Welcome to Energy Bytes with John Kalfayan and Bobby Neelon. Your essential guide to the intersection of data and energy. This podcast dives deep into the world of energy, shedding light on how data, AI, and technology are revolutionizing this sector. Each episode equips listeners with insights into the most efficient tools and resources, paving the way for a data-driven future in energy. From technical nuances to broader industry trends, Energy Bytes offers an unparalleled perspective on the evolution of the energy industry. Join us as we decode the algorithms of energy, one byte at a time.
Welcome to another episode of uh Energy Bites. Got uh I'm Bobby Nean. I've got the Rad John Calan here with us. How's it How's it going? It's going great. Been been a few weeks. We've been crazy busy. You've been moving. A full week of uh internet and power. So, yeah, a couple tornado and a hurricane and all kinds of fun stuff. All kind of good stuff. And then, uh we've got Steven Pace with us today from Snowflake. Uh you're a principal sales engineer, correct? Correct. And based out of here in Houston. So, really glad you could join us. Thanks. Happy to be here. Yeah. Yeah [snorts]. Well timed since uh as of today basically it's Crowd Strike has I guess they're slowly rolling out a fix or I don't know. Well, they didn't slowly roll it out. [laughter] It's true. Shoved it out. They're not pushing patches anymore. It sounds like Yeah. If you search for I don't know if you can edit in graphics as you're you're doing this, but if you look go to the Azure status page right now, they have a big long like normally it's like you know different colors or whatever. they have a big long thing of hey this is what's happening and uh that so it's yeah it's rough it's it's one area that you know snowflake is multicloud we started in AWS but then we went to Azure and it's our fastest growing platform and we support Google as well But what's, you know, kind of funny in this case is we, we can help isolate customers from these type of problems is because if you wonder like, "Well, if I can't get machines, how is it that Snowflake is running?" it's because we ahead of time are, are spinning up hundreds of thousands, millions of machines around the world. And so when we keep, you know, we have them. And if we detect there's a problem getting new machines, like if they would start, they would blow up. We don't give them back. We just keep them and then, you know, we stop giving them back and then customers are still running. They could have problems in other things that they have because they maybe tried to provision something outside of stuff like, but for the stuff we're running, a lot of times we'll just keep running. And so it's Yeah, it's it's interesting. It reminds people about the concentration of tech. Like we have customers like Caterpillar that are multicloud that practice failover between AWS and Azure and and do that. That's a hard thing to do, but if you're somebody of that size, you're it's it's worth it. But it's, you know, it's it's we all saw this and, and it's not just Azure that has this. But like, remember there was a story a couple years ago where like people's Roomba stopped working, and it's because there was an AWS region that went away, and people like, "Why is my Roomba even using the cloud?" It's usually US East one, isn't it? Like Yeah. So the most common ones that everyone Yeah. So other ones have had that and you know the last big Azure outage they had an active directory outage uh which then couldn't log in. So then you know nobody could use Azure and then people it gets people thinking about like their failover fail you know strategy big time. So because a lot of times when Snowflake if it just runs for a year and you never have one outage you're like well you know I'll put that off but then if you have something that impacts then you're like okay I I need to think about it. Yeah, because I think, at least last night, I mean, know, we're on Snowflake at work, and um I, don't I haven't dug in enough to know if it was on Snowflake side, but we were starting to get a bad gateway, you know, errors or connectivity. A lot of jobs were failing, but we have a lot of stuff tied through Azure as well. So it might have been tied to that but um because ours was on US Central which I think was the one that Yeah. Yeah. That was the most main everything was down was like just red for Azure Central but um I read it's the the biggest outage in IT history as of today. Oh I believe it. Yeah. And then I saw like banks I mean I think stations get up there a lot across the world too. I mean I was like saw a picture from Amsterdam this morning like Yeah. Berlin airport closed and I had a friend. Could you be could you imagine being at the airport and just seeing all the screens police [laughter] across be like this is encouraging. Yeah. Hopefully my plane's not running my phone, right? Yeah. I I had I had a friend from high school that on a big trip and they were their flight got cancelceled in Brooklyn. It's like why? Well, the whole airport's closed and so it was having to navigate that. How could the whole airport be closed? But it gets even just like I saw a like time-lapse GIF of the flights yesterday and it was like it reminded me exactly of what 9/11 looked like when it literally just shut down all of air traffic and it was just like oh man that's [laughter] a that's a hell of a mess up there. [snorts] Well awesome. So, you know [clears throat], diving into you know let's talk about what you're working on now, or what, you know, really I want to dive in about Snowflake because we hadn't really had anyone on to talk, you know, deep about Snowflake. But I mean, there's, you know, we're using it. I know there's definitely a more than a handful of people in oil and gas using it and across the energy space as well. Again, we're not exclusive to to oil and gas, but we do a lot of it. But um, just you know, if you want to talk about kind of what you do with them and like maybe some projects you're working on; but then also, I mean, you guys just had, what, Snowflake Summit and? you we can talk about some of the the new additions. But and maybe I don't want to jump the gun; let's start, like, if people don't know what say what Snowflake is like, can you? you know start at the highest level like what is Snowflake you know and uh why do I need another database? Yeah. So there's a lot of that and and I and that's really kind of where I came from too. Uh I mean stepping back a little bit um I [clears throat] can't you know I my first job was at Shell and wrote some software we you know part part of a team sure That was supported like data warehouse automation, and then we ended up using some VCs and we bought out Shell, and we became our own independent company. And so, I was in the data warehousing space for many, many years, and so, uh, I mean, I was very much in that camp of, you know, look"we, we have some good databases, why do we need something else?" And I think the thing that woke me up to maybe why you would need something different is I saw the evolution of NZA. Okay, so we we supported platforms like terod data and here's this upstart of like why do I need this in a teasa box and I saw how they sold it which was you know hey you got a slow Oracle mart we're going to roll this box in we're going to take all the data and we're going to run your query it's going to be faster and so we're you know solving a problem that people had And you know, that caused a lot of consternation for terod data. Then, "Oh, we're going to change our strategy; we're going to have a small box." But that at least woke me up to where I can see why you would need something, and I think this is an area where Steve Jobs is really good is: is there's periods and inflection of technology, and I kind of consider myself a technology person. Sure. That things are impossible until they aren't. And so he'll see something like he'll go to Xerox Park and say see a mouse and go okay well the world's changed. We're going to do something different. Or he'll see a flat screen and say oh [clears throat] this is going to be our phones and everybody's going to do that and you know Microsoft see Balmer will make fun of them and say oh you know we've been in phones forever. They're not there there's all these problems you're going to have to work through. But he'll go, "Yep, but we're going to work through them and then everybody's going to have an iPhone." Look up Windows phone. [laughter] So, yeah. And this is the thing. So, this is kind of the world I was coming at it from. And I don't know if you recall back in the Hadoop like, you know, maybe this is a little bit of the management by uh airline magazine kind of thing where um you know, people are like data warehousing is dead. We're going to just you know, there's this new thing Hadoop and you need to have and it's because why do we need it? Well, the the 3Vs or 4Vs depending on how you do it. We have the volume and the variety and velocity of data is just such that we regular databases can't hold it. We're gonna have to have this other thing, which I just thought was so funny because, like, and it and it drove companies like terod data crazy, like because "Big Data" is like the stupidest term there there could be, because they're like, "We've got multi-etabyte data warehouses and you're saying this Big Data thing is not?" And so it just it drove them mad, but it kind of was, I think, good from a marketing thing of, of like. saying it's different but also making the existing companies feel like oh you know are we legacy now and all this stuff so that that that world came up and one of our founders and this so I I kind of came at I I I knew two people. There was a guy in the Woodlands name is Kent Graziano. He was sort of our first evangelist for for for this and I was uh helping out with DMA here in town and we would have speakers and Kent came in. I saw very you know early on demo of this and and so I kind of it was on my radar and then I saw read a thing from one of our founders. He's a French guy. His name Benois, and somebody asked him about Hadoop, and he just you know like everybody gives these nice things and. And I think in, a lot of times in the media, they like, "We're going to get somebody with a controversial thing," and I just happened to see it because his was like...he basically said Hadoop is stupid. We can solve this another way but if you're going to move to the cloud you need a fundamentally different architecture and snowflake really is a completely different architecture. And so he thought through like if you're moving to the cloud what would you need? And then it's, you know, we so we sort of pioneered the separation of storage and compute and and these things and he just worked through that and um and so I when I saw that I go okay I'm aligned to that guy. So it was got on my radar we went out of stealth and then and I was then it was the right time it's like okay I'm going to move over and I was like employee 1100 or so like so still we were kind of sizable. Uh at that point we were just we're saying we're a cloud data warehouse. We were always meant to be a platform, but you can't just sell something that is just like nobody knows. You have to say we're going to do this. Be really good at this thing. Yeah. Yes. And we were we c like we could crush Redshift or Synapse and all these things that would just uh just let customers play with it. And even to today, you can go on the Snowflake website, you can go to signup.s snowflake.com, you can click, you can get a trial and you can just play with it. No, the simplicity is probably the most beautiful part of it. I mean, absolutely. I mean, I remember the first time I stood one up, it was because Bobby was like, "Just go stand one up." Yeah, it's like...it's literally a few clicks and you've got one stood up. Just, you don't have to have Google, you know? Because I mean, BigQuery is a fine, you know, warehouse, like, in, you know, similar vein, but, like, you have to have a Google Cloud account and then there's all this, you know, those. got to go set up all the APIs and the service agents and make sure all the permissions are correct and got to secure it. I got to tie it into my SSO and all this and it's like pretty soon it's like, well, I just want to go do this, right? I just need a database to hold [laughter] my data. I don't want to be an IT manager. [snorts] Yeah. So, I mean, you know, now if you look out, so if I was 1100, there's like we're more than 7,000 people now. So, it's just growing and growing. And so, I I was brought in to help. So, I'm in pre-sales, meaning I...there's usually, uh, the way it works in tech for pretty much all the tech is there's an an account exec and then there's a technical resource with them that understands the platform. And you're sort of a pair, and you go in there and you can talk about, you know, "Hey, how Snowflake works, how does it compare to what you have, how does it integrate with all your other stuff." And so, uh, my initial customers were brought in for Exxon and and some others. And I guess just proximity of being in Houston, you have more energy companies. Not that they're only energy, like we have Taylor Brands, like Joseph Banks and those kind of uh companies. Um, what's cool is as you drive down I 10, almost every big building now with the logo is now a Snowflake customer. So yes, if you look at like from from Katie on like you see Kico Phillips and Techn I think they just changed the logo in their building but if you look down the beltway as you pass past pass the beltway you see BMC or NOV off of there. U Sex Marathon is another one and they're right by us. Yes. So I mean it's kind of cool to see it evol how it's evolved but it's in every case it's like hey we we've got this problem. a lot of cases in the early times was moving from Hadoop uh a lot or if something very expensive like a HANA or something and it's like how can Snowflake help on that? Well, I was like, well, okay, kind of I think Marathon was sort of making the joke. They said, "You guys are $23 a compressed terabyte. HANA is a million dollars a terabyte." No, that's not I know it's a it's not unfair ones in memory and everything, but they're they're basically like it's very expensive and you guys are very cheap. So, yeah, I mean that's the mean a the storage is just a pass through, you right? I mean like it's roughly what a S3 bucket costs at the hot tier, right? Like and then but then the fact I think it's compression side of it is is the huge part. Yeah. And compression is very good as well. So if we're going to dive into that at all, but I mean like Yeah. So there was a kind of a funny thing that it kind of hit the, you know, the stock one one quarter because we improved our compression algorithm 30%. So one day you go to sleep, you come back the next day and now suddenly all storage moving forward is 30%less than even what it was. So I think our compression rates and it's different for every kind of data, but if you load a terabyte of data, it's not unusual that it'll be like 200 gigs after you load it. Wow. Yeah. So it's about six to 10 necks. Yeah. And we get a good ratio. So our uh commit to AWS uh this is public information I think is something like $2.5 billion a multi-year contract. So we get a pretty good rate and then we try to pass that on to customers and $23 is the starting rate. Yeah. So if you're a bigger Snowflake customer and it's right on our website. This is another thing I like. We're we're very open. You go snowflake.com/pricing, you can see the price list and it says $23, but as you start to increase, I think it goes down to like $13 or something like that, whatever. So, um, we're trying to be competitive for, you know, keeping it out and something else. So, you know, yes, you can keep it yourself, you can manage it yourself. We support iceberg tables, for instance, very popular open table format. If you want to manage it, if you want to take on that extra homework, you certainly can, but you got to beat that price. for for big customers, it's it's hard because it's like there's a lot of management that goes into it. It's not only that it's that cheap, it's that we're managing all the file sizes and the file structure and everything too, right? I mean, all the partitioning and like I mean that's not trivial. I mean like I remember y going through like an AWS uh data lake, you know, in a day kind of training and it was like it's you really got to be intentional about how you set this up and like how you partition files and do everything under the hood and then you got to secure it. I mean, there's just there's so much that goes into it to do it well and like like you're saying, can you beat Snowflake at their game? Like, if it's not likely, yeah, it's hard. I know we're going I'm probably giving too long of answers for podcast. No, that's but like uh just one example of that because we have a customer that said they wanted to take on that. And if you have a big engineering team, you can do it. But, you know, our micro partitions, the default size is 16 megs. And like imagine you change a file. Well, now you either gonna have a sparse file like let's say you add one record. You could have a a micro partition that's 16 megs and then one that's like one row. Yeah. But that that now you're joining things that are on equal size. So we're like constantly compressing them and like if you change a row, we're going to rewrite that micro petition and we're going to end the old one. And then there's time travel and things. So you could keep it around for a day up to 90 days if you want. But you [clears throat] just don't have to think about it. It just works. It just looks like it's a table. But behind the scenes we're leveraging the cloud providers to give like resilience for the storage because we're using S3 or we're using blob or we're using GCS but in behind the scenes and then our compute is commodity compute. Uh but over the years since I've been here like we've done the word interroad land is like a forcefully sweep. Okay. But basically you you you have a certain kind of warehouse a certain machine type. Next day you come in all the machines are better. And it's because we put all our machines that I'm talking about in these pools. Yeah. We send them back to Amazon or Azure and Google and then when we start provisioning those machines, they're a different kind of machine. Interesting. And they have maybe more SSD or more memory or whatever they are. But you don't and that's, you know, we have not changed the prices on that. Like so our prices over time, you look, we have something called a snowflake [clears throat] performance index. It tells you over time if you'd run a certain set of workloads and then you move forward a year and you run those same workloads, they're going to be faster, they're going to be cheaper. But even even at the same t-shirt size, right? Because that's kind of you all do extra small, small, medium, large, you know. So you could keep running things on an extra small year over year, it'll just get better because you're using better hardware. Yes. Because over the time we've changed like in my times, five and a half years, we've changed that at least two times on AWS. Gotten better machines. Now there are some other machines now. We have snowpark optimized uh servers which have 16x the memory. So if you're training models now we're doing more and more in AI front. Okay. Uh because you asked like three questions at the beginning and I haven't [clears throat] even got to the other two but no fair but but uh no I think let's just go organically. So so uh so sometimes when you're training a model you need more memory. So we now have snowpark optimized servers. We charge 1.5x for those. But you know now you could maybe use a smaller number of them because you don't need more machines to get the memory you need. You can have less but they just each machine has more memory. Uh and then now we have and it's not as visible to customers not because they're not in in those types of pools where you see where t-shirt sizes but we have something called snow park container services which is essentially containers and those containers can be GPUs. So we now have GPUs in those pools. Some of the machine learning models and things work better with GPUs. So we now have that as an option as well if you can find them. Well, we have them because we get we've paid Nvidia [clears throat] a lot of money and and and we also pay the cloud providers to keep them and then they're in our pool so that when you need them, you don't have to wait two weeks. I have had some customers who said, "Hey, I tried to get a GPU and it took me three hours or like I one customer said more than a day." Wow. But in certain regions, keep that warm pool essentially if Yes. And so there there and and there are some other things coming. I I don't know what things I could say, but um you know, so we have them in different regions, but some regions are harder. So, but you you'll be able to leverage some if you you know, if you approve um where we can help minimize those shortages. So, there's some Yeah, there's some things we can do. I hope Azure takes notes on that because took me a week of emailing back and forth with it just to get my allocation for one [laughter] GPU server a couple weeks ago. [clears throat] It's just So, I think there's one more. I mean, you hit on a little bit, but I want to go a little bit deeper. You almost like I would say snowflake fundamental at least. What y'all kind of sell there though is can you talk about um kind of the idea of like linear scaling and then kind of the warm pools like I mean so basically that you if I go from an extra small to a small I basically double my cores and it should run you know twice as fast. Correct. I mean in theory I mean yes I mean it's I've tested it's I mean like in multiple occasions it's pretty accurate. Yes. So this is something that Benois and Tiri our founders came up with very early is that since we charge per second yeah and if you can double the compute you and half the time the cost does not change. So imagine just assuming auto spend. Yeah. Yes. So the the rough math is for any given job like let's say a job took one hour on an extra small. Yeah. It would take 30 minutes on a small. So we charge so at enterprise level is $3 for on demand price. So imagine the $3 it take you three hour. So so with one hour it would cost you $3. But with a small it would take you 30 minutes. So it would be $6 for if it was a whole hour but because it's half the time it's still $3. So now a medium would be 15 minutes still $3. Yeah. Then a large. So you see how you're you've compressed the time. So suddenly you can do these huge like loads from Hadoop and money over into Snowflake. Yes. And you're like, I can compress this time. It cost me no more money. But you're you're exactly right that at the end of that process, if you flex up to do a big load, you should flex down back to your original size. And that is one thing that's really uh works very well in Snowflake is we auto suspend. So by default, you know, it's we're going to turn off the if nothing's happening, we're just going to turn off. And some people worry about that because in some worlds like like in a Spark world or something, you're like, well, I'm never going to shut my cluster down because it takes, you know, 30 minutes to start up or whatever. And so they just when they wherever they're specking like the price of things, they just like, "Well, just assume it runs 24/7." But the vast majority of Snowflake customers, it's off all the time and you're not getting charge. It should be and and again like then the crazy thing is it just it's on like when you need it or it scales up when you need it. Like it looks like it's on. See, this is the thing people that that shocks them at first is if you go in there like, "Well, but but I if I come in on the weekend and run a query, I don't want to wait 30 minutes." Like, nope. You're going to come in, you're going to run your query and we're going to provision that cluster to you in under one second. Yeah. And the reason we can do this is because of those free pools. So we're running these machines and our job is to make sure we have enough machines. So if a customer comes in and says, "I'm going to run my query or do my ingestion or need any warehouse at all," we're going to have them sitting in the pool. So we're managing it. We're we're we're we're essentially um we're a managed service. You ask what Snowflake is. I still have not answered that question, but uh you know, we're a managed service. So uh we need to make sure that when you run your query, we're going to provision it and you're going to run it. you won't know it's off, but there's a little light on your if if you're using Stoite, it'll be gray and when you run the query, all of a sudden you'll see the light turn green. That means we're billing you. And if you use a minute and a half and then you shut down, we're going to charge you for a minute and a half. And so that's different than a lot of services back in the day like I think in Synapse if you use ran one query in an hour, they charge you for the full hour because they're just they're not designed. And I think this is an area where you like for something like Spark for instance, it really is not designed to run uh in this. serverless managed way because there's so many knobs. There's, like, you know, 500 knobs and, you know, I think, you know, it's going to companies that try to do that, you know, in a managed way are going to really struggle. And that's what we've sort of seen ourselves is that people are trying some other services, and it just we've been doing it for 10 years it; it sounds like it's, you know, easy but, it it really isn't. make it look easy but yeah yeah and even this thing with crowdstrike today we've you know ourselves over the years as we've gone to more regions and now we support three cloud providers you know we've done things where we roll something out and we think, "Oh, there's no way that could hurt anything." We'll roll it out wider and it affects more people. So now you'll see when we roll out a new version of Snowflake, which is every week, we're essentially rolling out smaller regions first and we we have a bunch of canary and pro test processes and then if everything looks like it runs and it goes to bigger regions and bigger regions. Eventually, you'll see by the end of the week the all the regions will have it. But there's times where we'll sort of internally get notes like where we're going to roll something out. We start to see some kind of indication maybe something's not right and we'll just roll it back. Yeah. And that's the cool thing is we can upgrade a customer. They could be running a thousand jobs and we can downgrade them and running a thousand jobs and it should not affect them at all. That's wild. So it's it's designed that way by by default. And so doesn't mean we're perfect, but you know we're you know we're quite stable and that's what we just want customers is say it works. And that's why we've got a lot of takeup in the energy space I think is a lot of times energy companies are running lean and mean and one of my customers um you know I'll I'll say the name uh cord energy now they were they merger of Oasis and Whiting and when Oasis you started they had three people essentially on their team and you know one was like a player coach one had a lot of knowledge of the internal systems and one was sort of like a coder and they were able to move from mapar and SQL server and move into snowflake with not a lot of knowledge. They did it themselves. Yeah. And they're used to they were used to SQL Server. So Snowflake presents as a database and yes we can run Python and Java and Scala and everything but if you're used to SQL you can get stuff to run. Yeah. And they were very successful and came in on time on budget and did that. And then fast forward to now, like, you know, when when the merger happened, I think after they did a little bit of analysis, uh, I think what they told me was it took them two days to move all the Whiting data into the same, because, you know, you have this period of rationalization or whatever, and after to the two days they were able to report both the Oasis and the Whiting stuff together. So still very you know happy customer many years of operation grown um and I just you know I love those kind of success stories. Well I mean that was our experience too. I came in and I was the data guy, you know, like, you know, and we we pulled a lady over, Michelle, she's on my team now. Um, she was a reservoir tech, but new SQL, but I was able to pull her over onto my team. And we... but by the time we decided we're going with Snowflake and then we went with kind of a traditional modern data stack, you know, Fiverr, dbt, because we're just running so lean. But, I mean, again, all of our core data was synced within a weekend, and porting over some queries and you know, within a month, like all the, you know, monthly business review metrics and everything were coming out of Snowflake. And it was just, you know, the time to implementation and fact that it works, and it's fast it just, you know, it's just really impressive. So that's awesome. Yeah. No, we love we love to hear that. Um all right so I've been holding you up. What is Snowflake? [laughter] I I should have print my elevator speech better. Uh you know so when we started you know we said hey we're cloud data warehouse and that was just one use case and that was kind of how we got people to understand it and then we became the data cloud and here's another tiein to let's say the the Oasis and chord case. One of the things that pretty much every energy company does is they get data from IHS market. Now it's S&P Global or or Invar or RIAD and these things. Uh and there's also now like if you have like well view or something from Pelaton they if you need your data you know when you ran it on prim data because you had an Oracle SQL server version of wellview but now in the cloud how do you get access to that data There's a, there's a live share. So, uh, we started calling ourselves the data cloud, meaning we want to be the that one platform where all your data runs. We're going to secure, we're going to have one governance model, one engine, and that's a breakthrough for somebody like an Oasis where they can say, you know, I'm getting that data. That's a pipeline I'm having to maintain. Yeah. AI or whatever else. Yeah. Yeah. But the best pipeline is the one you don't have to do anything. It's that S&P updates their table and you query it 50 milliseconds later and it's there because essentially technically what's happening is they're giving you the right to decrypt the those blocks from their table. It never moves to your account. You're able to query it, join it to your own data, and it just works. So then we were became the snowflake. were the data cloud and that that's taking on board that idea that you know as more and more companies join Snowflake there's a gravity to data and it doesn't mean you have to share data but it means there's an opportunity to share data and it you think oh maybe it's just energy but like Exxon they're building they build these platforms and they have these big electrical distribution systems and then they're like hey we need telemetry out of this box and then they go wait a minute the company that makes that Pal is in Houston and oh they have snowflake. So maybe instead of having this API and working all this stuff we can just say well if if Pal loads it into a snowflake and we can get a live share then so that's a case like is that energy? It's it's it's it's more like it's part of the data cloud. So that's we would say we're the data cloud and you know uh we have a a CEO that's come from from Google. He's the guy that ran uh you know their ads business from like one and a half billion to 100 billion. He's super into AI. And there's sort of like there's been multiple phases of Snowflake. We started with a CEO. He was sort of the VC that was behind like that got Ben Wah to leave Oracle and and and make Snowflake. His name is Mike Spiser. Then we had uh Bob Ulia who came from Microsoft. He understood databases. He understood how to build teams. or uh where he is now. I don't uh I I have to I think I got that mixed up. We can Google it. But so Bob Muglia, right guy for the right time. Then when we went to IPO, we got uh you know, Frank Slutman, he you know, this guy is a rockstar CEO. He he's taken multiple companies public. He you know, data domain and then Service Now and then Snowflake and then, you know, that was another five years and now we're on to our next you know CEO. So,shredar he, you know, he's super into AI, super smart guy and he's like, "Okay, this is where it's all going because you can't have a good AI strategy unless you have a good data strategy." But if we're going to be the place where all this data is, you really don't want people extracting the data and running it because, one, it's...you...I saw a funny thing. I'm just interrupting myself; you know how, like, people with seismic data is so massive and people say it really isn't big data until you can't back it up. Yeah. You know this is the thing. So, but but when you have this huge amount of data, you it's hard to extract it because it'll be slow. You got to extract it. You got to run somewhere else and then maybe you're pushing the data back. So, if you can run it in the same platform, then you're going to be a lot better off. It'll be faster. And so, we we now call ourselves the AI data cloud, which is taking on board that we're still this platform for for your data. It's going to be secure. It's going to be governed. It's going to be one engine, not multiple engines as we've seen with some other providers. But on top of that, we're going to natively be able to run AI. So we have, for instance, our own notebook, which you've probably seen in your account. Uh you can, you know, run Python. We have Anaconda Channel with, you know, thousands of libraries that are there. You're not limited to those. You can [clears throat] upload your own Python libraries if you want. Uh but that means you're responsible for the whole dependencies with all the other libraries, but you can absolutely do it. Uh we bought Streamlit, which is a great Python library for applications, and now we can host that. So it doesn't mean you can't host your own streamllet outside of Snowflake, but if you've got one and we have a lot of quick starts where you can see with like a hundred lines of code, you can have a really powerful app. You can share to anybody in your company. You don't have to worry about turning it on where I'm going to host it. I'm going to stick it in an EC2 machine or something like that. Nope. You just have it. You can say share it to whoever you want. And just like the rest of Snowflake, when you go to run it, it's going to spin up. Now, it may take a slightly longer. Maybe that one will take a few minutes, but you'll no not a few minutes, a few seconds to start up, but you'll see it. Hey, it's starting. And then boom, it's there. And then when it when it's done, it just goes back to sleep again. It's wild. It is wild. I'm for one, I love that by default your stuff turns itself off because everyone who's ever touched the cloud has left something on on accident and gotten a bill and been like, "Oh, I thought I turned that off." It's like for no business value. Right. Right. It just sat there for three months for no reason. And then you finally checked the bill and you're like, "Oh [ __ ] hope it hope for me fortunately most of those have only been a few dollars at a time but even then you're still annoyed by it and it's just like it should turn off if it's not doing anything Um, [snorts] but I, as someone who's been working kind of on the language model space for the last year, I'm very curious. You know, like, Postgress has come out with PG vector; I'm very curious, since you're talking about AI and and that sort of thing, do you see Snowflake [snorts]? Because the really weird thing for me when I first started getting into vector databases and stuff was like, "Oh, so I have to stand up a vector database, but then I still have to have a relational database to understand what the hell is happening in my vector database." And it's like, well, that's annoying. It doesn't make sense in most situations because you're going most people are going to want to understand the relationships of what's in there and even just metadata and stuff like that about the vectors that are in there. And so do you see you know do you see Snowflake getting into that space at all or do you see it being more on the kind of enablement side like you're talking about where you can build leverage AI via Python etc building on top of existing data sets. [snorts] Yeah. So that that is a very good question. So I'll answer it in a few ways. So one way is so we have a bunch of quick starts. So we love people to get their hands on Snowflake, and one of the quick starts is a RAG quick start. And for RAG to work so uh, Retrieval Augmented generation meaning we're going to take data out of some of your documents because the lms weren't trained with your documents, and we need to store it somewhere. Typically, that'll be a vector database, and we we have the ability to run pine cone in a Container Services. But for a lot of RAG processes, you just need that vector column, so a Snowflake table now has a special uh column type which is vector. So, in the RAG quick start, we have it will just use a regular Snowflake table; we're going to put all those vectors in that table. And how I kind of explain it to people is, like, think of like if the LLM is your brain the, brain it it doesn't know about your your data and all those things, but we need to add some neurons to make to [clears throat] improve the answer. Because if that was trained two years ago, it it needs to know stuff from today, it needs to know stuff from your data, so we're going to basically write that those neurons into this table. And then usually you'll have some kind of front end. Uh we have a couple options for that, but one is the stream having streamllet with a chatbot and being able to a answer ask questions. And now you can choose which LLM you want. So we're hosting multiple LLM. So we have our own where we've trained called Arctic which is very good for like enterprise level stuff. It can generate SQL. It understands like metadata about databases but it can't write like a Shakespeare sonnet or something like that. Like it's it's niche as focused. Yes. is very it's very efficient in token wise. So we've got that, but we've got the some big models like mi, uh, Mistral Large. We've got mixtro, we've got some of the Google ones, uh, gemas. We've got some Llamas, like Llama 3es, uh, and so we have other ones out there. And then we have some other models, or access to other models, so we uh, Neo models from Nvidia, um, it's part of our partnership there. We bought a company, uh, that's and we have our model from that called Arctic Tilt. So it's it now will be in on in your account. It's called document AI. So if you have a PDF and you need to extract content from unstructured this could be in text or images. It will um basically allow you to write the output of that to a table. But you just ask questions in natural language of the document and it'll come back with the answer and then you can say these are the questions I want out of these documents. And if later the format that they move around or whatever it's is very good for a zero a oneshot or whatever they call it zero shot model where you know you can just ask the question and you'll be shocked like it'll give you the answer almost the right answer almost every time. If it doesn't in the training process we recommend you know 20 to 50 documents. You go through and you say yes this is the right answer or no this is not the right answer and then um and then after that then you run you can run your process again and it will it'll improve even more. So that's an example of tech text image layout transformer. That's our Arctic Tilmo. So that's one we own. But the other Arctic one is it's open source. We we also I think it cost us around $3 million. But to show you the end toend process of training your own from scratch and you know so we documented it. We have a huge research team of these and uh we put it up in hugging face and everything. So you you don't have to have snowflake to play with it. You can get it from hugging face and cool go play. Now that's interesting. Um, I've been talking about this for months, so I'm glad you bring it up. I still think people are sleeping very much so on language models for what you just described, which is data extracting data from what appears to be a structured document known as a PDF, which is honestly just the worst document format that really exists, as someone who's been beating their head against them for the last six months. Um but like yeah I using language models to extract data from your existing data sets into a new data set is is like a cheat code it really feels like these days. Um it's just because like alternatively I would have had to manually do these things, right? Or it's like or I can just write a loop that that does this across all my files and then go back and fix the ones that look wrong. Yeah [snorts], I I highly recommend people go take [clears throat] a look at document AI because they will if if this is the kind of thing you need to do. It's just so easy. Doesn't mean you can't run Python and do things and um uh I'm talking to a customer right now that they download data from the Army Corps of Engineers and it's and that data is in a terrible format as well because it's giant PDS, but they need to extract things from there. And so it's a combination of needing to extract specific small things and then maybe also the summaries. And so this may be a combination of a document AI to pull out the the small bits and and something uh like a very simple PDF thing, you know, extractor. So we have, like, PI PDF, and there's a bunch of if you look at the Anaconda libraries, there's a ton of them to go pull out, you know, something else that's needed and and maybe that, then, that's a RAG process where you ask questions of the data and and get an answer. And you know, we had one of our SES early on in this process took all of our documentation and shoved it into an LLM. And because sometimes the answer that a customer needs, it really came from three different pages. Sure. And one of the things so Shredar came from uh a company called Neva. And what they were trying to solve a problem of, you know, not only when I give you the answer when you search, but I need to give you provenence of how I got that answer. Sort of in this age of like how do I trust this response? it links back to where it got it. And you're going to see some elements of that with LLM because it's like, okay, how did the LLM get from there to this kind of weird answer, you know, why did it hallucinate in this way? And so you'll see like in a lot of our quick starts, it'll have here's the answer and then here's a link to the PDF or the the resource that okay, we did that. So yeah, that's one of the big things we're focused on, which is why I brought up metadata on those vectors is it's like, yeah, most people, especially engineers, want to know where the data where the answer came from, right? And so, especially when you have a rag where it is only pulling the answers from the data set that you give it, you should it should be expected almost in my mind that it should return where it got the data from, right? Should site the sources. And so that's a it's an interesting it's a fascinating area because there's so many different you know you've got the existing incumbent Postgress kind of databases trying to get into the vector database game then you've got all these new vector databases that are only vector databases and it's like can you survive as just a vector database because most people need more than just a vector database. Yeah, it's [snorts] pros and cons and like that they'll say, well, you know, I can get you speed and everything. Then you're right, but then you know, our whole thing is about unsiloing data. Yeah. And if if you take nothing away from like the problem we're trying to solve is we we're trying to break down all the data silos and it's it's really hard. That's an area where, you know, like Hadoop would have said, oh, you need a different date, you know, thing to store this image or this PDF. And we're saying, well, actually, you could have it all in one platform. And so, from day one, I'll just throw out, so, one of the other areas where that really helped us in as we grew was we could store structured and semistructured data so JSON files parquet, in the same manner and, and so, and there was zero performance degradation. And so, that was a big breakthrough, because some of our early customers were like video game customers that were like, "It's spitting out huge amount of JSON," and and everything like, "Well, we can't query this and our Hadoop is slow." And we'd like, "No, we just load it into Snowflake, query the JSON in place, and then you can address every tag in that." So, think of like a like a regular table that has like 10 columns. You know, you think about like how could you write that in a way that I can retrieve those columns and then as JSON maybe it's got a hundred different tags, but all we do and when we load it is we take take every tag and we sort of colarize it. Okay? So, we sort of present as a column or database, and so we were able to process, you know, thousands, millions of JSON and parques and everything at the same speed. And so, a lot of our early customers were moving Hadoops or, like, solving problems where, like in the ad tech space where also where you're trying [snorts] just massive amounts of data and just being able to just query it super fast and it just worked just like you would expect. you know, we extended SQL to where you'd have like the name of the column, double colon, the tag name, but otherwise performance is [clears throat] good. And so you could put a view on a table that has JSON and then point it at PowerBI or Tableau or Hotspot or Sigma and it would just work. And again, so that was another, you know, area where people migrated to us because we just were very, very good at that. Yeah. I mean, and I think that's probably another part of what made the whole ELT paradigm, you know, more doable. like I just put it in how whatever format I have and I still [snorts] have the granularity but I can get it and it doesn't just have to be rectangular data. It can be that more unstructured data. Yeah, I mean I think well I've tried and failed but I mean messing around with XML was a little more crazy. I think the lateral flatten and everything was a little a little hairy but yeah I mean technically we do support doing that but yes it's it's it's talking about terrible formats. XML is a terrible format. JSON is is much better. So I would always try to encourage people to get away from it. But back to your other point of like how can you survive if you have to have a different silo of the data. I think it is going to be hard for some kind of companies and they're trying to find you know their their all of them are trying to find their niche. No, I think there's this weird like especially in tech you know early on it's just like the tinkerers and just you know us nerds playing with new [ __ ] because it's new, right? Like only because it's new, and then you know you go through the whole, uh, crossing the chasm cycle. But there's this weird thing about a lot of new tech and the fact that it's like it feels so easy to the user that they don't realize that, like, hey pine, cone would be is a good, you know, uh, vector database if I just am trying to spin up a very specific RAG model for a very specific thing, probably not largecale production enterprise grade stuff. But if I'm just trying to do that internally, not a problem. But then you start throwing more stuff into it and there's more context overlap, then it's like, oh well, now I need all this additional data that I can't even put into fine cone. So now I have to set up an actual database and then link them. Even though they don't want to be linked, it's just this messy. You're you just I see it and it's just like, oh, okay, now I see why. Because it's like they make it easy to quickly go look like you've achieved something, which you know the value of that is whatever. But then it's like that jump from that to a production scale version is so big but, you but they make it look like it's going to be really easy because that first little thing, that iteration that you did or their example, made it seem like, "Oh [ __ ] well, I can talk to all my documents now, that's awesome. I can do this with all my documents," you know? It's like it's very different scenario once you start getting into production worthy type type software and stuff like that. So I'm very curious to see how that works. But I feel it's the same way with the language models as a whole, right? Is it's like GPT has put out all these products or open AAI around building your own GPTs and your rags and your own agents and stuff. It's like again at small scale those I think work fine but they're not it's very hard for someone to go out and productionize something like that into you know the entire community to make it worthy of people paying for it and things like that. So it's it's this weird like dichotomy where it's like the facade is oh it's easy and everyone's doing this and everyone should be doing this and then the actual thing from an you know enterprise corporate perspective is like wait a minute how much is this going to cost me how can I monitor it you know That's why, like, RAG, in my opinion, one of the reasons why RAG has is, and I think still will be, very popular is because, hey, tokens, right? Instead of training a custom model that's just an endless, neverending process and expense essentially, because you always have new data coming in, so you're always having to update it or [snorts] we could just do a RAG where we just build it one time, we don't have to have all this custom training and spend millions of dollars doing it and have the compute resources. Even like, that's going to be a harder and harder thing, I think, moving forward, so it's but I think that comes back to what he was saying earlier you guys have the compute resources they're just sitting there waiting for you to use them you know like and but you're not paying for them until you access them exactly Exactly. And and so fine-tuning is another [clears throat] area where that that potentially fits in there. We now have fine-tuning of one command, you know, one SQL statement. You can fine-tune one of the models to kind of do more what you want. And there's this also interesting thing whereas maybe for the like let's say you're trying to classify something with the using a model, you might use like something like a mixture large to train get the output and you're then train a cheaper model to go hey this is what I want to see. So you're doing a fine-tuning and then using the the cheaper model so with a lot less tokens to do your day-to-day things. So there may be a process and the benefit of having some of these other models there is you can try them out and see which one has the best result but you might use the output of one to save you money downstream. So it's real interesting. Yeah, there's a weird It's not weird. It makes sense when you think about it but it's fascinating on the language model side. It's either really big, getting bigger all the time and more, or it's complete opposite. It can run on a Raspberry Pi. You know, everybody's going one of two routes. And I was talking to our intern about this yesterday, and I was like, I genuinely think like the way things are going, you can you can kind of see how it's like, oh, well, OpenAI can't keep running all of this stuff on their own stuff forever. Like, that doesn't work. And then you think about like, oh, Apple's got I don't know how many millions, hundreds of millions of devices worldwide. And then I was like, well [ __ ] Apple just rolled out an IO, you know, an iOS update with built-in onprem language model. They now have the biggest distributed computing network in the world just like overnight basically. And it's like, oh, okay. So, the idea here is distribute the compute to the user so that they don't have to pay for the cloud compute resources to run and maintain it. That actually makes a lot of sense as far as like it combine that with all the edge compute stuff. And it's like, okay, I can kind of see that now. But it was really weird for me initially when they first started rolling out these smaller models like why are they going smaller? But then you like, oh yeah, so it can run on your device and not theirs. Yeah, Arctic Tilt is considered a smaller model, but it's a but it's like I said, it's very effective, but it knows doc it can pull apart a document. Like it can get the handwriting off of a document. And I was like, how did it do do that? But it's just like the way they trained it, the team that did it. That company was called Applica we bought. Okay. And is now it's part of the guey. So that's kind of our model is we're we we buy companies, we roll in the tech, and we stick it as part of the guey and just try to make it available and and we're a consumption model. So you pay for it if you use it. If you don't use it, you don't have to pay for it. Another example of that, we bought a company called Simuha. They're they do clean rooms. So if you're into clean rooms, there's now it's all part of the guey. You can just turn it on and uh then we integrated other partners into that. So if you need to do like identity resolution of who this person is, then you can choose from a number of options and maybe you're already using one of them and just slot it in and then you can compare. And you see this sometimes with mergers. It's like there's maybe I can't see all the data. You know, the sure the government's going to say, okay, maybe you have a small group of people that can see things, but most of the people can't. So you can agree on like where we're going to allow you to query data and see an aggregated result, but you can't see the lower level PII. So, so Disney and NBC Universal, they're building their, you know, platforms on on Snowflake because then now they can share with other people that, you know, content or whatever. Like I need to to to know how to target somebody that, you know, right, went to a Disney park and watched Moana or something [snorts] and then they can say, well, I'm not going to tell you who they are, but here's some slots where you can target those people. So, yeah, there's a lot of really interesting things happening there. Nice. So I think one or two more things I want to get into with Snowflake before we you know um because we're already getting even close to an hour. It's crazy. [snorts] So this is three hours, right? Yeah, exactly. Um we have had those requests is uh so I mean again again I think a lot of people think of Snowflake and they think of SQL obviously, right? You know because like a data warehouse that's what it originally was but can we talk about like snowpark and then also one thing I'm excited about is the new like snow park pandas I think where you all integrated like you know moden and stuff like that. So you know what those are and then maybe maybe if we have time to get into some of the data ingestion the snowpipe and the streaming and all that but Okay. Sure. So yeah so originally we you know we have an engine and uh how I think of it is like there's it
SQL is nothing runs it natively right like at the end it's getting down into ones and zeros or whatever. So we have an engine that you know most people think about it just could be SQL and so it compiles down into the bite code that goes and runs [clears throat]. So that's SQL but there doesn't have to be SQL and so we've made versions of that for Python, Java and Scala to but it's still the same engine. So it looks to you like oh Snowflake's running a bunch of different languages. It is but it was always compiling down into something else and then running. So that's how this works. So we have SQL we have Python. Those are I think probably the two most popular, but we also run Java and Scola. And and here's an interesting thing about how we've done this and and people wondered, well, why did Java come out first? And it's because we're hyper, you know, aware of security and things like [clears throat] that. And so one of the things that Java had is it's naturally sandboxed. Yeah. So you could say, hey, you're not allowed to go do these things. And and Python does not have that. And some enterprises get very worried because in a world where you can pip install anything from some dude who wrote it in on the internet, it just works, right? If you go on your laptop and you say pip install this, yeah, homebrew that, you know, you're it's going to pull it down and it's going to run. And we had a very famous case from a few uh months ago where somebody modified an open source library and almost, you know, probably could compromised every, you know, Linux machine in the world. Luckily, at the very last second, same kind of thing of the book you were talking about earlier where, [laughter] you know, uh, somebody was like, "Hey, why is this a little slower?" And then figured out, "Oh my gosh, there's this thing that happened." Oh, man. So, we spent and this took us a long time. Uh, but we our Python runs in a sandbox. So, you can't mine Bitcoin or, you know, xfiltrate data, you know, in in in those ways. But otherwise, it works with Anaconda and you can have your own libraries and do all that. So it's in a more enterprise friendly way of running Python. Uh we try to keep as much of the syntax as as you can. So you can move very easily. Our our syntax is sort of bit based on pispark. So if you came from the pispark world you should be very comfortable doing that. And now you get into what you just asked which is okay how do I get performance on all these things? How do I take advantage of multi-mp compute? So, you know, by default, we're an MPP database, meaning, you know, we can if you have one extra small machine that's one thing (although maybe there's eight cores), but uh, but we can have up to 6 XL, which is 512 machines, but certain things will parallelize and certain things will not. And so, we uh purchased Moden and then we've integrated that in. And so, now there it just makes it so where certain things can run a lot faster and parallel parallelize better. And I guess for people who don't know about Modden, it's just a Python. I mean, not just I mean, there's a lot going on under the hood. But at its core and I think I first used it at RDS, we worked together previously but it was like you just take your pandas code and, with like what, one line of code, you just install mode or whatever, and just your pandas works, but it goes way faster, uses all the cores, and can parallelize and. Yes. Um but then but now you all have that in you know like you can run it like within a Python notebook or snowbark pandas and it it will just automatically paralyze your pandas code or Yes. Yeah. And so that's what we're constantly taking feedback. So uh we would love to hear [clears throat] how people are using it if things they like things they don't like but that's the idea is it should your code should run the way you expect it to. It should run fast it should just work. Uh, one of the things that is coming in now, uh, I just...it's very hard as SE to keep, keep up on everything. And we're sort of, they're breaking us internally into colleges, and I'm in the Data Science college. But, uh, you kind of have to know about everything, and we're getting to the point where it's, it's hard to know about everything. But one of the things that, you know, I just took training on is observability of all these things. So as you have uh store procedures or or these things how do I you know know like am I running out of memory or this or that. So we have a concept of event tables. Um and there's one in your account. There will be a default one that will just be there but uh now you can create it. Um but what we're doing is we're making it much easier for you to see and we're using uh an open-source uh observability format uh so that if you have like a graphana or or data dog or something like that that can read those formats, it'll automatically work with those tables. Uh, but we also have it as part of the guey. So you can see, hey, I've got a, you know, here's my my process, but it spun out these eight different processes and how long did each process take and where can I go debug these things? And so we're making that significantly easier within the platform. Um, and so yeah, happy to get any kind of feedback on how people are liking that or there are more things that they want, but we're we're making that much easier. You asked another part of the question which I don't remember now. Well, I mean, I think the next part was just maybe if we want to get into like some some of the data ingestion, um, side, you know, you know, just maybe high level: What is Snowpipe? And then I know recently the Snowpipe streaming I think people are getting a lot of, kind of, realizing a lot of even cost benefits from that, and. Yes, and I think that's how you found me originally. When I looked in my message thing, you had some questions about it. So, so basically, so what Snowpipe is, is serverless ingestion. So, uh, what we mean by serverless is there's a concept of a warehouse in Snowflake, and this is like the t-shirts you were talking about. So extra small or whatever and so the price to beat is the the warehouse price. So let's say you're running at enterprise level and it's $3 an hour. Mhm. So if you're continually sending data to Snowflake [clears throat], if uh your serverless process is more than $3 an hour, you wouldn't use it. You would just switch to a warehouse. But the idea is if the warehouse was $3 an hour, uh what Snowpipe does is it charges you per the size of file and the number of files. So imagine a case where that is now a $150 or a dollar, you know, that's significantly cheaper to move to serverless. You should switch over to that. We also have connections to things like Kafka and Kafka will also have different parameters. So Kafka, I think is, you can say number of rows, time, or number of bytes, something like that and so you can tell it to, "Hey, build files of a certain size that make it so that Snowpipe is, you know, very .costefficient" But when you start getting into telemetry, real-time telemetry, the number of files that would be so many of them that that your Snowpipe costs would be too high, and so you'd say, "Well, you could go back to a warehouse." but customers didn't necessarily want to do that they still want a serverless way and so that's where snow uh snowpipe streaming comes in. So it's basically an API and because also because of our size and scale all all of our partners you know will also support these kind of things. Um this is why for instance on the BI front if somebody says oh I have PowerBI or Tableau or Sigma or whatever uh you know it just works because so many there's so many customers on the platform that they want to make sure that their things work well with us. Same on the ingestion. So we would say uh yes you need to stream the data in you can call this API and then behind the scenes we're writing it we're storing it in a different way but then we're writing out the tables that you would normally expect and you just don't have to worry about it. You just call the API writes it and that could be you know everything's different you have to look at your profile but like it could be 10%of the cost of snowpipe. Yeah. So, it's, it suddenly we're seeing this huge shift, and you can move gigabytes per second or something. And I, and I don't know, like ours, I think I could have the number wrong, uh, but something like, I think we load like 70 terabytes of, like, data security logs a day, like, it's, it's insane. But, like, so because you think of millions of machines, you're looking at when do they start, when do they stop, what happened, so you're writing a bunch of logs, and it just, we're just absorbing this huge amount, which is why a lot of companies do use this for their cyber security data links, right? Because again, it's like, do you want to have regular data with your cyber security data? It sounds like traditionally that's been two different things. Like maybe it's a Splunk, but then there's areas where the things come together. It's like, hey, maybe that is normal. But if I have to know that this employee why does this employee who left our company two months ago clicking a link that, you know, like maybe that's going to, you know, that should be an event I look at depending on whether it was a somebody hadn't left kind of thing. So there's a there's a good reason I think to have your cyber security data next to your structure data because there's times where you need to kind of have a look across. So um so that hope I know hopefully that answers on streaming. So it's a way to very efficiently move in time series data and there's so many things happening there. I made myself some links. So MQTT I went to the intellect cons conference here in Houston. It's a lot of like IoT and how do I move data over and stuff and so there's like uh there's a guy Arlon Nipper. He he I he was part of the MQTT standard, and they have things like Ignition and Cirrus Link and standards like spark plug where you can just like turn on, um, you know, a remote site and it sees all the tags, and the tag it it builds up basically what your plant looks like, you know, just as part of this standard. Because the kind of the tag has the metadata, how it relates to other things, and so you can suddenly see it. And if later that you swap out the sensor, it sort of reregisters and it says, "All right, now this thing is gone from version one to version two, and maybe has these other properties that it didn't have before." So, it's it's kind of cool to see all that. And so, yeah, we were there as part of part of that. Very cool. The uh the security logs made me think of [laughter] the book Cuckoo's Egg because that was uh what was it? He would go in and he would he would find like old emails that hadn't been logged in and then he would log it. he would hack their password and then he would go into that account and then use that account which is now a permissioned normal account to create his hacker account that was outside of the network. And it's like yeah had someone just been paying attention to the fact that these people who left the company six months ago are logging in and doing creating users that never would have happened. [snorts] [clears throat] So, I think we're getting getting close. Like, one thing I want to do before the speed round because we didn't really get to talk a lot about use cases but I would say, you know, is there you have like a kind of a favorite use case that you've worked on it in like energy or oil and gas? Like, where like... or kind of a unique one where like I never thought about using it snowflake in that way. Oh there's there's so many. I'm just trying to think of one that I one of my customers worked on. I mean I love the the takeouts. [clears throat] You know, to me it's like where somebody had used Hadoop and they struggled with Hadoop. Yeah. In our region, one of the first, and this is where it gets into like, you know, oil and gas is a tech world, right? I mean, you're like, you're drilling deeper. You you have to have a lot of tech to be able to do stuff and directional wells and all these kind of things. And so, like a Devon and a Chesapeake were in our region, the two first snowflake us. I think they joined in 2018 or something. Referenceable customers. They got case studies that's there. But like Chesapeake, they push us in certain ways as well. So to give you an idea, they we're their historian. And I'm not saying that we're their historian. I wrote a blog and maybe this is how you found me like a couple years ago like you know, you know, is somebody had asked the question, is Snowflake the best time series database in the world? And and you know, we're not a time series database, but we're really good at time series use cases. And that's what Chesapeake found. It's like as they move over these massive Hadoopes, they're like, "Wait, I can query this at scale and I can have all these tags and it just works." Um, we were they're like, "Hey, we're not getting the benefit of our historian." Snowflake's our historian. So, we're like, "Whoa, wait a minute." You know, another word. Yeah. But, uh, and so they pushed us and and so you recently you've seen some new features come in like as of join. So you know things you would see like in a time series database that we now support because customers are using us in these ways that it's some cases stretch us and then we're say yeah that's something we should be doing and we we put it into engineering and then you know we'll come out and they'll test it and give us feedback and things like that. So I think those u are some of my favorites where you know we've we've solved a problem they had you know they struggle with Hadoop or something like that and then suddenly everything just works and people are happy. Uh I did a um I love to do hands-on labs. We call them zero to snowflakes. We did one for a big energy, uh, company. And, um, they they had this massive Hadoop cluster and they were moving away from it. And I, I, we did the zero to Snowflake. But then they also have some of their own data there, and this is when, kind of, the early days of our Iceberg support. And so, this is some data they're like, "We're never moving into Snowflake." And people ask the question, "Well, why would Snowflake support Iceberg?" because there's some workloads that customers say, you know what, I've got it over there. It's too much. I don't I don't want to move it into Snowflake, but I'm struggle with the compute. And we want customers to be able to query that data and have it perform well. And so, we ran this query on their data at the in this session and it sits there when you're in front of a crowd of people and it's taken like three minutes and I'm like I'm up there [laughter] going, "Oh, you know, well, you know, I don't know. I need to look at this." And part of it is like the sizing of the files and everything is like, you know, because they they're in control. In their case, it was unmanaged iceberg. So there [clears throat], you know, they have to control the size of the files. And it came back after 3 minutes and they they there was kind of a reaction and I'm like, well, you know, we can look at the tuning and they're like, you don't you don't get it. Said this query would have run six hours on our cluster and most of the time it dies. Yeah. And you just ran it in 3 minutes [clears throat]. This is like amazing. So those are the kind of reactions that I like. So, I I think some of those where there's a big takeout where the customer's happy and um that just a good feeling when they say, "Hey, I've got 99 problems in it, but Snowflake isn't one of them." Yeah, absolutely. I mean, just just being able to take customer feedback like you were talking about about the historian. I know almost every oil and gas customer that I have personal relationships with hate their their historian. And it's like that is the incumbent historian that has been there. You have to move it off of your story into something. Exactly. And it's like, yeah, that's exactly it gives me the exact same vibes that the vector database and database stuff gives me. It's like this doesn't make sense that we have to do it this way and at some point someone's going to come along and figure this out so that you don't have to do it that way and they're going to win. And so it's just Yeah. you know, SCADA systems, like that was always one where I thought, well, oh, SCADA systems will never move because like like if the plant blows up, if we can't get the data, I don't want to be that system. But there's some flavors of it to where it's like, okay, but most stuff could run out. Yes, you do need shutdown and safety and things, but we're starting to see that as there's SCADA companies moving to the cloud. So, people want to have access to their own data. They don't want to, you know, be say, "No, no, everything should move here." It's like this is a a a conversation that I had with one of those data providers. They're like customers were asking them for shares and we collected a list. I want to say like hundred customers are saying they want a live share like but don't they know that there's just a very easy API to call and they could just download the data. It's like but it's not as easy as just being able to query the table. And so eventually they get you know most API suck the timeliness and you have to babysit. That's another thing you have to babysit you know like no it's Yeah. Then they change they get they update their off to from zero to one or two and then you have to rewrite the whole thing. That's like Bobby turned me on to fiverran. I know they use fiverr for that very like yes fiverr then keeps the pain of that because they keep up with that. They manage the APIs. It's like hey I don't have to rewrite it and it just works. That's that's what most people want to pay for fiverr. Yeah. Um, Well, but even then 5T is a fraction of what it would cost me [clears throat] to have a data engineer. Just, but I mean like, I mean one thing we didn't really get into and again, we're running up on time so I want to go too deep but I mean I know that you guys have some native connectors now too. Like, I think you can connect to my SQL and Postgress, and I think they're working on SQL Server is what my SE told me, but where you can just point Snowflake at your SQL servers and it just ingest directly. it just or my SQL will never cease to blow my mind that so many tech companies don't understand that you're only as valuable as how your clients are wanting to use you. And if you can't integrate your product with their workflows, they will find a different way or someone else will build it integrated. Like it's there's only two options there. And it's [snorts] like yeah, you might be the incumbent for 101 15 years, but at some point someone is going to get frustrated enough as an end user that they're going to try and figure out how to solve it themselves. And for sure. Yeah. Yeah. So yes, we do house connectors. Uh there's some that we own, some that other people own. There's native connectors. Uh so uh Service Now, um Google Analytics 4, there's a connector, you've probably seen it up there. Uh so yes, in public preview, my SQL and Postgress. So it's the first thing that really steps us out of the cloud because it's a Docker image that runs on your site that can read your it can read the logs of the database and then push it to us. And this is just customers pushing us to like, hey, we need to do this and you guys should know this better. And so again, it's I'd say, hey, get Fiverr. They do a really good job of it. Click has a tune or used to be a tunity replicate. Now it's click replicate. There's other there's things that do it really well, but they're like gosh, you could just do this one thing standards. So we have done some of that. You probably see some more coming over time, but for the most part, you know, it's just we're working on the platform and yeah, we've got a lot of great partners and solve the problem in the cheapest way you can. Absolutely. Like perfect example. I am now going to go look at that connector because I don't have to pay Fiverr to to replicate it over to to Snowflake instead of which is ironic because you use have been using BigQuery and but Google Analytics. I also like that was a thing. So BigQuery, you know, we would I think we want to work better with Google. So I would say if you look at it, you know, AWS is where we started. That's where the most of Snowflake runs. Uh Azure is a f huge fastest growing what I don't know the exact percentages. We've talked about it on earnings call, but let's say it's 25%. So it's like 74% 25%and like 1% Google. And that's because Google like some cases, yes, we have an engineering relationship, but they're very much like, hey, we're going to do it. And so that was a case where customers are like, I don't want to have a big query, but you're making me have a big query for Google Analytics 4. So we this connector essentially runs all that. So you don't have to have your own big query. But we would have much rather had a live share or something like that. It just we didn't have an option for it. GCP is such a weird company. Just like hard stop. I feel like they do some really interesting cool stuff and then they do some really dumb stuff that you're like, how the hell did this happen? When Llama 2 came out, for example, they came out with Vert.Ex Studio and all this stuff. I was like, "Oh, hell yeah. I'm going to go spin up a Llama 2 instance." I couldn't because Have you seen Gemma? Yeah. Well, I [laughter] haven't even tried. I just quit. I just quit trying to spin up anything on GCP because I found through Google support forums that, oh, GCP doesn't have any of the resources that will run these models. Like, then why did they roll the damn thing out? Like, it doesn't make any sense. [laughter and clears throat] You're just making yourself look bad. Like anyway, GCP rant over. I still use GCP all the time, but you want to do Yeah [snorts], um we just kind of do like a little lightning round kind of thing, but nothing nothing too crazy. But since I've been talking about books, I'll go what's your favorite uh either tech or or just book. A book. A book. Uh I I just got the new Elon Musk doc uh uh thing. Um, uh, I haven't read it yet, but maybe I'll read it on my my trip here. Um, I I I have all I like all the old books like, uh, because I'm a data warehousing guy, like data warehouse toolkit, the R Kimble books and, you know, Inman, corporate information factory, those kind of things. That's been a while. That's many, many years. That dates me, but you can see I got some some gray hair. Um, yeah, like I mean that that's that's the first things to pop into mind if it's a workbook. Yeah, that was I'm glad I'm not the only one listening to books like that. My wife, we were driving over the 4th of July and I started listening to that Cuckoo's Egg book and it was, you know, like 11:30. We got we ended up pulling up to the house at like 12:30 at night and she had fallen asleep and she woke up and she was listening to the book with me. She's like, "Do you listen to this to stay awake?" [laughter] Like, now that was that's a classic one. Another one is is it hackers? I can't there there's in that vein or whatever, but uh yeah, that was a great book. I'm glad I'm not the only one nerding out to uh fun stuff. Um and I guess I'll finish up. What's your favorite o uh open source library?
We come with the hard-hitting questions here. Yeah, you [laughter] can you can edit out my pause of whatever. I mean, you know, just just recently cuz I was looking at it. It was uh PI PDF because I needed to that for the quick start or there's another one. Is it uh is it tesseract or I forget test OCR? Yeah. Yeah. Because somebody said, "Oh, I have this PDF." But then when you looked at it, it it looked like text, but it was really images. So, you have to like use Tesseract to pull it out. So, uh those are all part of our Anaconda distribution. So, you don't have to do anything crazy. You just have to go into your thing and say add in this one and then you can start using it. So that I'd probably do those. Yeah. Well, sweet. Thank you so much. Thanks for taking time out for us ahead of your awesome trip. So, no worries. Uh I'm uh I'm I live nearby so you know if you ever you know we can if we didn't get to everything you wanted to. Happy to come back any question. Perfect. Sounds good. Thanks for uh joining us. Shout out to Zach and the guys at Forran for the or forecast for Forecast. 4chan is a very different organization. I thought you mixed Fiverr and for 4chan and Fiverr. I did. That's exactly what 4R was a very old language. Uhhuh. Um for uh for the swag and uh we'll see you guys next time. Appreciate it. I tell.