

89:39specifying

89:39which database to use which schema

89:41essentially which part of the database um

89:43it's creating a

89:45table and then it's putting some

89:48information um from from the model into

89:51that table that's what

89:53the first sort of nine lines are doing and

89:56then from line 11 onwards we're selecting

89:58specific

89:59columns from that model and those columns

90:01are the ones that we just saw so the

90:03inspection date

90:04inspection grade inspector's name and the

90:06defective parts it's essentially selecting

90:08those

90:09columns out of the data set and then here

90:11you're basically saying from model results

90:13model result

90:14is going to be um if i get my annotator out

90:17so uh model results is essentially this

90:20table that

90:21was created up here and uh that's

90:23essentially essentially it um and it's

90:25ordering it by

90:26inspection date so basically deciding the

90:28order and then you get a table at the

90:30bottom and then

90:31you've got the uh responses the prediction

90:34as a json so it's almost storing it as a

90:36semi-structured

90:37data so um this is this is quite

90:38interesting quite powerful i really like

90:40this actually it's really

90:42it's really intuitive actually it doesn't

90:44it doesn't feel like broken i mean sql is

90:46far from

90:47intuitive but for a new feature uh someone

90:50who has to work with sql would immediately

90:53know how

90:53to utilize this so it's pretty good got

90:57some results and we can see that every

91:00three months

91:01the injection molding machine has failed

91:04inspection so this is really interesting i

91:07can share this

91:07insight with my quality engineering team

91:10and they can use this to run maintenance

91:13every two to three

91:13months so we can stop the time i've already

91:17I just want to go back a second what i don

91:19't see here is

91:22you know when you get a table like this it

91:24's very easy to jump to the conclusion this

91:26this is final

91:27date this is like data but because it's a

91:29prediction i think the bit that i'd love to

91:31see

91:32more of is like you can see the um score

91:35the confidence score essentially and how

91:38confident it

91:38is i would love to see some sort of

91:41aggregate confidence across this whole data

91:44set so right

91:45you know um how many false positive how

91:48many false negatives how many you know all

91:51of these things

91:52that are going on like what is the makeup

91:54of that and as people report these issues

91:56over time

91:58obviously this model will change in the

92:00future so if i base my analysis on this and

92:02then you know

92:02a week a week further down the line emily's

92:05gone in and gone looked at the data and

92:08said actually

92:08no this one did pass and it's just you know

92:11it's it's the ar model has read it wrong

92:13and i'm basing

92:14a report of this particular table i'm

92:16hoping to i'm hoping to get a sort of

92:18expectation that this

92:20has changed maybe some alerting all of that

92:22good stuff that comes after this particular

92:25journey i

92:25think is super important so um the insight

92:28is is good but then there's always that

92:30little element

92:31of polish on top of that just to make it as

92:33something that i can do once and forget and

92:36know

92:36that it's going to tell me if it finds a

92:37particular issue especially when these

92:39things are constantly

92:40changing molding machine has failed

92:42inspection so this is really interesting i

92:45can share this

92:46insight with my quality engineering team

92:49and they can use this to run maintenance

92:51every two to three

92:52months so we can stop failing these

92:55inspections and other data engineers

92:58analysts and developers

93:01at the company can use my model directly as

93:03well so that's really cool very good and

93:07now i want to

93:08be able to use this to process new document

93:11inspections as they come in well this is

93:14really

93:14easy because it's fully integrated into the

93:17snowflake platform so i can create a

93:20pipeline

93:21using streams and tasks to process these

93:24new documents and every time a new document

93:27comes

93:28in it will run and i've actually set up an

93:31alert to send me an email anytime a new

93:33document fails

93:35okay so we can even use document ai on text

93:39heavy documents like warranties and

93:42contracts and we can

93:45see here that the injection molding machine

93:49's warranty says that this part is actually

93:52good

93:53until november 2023 this is crazy so we

93:55might actually get some money back to recap

93:59what we just

94:00did we analyzed pdf documents using snow

94:03flake's first party large language model and

94:06then we were

94:07able to extract information with regular

94:10english questions and a few clicks rather

94:14than writing

94:15code and then i was able to fine-tune the

94:18model to improve my results and then i

94:21could publish

94:22that model for anybody in my organization

94:24to use and then i created a pipeline using

94:27streams and

94:27tasks to process new documents as they come

94:31in and send me an email when something

94:33fails that's very

94:35good join me for the what's new document ai

94:38and unstructured data with snowflake

94:40session to learn

94:41more thank you dash christian back to you

94:45that's pretty cool i love that that was

94:48really nice it's

94:49a simple problem simple solution is it cool

94:52very cool yeah yeah from the early days of

94:57snowflake

94:58governance was a priority and you saw it

95:00from frank governance is a big reason to

95:03have data in

95:05inside of snowflake security was where our

95:08founders started on day one but over the

95:11last few years

95:12we've been investing heavily on privacy

95:15which is another aspect of governance we

95:18talked about this

95:19actually in there separate not only because

95:21there's a lot of regulation around privacy

95:23but in this age of ai and gen ai you feed

95:27some pii to one of these models and god

95:31help you when that

95:32pii is going to show up who knows in which

95:34context and from that perspective we want

95:37to give you the

95:37most comprehensive platform to secure and

95:41protect your data all the way from

95:44classification of the

95:46data what is sensitive what is quasi-

95:48sensitive masking of the data we have

95:50ability to do private

95:51data products private machine learning and

95:54of course be able to audit the entire

95:56process

95:58and we have three exciting announcements

96:00for you today first one we're introducing

96:03in prior preview

96:04what we call query constraints which is a

96:07policy that you can set on a data set and

96:10say what types

96:11of queries are allowed to run maybe i don't

96:14want some columns to be selectable which is

96:16a projection

96:17constraint or maybe i don't want some

96:20columns to be a queried without an

96:22aggregate very exciting

96:25we're also introducing and integrating to

96:27snowflake differential privacy wow that's

96:29pretty

96:29cool you can introduce noise and track a

96:32privacy budget again to protect re-ident

96:35ification of

96:36sensitive data yeah we'll start with python

96:38the first time i heard of differential

96:40privacy was

96:41apple when they were um talking about the

96:44way their keyboard learns now this failed

96:46epically

96:48because the problem what would happen is

96:51that um you get a word that would briefly

96:55enter sort

96:55of everyone's vocabulary and it would make

96:58it through the differential privacy system

97:00and then it would become an autocorrect and

97:02this autocorrect could only really get

97:04fixed with a

97:05software update essentially someone

97:07actually manually taking it out and fixing

97:09it um so the

97:09way it works is essentially um let's say

97:11you're collecting data so you collect data

97:13from 100

97:14people what you then do is you add a bit of

97:16noise to that so you can't tell which

97:19hundred people

97:20created that data those data points and

97:23then uh when you're then analyzing this you

97:27've not really

97:27changed the makeup of the data but you've

97:30you've removed the ability to trace who

97:32exactly said

97:33something or who exactly uh provided a data

97:35point that's that's that's like a broad

97:37brush example of

97:38how it works and so the ability to do this

97:40here is actually quite nice because um the

97:43way this

97:43would work is you don't want to change in

97:45the source data it's kind of has to be a

97:47layer on

97:47top of your data when specific people are

97:49actually accessing it or querying it and

97:52who don't maybe

97:52have the right levels or right roles to be

97:54able to sort of work with that data in that

97:56way so that's

97:57uh that's a nice touch we'll do it into the

98:00core sql engine and last but not least we

98:05continue to

98:06invest in our clean room capability not

98:08only making the platform better there is

98:11some ui and

98:11we're working with amazing partners to

98:14deliver industry-leading privacy multi-

98:17party computation

98:18now i'm going to shift to the core of the

98:22engine right and to do so to to share some

98:26of our

98:26innovations i want to invite one of our

98:29founding engineers allison lee please join

98:31me in welcoming

98:32allison allison it's great to see you all

98:42it's great to see all the excitement for

98:44snowflake

98:44one of the things that i love about being

98:47an engineer at snowflake is how focused we

98:50are on

98:51making our customers lives easier this has

98:53been true from day one and it's a key

98:55reason why

98:56snowflake is a single product with one core

99:00engine so whether you're using sql or snow

99:03park it's all

99:04powered by the same engine so every

99:06enhancement that we make is applied across

99:09the board

99:10of course this makes it critical that the

99:12engineering team focuses on the most high

99:16impact areas of development and this is why

99:18we have a data-driven approach to

99:20engineering

99:22since snowflake is a single product we can

99:25easily analyze how snowflake is being used

99:27and figure out the best things that we can

99:30work on for you guys without you having to

99:32tell us what we

99:33should be focused on and then once our

99:35engineering team makes an improvement that

99:38is actually a

99:39pretty important point um a lot of products

99:41let's say take oracle you take any of these

99:44solutions

99:45they're actually a family of products each

99:47with their own code bases each with their

99:48own use cases

99:49because snowflake lives in the cloud it's

99:51one code base everyone's using the same

99:53version everyone's

99:54getting the same benefit and because you

99:56have that critical mass of people um you

99:59you get better

100:00telemetry out of the platform and therefore

100:02you're actually able to focus your efforts

100:04and energy on

100:04things that people are actually using

100:06versus things that you think they should be

100:08using

100:08but also you learn from quirks you learn

100:11from you know customers who are at the

100:13bleeding edge

100:14and they all surface new challenges snowfl

100:16akes will fix those challenges and then

100:18people who

100:19come later down the line don't even

100:20experience those challenges because someone

100:22else has already

100:23sort of set the trail as it were to borrow

100:26a phrase from south force um and uh kind of

100:29put

100:29people in the right direction so um i think

100:31it's a very important point easy to skip

100:33through in the

100:33keynote but i think it's actually very

100:35important sort of thing to call out it it

100:38just shows up

100:38there's no need to enable a new feature or

100:41tweak some parameter to get the best out of

100:44snowflake

100:45we want it to feel like magic to you and

100:47just work simple as that of course our

100:52performance work is

100:52never done and this data-driven approach to

100:55engineering works particularly well for

100:58performance work we're constantly looking

101:00at your experience with performance and

101:02snowflake

101:03and figuring out how we can improve it and

101:05assessing the resulting impact i'll be

101:08talking

101:09more about how we're assessing the impact a

101:11little bit later but first i wanted to tell

101:13you more about

101:14some of the new advanced analytic

101:15capabilities that we've been working on for

101:18all of you

101:18since you last heard from me we're

101:21continuing to expand snowflake's robust

101:24support for geospatial

101:25data that's pretty good as part of that

101:28geometry support is now generally available

101:32nice this means

101:33you can ingest any type of spatial vector

101:35object in snowflake and do your analysis on

101:38it whether

101:39you're operating on a spherical or flat

101:42surface additionally now in public preview

101:45you can easily

101:46switch between different spatial systems

101:48for instance if you're switching between a

101:50state-level

101:50system and a global system and now

101:53generally available we say like that to

101:57break that down

101:58geometries are types of spatial objects

102:01essentially and regional versus global

102:05systems what they're

102:06talking about is sometimes when you work in

102:08geospatial analysis you can have what are

102:11known

102:11as localized versions of of mapping if that

102:15makes sense so a global system for example

102:19the one that

102:20everyone's mostly familiar with is latitude

102:22and longitude right there's basically a

102:24bunch of

102:25coordinates around the world now if you

102:27live in the uk you might have something

102:29different called

102:30eastings and norden which is a different

102:31sort of grid-based system here in the uk if

102:34you go to

102:34america different region different systems

102:37sort of pop up depending on how that works

102:40and so

102:40you do need a central understanding of all

102:42of these systems to be able to map them on

102:44top of

102:44each other they all kind of have to be

102:46translated in order to work together

102:48because sometimes you

102:49get data from one place and data from

102:50another place and they're working with

102:52different systems

102:53so being able to support those systems in

102:56one context is super important for use

102:58cases like

102:59intersecting coordinates and shapes and

103:01invalid shapes but when you take all of

103:03these things

103:04together it means that it's much easier to

103:06migrate your spatial data into snowflake

103:08whatever that data looks like another area

103:11that we're focused on is giving you the

103:15ability to work

103:15with ml models directly from sql now in

103:18public preview we have a set of ml powered

103:21functions

103:22that allow you to build more reliable time

103:25series forecast quickly identify what's

103:28contributing to

103:29a change in a metric and detect anomalies

103:31and trigger alerts and the coolest part of

103:34all of

103:35this is that you can do that without any

103:36machine learning expertise in addition to

103:41expanding our

103:42advanced analytic use cases the team's been

103:45hard at work on something which for me as

103:47an engineer

103:48is the most exciting thing and that's

103:50making snowflake faster for all of you

103:53this is a lot of stuff we deliver

103:54performance related enhancements with

103:57nearly every release

103:59you might not be familiar with all of these

104:00on screen and you shouldn't have to be that

104:03's part

104:03of the simplicity of snowflake a lot of

104:05these are pretty geeky but each of these

104:07had an impact on

104:08the customer's performance experience

104:12however it can be challenging to understand

104:16the overall

104:17performance impact when you take all of

104:20these things together and this is why we've

104:22developed

104:23the snowflake performance index or spi this

104:27allows us to assess the impact of all of

104:30the

104:30great performance enhancements that we make

104:33across the entire year since we first

104:37started tracking

104:38the index back last august through the end

104:40of april so that's about eight months we

104:44found that

104:44query duration for stable workloads in snow

104:49flake has improved by 15 percent by stable i

104:53mean

104:54recurring workloads that are consistent and

104:56can be compared over time but what's most

104:58important

104:59is that this is all based on actual

105:01customer usage so this is based on your

105:04workloads right

105:05some of you are probably familiar with

105:07industry benchmarks such as tpcds these are

105:12commonly used

105:12to analyze performance for certain types of

105:15use cases and they they definitely have

105:18their uses we

105:19we use them as part of our development

105:21process but when it comes to assessing the

105:24impact of the work

105:25that we've done and asking the question

105:27have we made our customers lives easier

105:29these benchmarks

105:31aren't tied to any actual customer usage

105:33and so they really don't cut it and this is

105:35why we've

105:35developed the snowflake performance index

105:38because what's most interesting to me and

105:40our engineering

105:41team is analyzing how you our customers are

105:44using snowflake in the real world and

105:46making your

105:48workloads and your queries faster so i've

105:51talked a bunch about performance and query

105:54duration but

105:55don't forget that when performance improves

105:59in snowflake that's closely tied to cost

106:02and that

106:03means that when we make performance

106:05optimizations in the system your costs can

106:08go down it's free

106:09it's free money as frank would say all

106:11right well with that in mind i'm going to

106:13hand it back to

106:14christian and he's going to talk a bit

106:16about the work that we've been doing to

106:17help you with cost

106:18predictability and control so that was a

106:21tough section because it's one of the bank

106:24less things

106:24that goes on you don't get a gold star for

106:27doing it and when you make it work things

106:31go faster

106:32most of the time it goes under the radar

106:33you don't even see it so you have to be

106:36really paying

106:36attention to notice some of some of these

106:39performance improvements and because snow

106:41flake

106:42maps quite closely to cost the more time

106:44you're spending doing something probably

106:47the higher cost

106:48the more money you're spending with snow

106:50flake and so what is interesting is i'm not

106:53too f-a with the

106:54sort of community on this one but i think

106:56there's probably an opportunity where it's

106:58possible to see

106:59the impact of performance improvements from

107:02release to release just by comparing

107:04workloads

107:04that broadly haven't changed between

107:06releases so if you're seeing a workload

107:08that was taking x amount

107:09of time and then after an update it takes

107:11this amount of time but between those two

107:13nothing has

107:14changed you haven't changed your ware

107:16housing you haven't changed the um the setup

107:18of how your query

107:19runs maybe the exact same number of rows

107:21are coming through it there you may be able

107:24to see

107:24some more subtle changes but it's not

107:26really something you can demo nicely

107:29because what are

107:29you going to do show someone a query

107:31running faster and we're talking about

107:32milliseconds

107:33but those milliseconds add up if you think

107:35about the scale that snowflake's operating

107:37so

107:37it's it's a super tough gig i always say

107:39when you talk about performance

107:41improvements

107:42you get this in consulting as well um you

107:44know then you don't really see them you

107:46just you just

107:47experience them and sometimes even that

107:49experience doesn't really sort of aggregate

107:51up to this like

107:52wow moment if that makes sense thanks thank

107:56you

108:02lots of enhancements and we work very hard

108:04to make it transparent for all of you

108:06so things get better without effort if you

108:09look back at the old days planning for an

108:11upgrade

108:11that was crazy talk in this day and age all

108:15ison rightly said that we are focused very

108:20much on

108:21helping you govern and manage your spend on

108:23snowflake none of us at snowflake wants you

108:27to overspend we want you to be aligned with

108:30the value you get from how you're consuming

108:33resources

108:34in snowflake and the framework by which we

108:36enable this we want to give you visibility

108:38into how

108:39resources are being consumed we want to

108:41give you control and policies to manage

108:44that resource

108:45utilization and we want to give you

108:47optimizations and allison just covered many

108:50of those

108:52but in reverse order let's talk about

108:55control and today i'm very very excited to

108:58announce that

109:00the core budgeting capability is now going

109:02into public preview which will let you

109:06specify for a

109:07subset of resources of your choosing a

109:10budget that you want to track against and

109:13not only get an alert

109:14when you've exceeded the threat the budget

109:16because that's sometimes or oftentimes too

109:18late but also

109:20be able to know when you're on track to

109:22exceed that threshold nice so there's in

109:25public preview

109:26now and in the topic of visibility i think

109:30there's only one item that we've heard very

109:34very consistent

109:35feedback from you and it is the ability to

109:39have a warehouse utilization metric nice

109:44what this lets

109:44you do is specify how utilized is a

109:47warehouse a cluster in a given point in

109:49time which will help

109:51you optimize i hear questions from you all

109:53the time should i have a larger warehouse

109:56should i

109:56have a smaller one can i consolidate this

109:58is the building block that lets you see all

110:00of this okay

110:02very exciting nice you just heard many new

110:05capabilities and enhancements that we've

110:08done

110:09the core platform but for us the real

110:12reward the real benefit is when we hear

110:16from our customers

110:18getting that value from the innovations

110:21that we do and we want to invite one of our

110:23partners and

110:25customers onto the stage please join me in

110:27welcoming mihir shah he is from fidelity

110:30investments

110:31are you going to hear from him a little

110:33more mihir welcome okay so i'm going to

110:34skip this bit because

110:36um whenever i've had this experience in the

110:38past with other conferences but in essence

110:40whenever a

110:41customer comes on stage they agree to this

110:43setup here with snowflake but they don't

110:45necessarily

110:46agree to being critiqued by people like me

110:49so um having learnt my lesson having got a

110:52slap on the

110:52wrist in the past i'm going to skip this

110:54section and go to the to the bit after this

110:56so i'm just

110:57going to fast forward uh passes if you want

110:59to watch this go watch the keynote i won't

111:00really be

111:01commenting on this

111:12okay so this section is just about to

111:19finish and we're going to carry on there

111:25thank you very much

111:27thank you very much

111:28thank you so thank you so much to to me

111:33here that's warms our hearts to see the

111:39impact that

111:42we can have an organization tearing down

111:44silos and consolidating data so this is the

111:47end of

111:48chapter one i can move to part two of our

178:08conversation today wow and it's all about

111:55how do we deploy distribute distribute and

111:58monetize data products maybe a data set can

112:02be a native app as you heard and the face

112:05of all of this is our snowflake marketplace

112:09and the momentum that we have is is is

112:11amazing and we continue to to innovate

112:14lots of launches and you'll hear more

112:16throughout the conference the ability to

112:19have

112:19public listings and private listings the

112:22ability to automatically fulfill products

112:25across regions

112:26and across clouds that's going generally

112:29available at the conference but at the end

112:32of the day

112:32what really makes a marketplace like this

112:36interesting is the content itself

112:40and we have some exciting news to share

112:43with you today and to do that i'm going to

112:47invoke the help

112:48of ai in person and by ai i of course i

112:53mean alex isidorjek alex welcome to the

112:57stage

112:58okay so found a cyber cell i think i'm

113:02gonna have to skip this as well

113:04am i gonna have to skip this you know i'm

113:07gonna find out all the time i was born for

113:09this

113:09so you are well known to some people but i

113:13don't think everyone here knows you can you

113:16give us a

113:16little bit more about your background

113:19absolutely and first let me just say that i

113:20'm thrilled to be

113:21here the energy is absolutely palpable um i

113:25spent the first six and a half years of my

113:28career and

113:29the last six and a half years of my career

113:31at a hedge fund called cotu focused on

113:33using external

113:35data to make real-time predictions about

113:37the economy what is inflation doing what is

113:40consumer

113:40spending and so on and now i've started a

113:43new company i have to cut the sides it's

113:45help called cyberston which is a data as a

113:48service company that's native to snowflake

113:51and provides content for the marketplace

113:53that's super cool why snowflake you made a

113:57choice why

113:58it's a great question besides just you and

114:00your charming personality um i mean if you

114:04're looking

114:05at this and you're wondering what's going

114:07on here like what like this this company

114:09that's basically

114:10only working on the snowflake platform and

114:12selling data in the marketplace and this

114:14diagram sums it

114:15up yes they are scraping or getting data

114:17from these uh various sources they are

114:20putting a little

114:21bit of magic sauce in the middle they are

114:22processing it they're probably doing data

114:24science they're probably doing a whole

114:26bunch of workloads in uh snowflake itself

114:28or maybe outside

114:29of snowflake as well and then they are sort

114:31of packaging that data and selling it on

114:34the

114:34marketplace so you and i can go and

114:36purchase that data as context for our own

114:39business models our

114:40own bit of analysis and of course because

114:42you're doing everything in this single

114:45platform it's a

114:46fairly seamless experience to be able to

114:48use that their data inside of your platform

114:51and and the

114:51pricing and everything is all managed by

114:53the marketplace so i think this is a nice

114:55sort of

114:55summary diagram to show you what's going on

114:57there's two reasons at the big level we

115:00have a

115:01shared mission cybersyn's mission is to

115:03make the world's economic data available

115:05and usable to make

115:06it mobilize the world's data that aligns

115:09with snowflake's main mission as well and i

115:12would

115:12point you back to frank and jensen's note

115:14to thinking back what the most important

115:17table

115:17in your organization that lives in snow

115:20flake well what about using everybody else's

115:23most important

115:24table too that's snowflake data sharing and

115:27that's what cybersyn is helping to enable

115:29by providing

115:30this content on the marketplace right at

115:32the tactical level snowflake gives us the

115:35distribution

115:36it lets us connect to all of you and lets

115:38all of you access our content in one click

115:41that's awesome

115:43so where are you is this just like a good

115:46idea or do you have customers or any

115:48customer stories

115:49we're rolling it wouldn't be a good idea so

115:51far we've released a series of public

115:54domain data sets

115:55gaps in the snowflake marketplace where we

115:58've provided original content to fill those

116:00gaps

116:01things like inflation data population data

116:04so on we've had more than 500 snowflake

116:07accounts

116:08snowflake customers sign up and some

116:10organizations i'll call out blackstone as

116:13an example have used

116:14some of our data sets such as our sec feed

116:17data or inflation data for their own data

116:21science use

116:21cases nice i'll point out these

116:23sophisticated organizations it's not as if

116:26they cannot get this

116:27data on their own they can but with the

116:30power of snowflake data sharing we save you

116:32that etl work

116:33and i'll allow you to sort of focus on the

116:35downstream value add exactly but i'll

116:37summarize

116:38it pretty well i think that's not true of

116:40all your data they can just get it i think

116:43you're

116:43leveraging snowflake summit to announce

116:46some new data products exactly as frank

116:48said yesterday not

116:49everything can be free so we are launching

116:52two proprietary products we call these

116:55products

116:56foundations because there's something you

116:58can build on and the keyword is product

117:00they're not

117:00just data sets they're data sets and native

117:03applications in the form of streamline we

117:06're

117:06launching a consumer spending data set and

117:09an e-commerce data set that we think will

117:11be useful

117:12to retail cpg and financial services

117:14clients that's so this is like this is huge

117:17oh my words

117:18i mean i don't want to sort of um over over

117:22egg this in terms of like excitement what

117:27essentially

117:28they're doing is they're launching

117:30proprietary um data sets and applications

117:32in snowpark so they're

117:33not just saying here's the data they're

117:34also saying here are a couple of other

117:36things you can

117:38plug onto this data set alongside of your

117:40own that give you a capability and they're

117:42targeted at in

117:43this case specific industries so retail

117:46fast moving consumer goods why because of

117:49course um it makes

117:50sense to start there because these are

117:51going to be the industries that probably

117:53have the biggest

117:53propensity to spend on this data so and

117:55they have some general data that's already

117:58free i think they

117:59talked about inflation and population that

118:01sort of table stakes think of that as a

118:03marketing play

118:04you get to start to use this data you learn

118:06how they work you get a sort of interaction

118:09with the

118:09company they reach out to you they talk to

118:10you more they understand what you can do

118:12but in the

118:13background they're building a much bigger

118:15suite of paid products and um this this

118:17feels nice it feels

118:18it feels like a good use case for um how

118:19how they're working with the with the snow

118:21flake and

118:22actually it's a really good play into this

118:24concept that like snowflake has sort of

118:26transcended and

118:27it's become a platform right it's that's

118:29exactly what platforms do they create a a

118:31place where

118:32people can start to do business on top of

118:33what they're already doing and snowflake

118:36can just fade

118:36into the background as the place where that

118:38canvas sort of happens it's awesome and the

118:40question that

118:41i think all of you may be thinking of and i

118:43'll ask on behalf of everyone here where

118:46else can i

118:46find your cyber scene products nowhere we

118:49're exclusive to snowflake

118:50it was like so crap you have a booth at the

118:56conference you're you on your team

118:59i wonder if that's like a session if you

119:01want to learn more if that's like a paid

119:02thing if

119:02snowflake you're doing great stuff how much

119:07can we pay you to stay here thank you or

119:07maybe it's

119:08just that it's really hard to do maybe

119:09maybe the cyber sin have just decided this

119:11is actually really

119:12hard to do as as frank alluded to in his

119:15opening remarks the marketplace is not just

119:19data last year

119:21we shared a broader vision for an

119:24marketplace for native apps and today we're

119:28very excited that the

119:30native app framework is going in public

119:34preview and in the last few months that we

119:38've been only

119:38in prior preview the momentum and the

119:40excitement is through the roof you see some

119:43of the logos

119:43everything you see here is an app that is

119:46actually published already in marketplace

119:49for more than 25

119:50providers over 40 apps and i was chatting

119:53with our marketplace ops team i think there

119:55's another

119:5680 or 90 in the queue ready to get approved

119:59so there's a lot of momentum and some of

120:01the names

120:01here like bloomberg and others it's

120:03completely amazing so we're very excited

120:05about the products

120:06in the last couple of years we've been on a

120:10mission to simplifies monetizing data

120:14products

120:15and everything we did originally was around

120:18usage-based business models

120:22billing by number of queries billing based

120:24on time and of course all of you and many

120:27of our

120:27partners are so excited that you came up

120:29with 50 different business models and we

120:31cannot just

120:31implement each one of them let me check i'm

120:35completely excited to share the

120:37introduction

120:38of what we call custom event billing which

120:42is that for any app or any data product you

120:45will be able

120:47to bill in whatever the right units are for

120:50your business if you carry the apple

120:53analogy this is

120:54like giving the user um the option between

120:58a one-time purchase of an app a

121:01subscription to an

121:02app um or like a recurring subscription if

121:05that makes sense so like a usage only means

121:08you pay as

121:09you use um that doesn't actually exist on

121:11the app store um maybe it kind of does with

121:13apps that are

121:13made by apple for example because as long

121:16as you're um buying an iphone you're

121:18getting access to apps

121:20like imessage and so on and so forth so you

121:22kind of but not really uh base monthly is

121:25monthly

121:26subscription one time is just a one-time

121:29payment um and you've also got these events

121:31which is kind

121:31of nice so you can you can actually target

121:34specific kinds of events upon which you

121:36build

121:36your business on so it's really good when i

121:39do a one-time bill per time per user

121:41whatever you want

121:42usage based or not we will simplify the

121:45billing of um whatever your model is the

121:49other thing

121:51that we hear in the concept of billing more

121:54on the other side on the purchasing side

121:56the way i hear the question or some of you

121:59express it is i want to buy data or apps

122:02with snowflake

122:03credits and we quickly say well i don't

122:05think that's what you mean because snow

122:08flake credits

122:09is sort of like a unit of compute but we

122:11got you we do understand what you mean and

122:15we are also very

122:16excited to announce the introduction of the

122:20ability to buy from our marketplace by

122:23drawing

122:24down from capacity commitments to snowflake

122:26interest this is now generally available

122:29for all

122:29of our customers in the us and will

122:31continue to expand later on and what you

122:33can do is if you've

122:34made a capacity commitment to snowflake you

122:37're going to be able to deduct some for apps

122:40some

122:40for data sets and of course your

122:41traditional consumption of resources so are

122:44we ready for

122:45another demo interesting you want to see

122:48native apps okay i see commitment and try

122:50to understand

122:51what that actually means i think at the

122:53moment what it means is that um when i look

122:55at myself

122:56as a customer i pay every month for what i

122:58use and i think what they're saying here is

123:00instead

123:01of doing that i can buy a commitment up

123:04front um and let's say i can say i'm gonna

123:07buy a million

123:08credits uh up front and um i pay for that

123:11now um that's kind of useful because it can

123:14protect you

123:15from things like inflation and all the

123:17other stuff so you buy it at today's price

123:19it's great if you've

123:20got a spare budget going and you kind of

123:22want to you know buy ahead of time or maybe

123:24you've spun

123:25up a data engineering project or you've

123:27spun up a data science project and what you

123:29'd like to do is

123:30to buy a commitment to capacity um so that

123:33uh they can draw down from that capacity

123:36and essentially

123:37um you know when it gets to zero you you

123:39have a natural point at which you decide

123:41whether it's

123:42actually profitable or not so it's it's

123:44also really nice sort of piece of

123:45flexibility for

123:46for companies to try things out actually

123:48and and know that the commitment is capped

123:50rather than

123:51the current setup where you can kind of say

123:53yes we'll get it and then all it takes is

123:54for one

123:55person to do something really large on the

123:57data and actually you've gone over so um

123:59yeah it's

123:59really nice let's see a demo i'll say our

124:02engineer teams did not all come here many

124:05of them are

124:05watching on the on the live cast and they

124:07want to hear you are you excited about them

124:10yes i'm glad

124:11you said that because i i've been saying it

124:14earlier right now this is a tough crowd

124:16like

124:17come on guys all this stuff should be

124:19exciting me thank you gristan we are super

124:26excited to bring

124:28the native applications framework to

124:30developers around the world with the public

124:32preview launch

124:33in aws today i'm going to show you how you

124:36can build your apps with snowflakes highly

124:40reliable

124:40and global multi-cloud infrastructure how

124:43you can build your businesses with snowfl

124:45akes global

124:46marketplace and flexible monetization

124:49models and how you can deploy your apps

124:52close to the

124:53customer data while retaining full control

124:56over your intellectual property let's see

124:58all of this

124:59in action by building an app that predicts

125:02lead times for manufacturing orders dash

125:04are you ready

125:05all right we want you to bring your

125:08favorite developer tools and best practices

125:11to snowflake

125:13i personally love vs code so we are going

125:15to write our application with snowflakes vs

125:18code extension

125:20this is quite nice i need to try this

125:21actually i made a video about using i have

125:23multiple

125:24snowflake in vs code before this extension

125:26existed but now it's possible it's all i

125:28need now

125:29is a manifest file for the app config and

125:33the setup script that installs the

125:35application in

125:36the consumer account now we're ready to

125:39package these code files that i've already

125:41uploaded to

125:42my snowflake account let's head over to

125:45snow side and create an application package

125:48which is an

125:49independent self-contained unit of code and

125:52data that you can share with your customers

125:55without

125:55exposing your intellectual property before

125:57committing our code changes and creating

126:00the

126:00first version we should test our

126:02application with native apps framework it's

126:05super easy to do

126:06all you need to do is to install the

126:08application package that we just created in

126:11the same account

126:12once the installation is done an

126:15application instance is created while this

126:18app is installing

126:20dash already has one ready to go this is

126:22what our application built with snowpark

126:25and streamlet

126:26looks like i think it looks amazing dash

126:29what do you think let's commit our code

126:32changes by

126:33creating the first version so earlier on i

126:37think in a in a in a separate um in the

126:40previous video

126:41when i was talking about apps and yeah

126:43applications i said what what is an

126:45application and it's it's

126:47super interesting if i just go here and he

126:49says uh look what do you think of this well

126:51this to me

126:52doesn't look like an app it looks like a

126:54dashboard right and i can't it's it's it's

126:56just just as a

126:58like a lay person who's you know doesn't

126:59know how to code anything is just sitting

127:01here like

127:02you know drooling at some of these features

127:04and looking at this and you go well what

127:06makes this

127:07an application like how is this dashboard

127:10to me or how are these four charts any

127:14different to

127:15a dashboard and for the record these are

127:17not um you know these are not

127:19groundbreaking charts either

127:21like i'm not seeing any sort of interact

127:23ivity i'm not seeing the ability to kick off

127:25anything else

127:26maybe we're about to see that but um like

127:30why is this an app and not just a dashboard

127:33and if i've

127:35misunderstood this entirely let me know ash

127:38what do you think let's commit our code

127:41changes by

127:42creating the first version with versioning

127:45built into the framework you can increment

127:47ally release

127:48new features or bug fixes alike you can

127:51target releases to your customers and

127:54confidently and

127:55safely deploy changes while so okay let's

127:59while this version is created i want to

128:04show you

128:05something super important today as

128:07application builders you have to spend a

128:10significant amount

128:11of time and money getting your apps ready

128:14for security compliance snowflake

128:17automatically

128:18reviews every single version of publicly

128:21shared apps for security threats and abuse

128:23we believe

128:25that this is going to accelerate the sales

128:27cycle for you and time to value for your

128:29application

128:30customers as you can see in this example

128:33this version was submitted twice it failed

128:36the first

128:36time because of a security issue which is

128:38then corrected and now we have an

128:40application that's

128:41ready to go let's publish our application

128:44on the snowflake marketplace and for that

128:47so there's a

128:48small thing like if the first time it it

128:50passed and it failed so that's two things

128:53why in the

128:54second security scan is it not two things

128:56passed i don't know to me that's a pretty

128:59straightforward

129:00bit of communication issue in the interface

129:05but yeah that's yeah that doesn't make

129:07sense to me

129:08we're going to head over to the provider

129:10studio where we already have a listing for

129:12the application

129:13you can monetize your applications right

129:17here in the data cloud without having to

129:21set up or manage

129:22complex billing infrastructure snowflake

129:24gives you a range of flexible monetization

129:27models to

129:28choose from as you can see in this example

129:30we are creating a completely custom billing

129:33model

129:34where we can charge the customer for lead

129:37time prediction and we'll also include a 30

129:40-day trial

129:40now we are ready to publish our application

129:43why 30 why not 20 or 10 or 7 in just a few

129:50simple steps we've published our

129:52application to

129:52snowflake marketplace where more than 8 000

129:55customers can instantly discover your

129:57applications

130:01let's look at what the customer experience

130:02for using and discovering these

130:04applications looks