Orchestrate all the Things - Lightning-Fast Python for 100x Faster Performance from Saturn Cloud, now available on Snowflake. Featuring Saturn Cloud CEO / Founder Sebastian Metti

Episode Date: December 16, 2020

Python, the most popular language for data science and machine learning, gets a huge boost from Dask, an open source framework for running it in a distributed way on top of GPUs. Saturn Cloud, a startup offering Dask as service, is now a Snowflake partner, making Dask available to the masses. We discuss all about Dask and Saturn Cloud with co-founder Sebastian Metti. Article published on ZDNet

Welcome to the Orchestrate All the Things podcast. I'm George Amatiotis and we'll be connecting the dots together. Python, the most popular language for data science and machine learning, gets a huge boost from Dask, a framework for running it in a distributed way on top of GPUs. Saturn Cloud, a startup offering Dask as a service, is now a snowflake partner, making Dask available to the masses. We discuss all about Dask and Saturn Club with co-founder Sebastian Metti. I hope you will enjoy the podcast.

If you like my work, you can follow Link Data Registration on Twitter, LinkedIn, and Facebook. Well, welcome. And a typical way I start this conversation is basically by asking people to say a few words about themselves and the occasion and the company they're with and what they're announcing and that sort of thing. Sure. Yes, I'm Sebastian Medi. I'm one of the co-founders of Saturn Cloud. We are a data science and machine learning platform in Python. And our focus is to crush the most difficult problems in data science,

which is a lack of speed and having hard to use tools. And so our platform offers 100x faster data science and machine learning, all in Python, which is the most preferred language in data science. Okay. And that's already interesting, but actually I think you left out a part which is also very, very interesting. And the occasion is that you're partnering up with Snowflake, right? And that's correct.

Yes. So we've been working with Snowflake on integrating our products and having a joint go-to-market effort where we're scaling across their customer base and they're scaling across our customer base. And the reason why it's significant is because they have a massive presence in the cloud data world and cloud data science. And so we're going to enable their customers to enhance their SQL analytics, to go into Python and do really the advanced analytics for machine learning, which will be again, up to a hundred times faster than previous tooling that they've had access to. Okay. Okay.

So we'll get to the specifics of the partnership a little bit later, I guess, but since we now have a kind of at least fundamental understanding, let's say, about Saturn Cloud. I was also taking a bit of a look around before the conversation and trying to figure out one liner, let's say, of what you do. And obviously, you described it better than I would. I mean, you're the founder after all. But one thing that kind of attracted my interest was that you explicitly mentioned Dask in your product description.

And for people who may be listening and are not necessarily familiar with Dask, Dask is a framework for distributed Python, to put it as simply as possible. And I guess this may be the reason why you mentioned that, well, you do Python faster than the rest, basically. And to me specifically, it was interesting because I'd like to think that I kind of had a little bit to do with popularizing tasks, because at some point, maybe it's been already like a couple of years by now, we had a very interesting conversation with the person who's leading or at least was leading back then because I think now maybe he's more into other things,

but the person who originally founded Dask. And it's a very interesting project. And yes, indeed, the whole idea behind it was to simplify distributed python so would i be too far off if i said that perhaps what you do is like dask as a service that that's a great way to frame it um the yeah the simplest way for someone to understand what saturn cloud does if they're familiar with the open source world, is to think of us like Databricks for Dask. Dask as a service, right? And so Matthew Rocklin was one of the creators of Dask.

And he worked at it while he was at NVIDIA and Anaconda. And as an open source project, as much as they can do in the open source form, that's hard for enterprise to adopt sort of just off the shelf. There's a lot of tooling needed around it, right? And so that was our vision was to say, companies are not effectively analyzing their big data as much as they could. We had seen companies waiting one week for their model to run. And if they were to have an enterprise version of Dask with GPU acceleration that NVIDIA provided

to the Rapids project, we could take that one week and make it in one hour. And so Dask had been around for several years now, but companies were adopting it at a certain pace. We saw a path for an enterprise adoption that could be much faster. Okay, okay. Good to know that I wasn't entirely of them.

So when did you first come up with this idea and who has been with you in that journey? And actually, where exactly are you at that journey today? So can you mention a little bit of company background? Like when did you start? Who did you start with? did you start with, what your funding is, headcount, this type of thing? Sure yeah, so we started a little over a year ago, so we're fairly new. My co-founder Hugo Shi was one of the co-founders of Anaconda, which is one of the largest data

science platforms and really took the on-premise market. And the vision sort of was a result of a lot of experimentation, talking with different customers, his vision for a world with faster data science tooling, and the right timing between cloud adoption, adoption of Dask, Rapids, open source adoption. And so all of these sort of stars aligning simultaneously with the Python data ecosystem being mature enough to really sort of sustain something that wasn't Java or Scala based. We started last year pretty quickly after our founding, we closed a seed round,

$4 million with SignalFire, a large VC there in the Valley. And since then, we had scaled to now about 20 people. We launched the product, really gained traction launching it in summer this year. And as a usage-based payment model, we don't sell subscriptions. We just charge you for how much you use. And so since launching now, we've seen over 200,000 hours of compute run in which things like GPU computing, which is the fastest of the fastest, starting off at 10% of our total compute and is now sort of peaking at 30% total compute. So the speed is really, really resonating with the market beyond just sort of data science adoption.

Okay. Yes, you did mention your business model, which indeed seemed interesting to me, and I wanted to ask you about it. So since you mentioned it now, now's the best time to ask. So one basic question that i had was well in some way it seems like you're kind of reselling compute let's say with your additional layer on top so one thing i was wondering was like okay so that's that's straightforward enough is it possible to add on-premise compute to the

mix so for example if, if I have a data center sitting somewhere in the background doing nothing, can I run my own task workloads there and add some of the services that you offer? Of course, it's possible. Yeah. Any of that would be possible on-premise, any cloud. Today, we're partnering with AWS, and so we're really focusing on that angle. So all of our customers that adopt us will either spin up something on AWS or they currently use AWS. In the future, we do look at expanding to other clouds and other solutions if it's on-premise. Okay. So if I got it right, right now you're on cloud only and actually specifically AWS only, right? That's right, yeah.

And the reason why we pursued them was because in three years, I want to say, after the founding of Databricks, they were really gaining traction. And then they partnered really closely with Microsoft Azure. And that was a huge launchpad for them to accelerate their growth. Similarly, we've pursued the same with AWS and have now started to reach the highest form of ISV partnership, which is what the sort of tender with them get. And that means we can offer free AWS credits to people that are trying out Saturn. We can offer AWS support and expert advice,

a whole number of things that just take out the pain in trying a new platform. Its billing is all done through AWS. So the data scientists don't have to worry about procurement. It's just pay-free, 10 minutes, and now you can start building your model. Okay, okay. Actually, those are both important points

because that was going to be my follow-up question because, yes, you did mention that your business model is not subscription-based, so it's pay-per-use. But then I was like, okay, yeah, that sounds good in theory. But what happens if people, for example, want to control their bills so they don't get charged above a certain threshold or all these kinds of things? But I imagine since you are, well, A, an important fact that you mentioned

is that you are actually partnered with AWS as well. And B, since you are, it means that people are able to, for example, to use the tools they use on top of AWS so they can have one common billing and all of these things, right? Exactly. There's a lot more cost transparency that we can provide that aws may not provide an initial interface so we can show you know teams who is spending the most at your company who is using you know resources uh this way or that way and so then they can have more transparency

into the bill breakdown that aws provides them at the end of each month. So we're able to recover. Okay, that's an important point and good to know. Okay, so another thing that I wanted to ask about was like, okay, that far it's relatively clear, let's say. I was wondering, okay, so in addition to offering Dask as a service, I also have the impression that you also offer some services on top of that. So having to do with deployment and version control, possibly, and potentially something else that I'm missing. So do you want to elaborate a little bit on the add-on services, let's say? Sure. Yeah, Dask as a service is one of the sort of many components

of our overall vision. So, we try to think of ourselves for our customers as we're the solution for lightning fast Python for data science and machine learning, where Dask is instrumental to that, as well as RAPIDS, the GPU acceleration framework that makes Dask scale out. In addition to that, we're meant to be end-to-end. We work with enterprise. So people have to collaborate on projects. So you have to have version control sort of capable there, sharing your content with your colleagues. We just want to cover the entire spectrum

from starting a project to deploying it. And so we have customers that have dashboards that are productionized and sort of different kinds of deployed models, which is meant to provide the end-to-end solution. Okay. How much of that end-to-end solution that you mentioned would you say that you cover at this point? I think every month we finish a sprint and we're saying, okay,

I think that was the whole end to end enchilada. And then you discover, Oh, there's an MLOps thing that you can add. And then there's this, and then there's this. And so it's, it's, it's never ending almost. And then you see companies like Databricks, which is like, they pretty much flushed out end to end. And now they launched the Delta Lake product, which is like, they pretty much flushed out end-to-end and now they launched the Delta Lake product, which is like a Snowflake competitor. And so you can continue going deeper into the data or deeper into production or collaboration. And so in terms of the enterprise offering, we cover the full end-to-end scope, but of course you can start going into ML ops and

data warehousing, which is not on our roadmap anytime soon, if that's going to happen. No, no. I mean, that would be a bit overly ambitious, I would say. I was just wondering about the end-to-end flow, let's say, from the moment, I don't know, you build your Python program to the moment you're ready to deploy it in production. Yeah.

So, I mean, from the ETL component to the research and experimentation to the collaboration and through deployment, it's all sort of capable on SoundCloud. And making that a very easy experience is like another layer to that. Okay. You also mentioned at some point dashboards and people building dashboards. It's not clear to me whether this is something that comes as part of what you offer or something

that people can integrate, I don't know, if they're using some other dashboard solution. So actually my co-founder was one of the core maintainers of Bokeh, which is a major sort of Python analytics and dashboarding tools. But we have a ton of other ones that integrate. Since we're operating the Python ecosystem, the open source tooling is as lively

as it could ever be today. And so we can tap into any Python centric tool for dashboarding and all that. So Bokeh, Tableau, you name it, they're all sort of being run in production with Sennheiser. Okay, I see.

And, well, speaking of integrations, I guess that's as good a point in time as any to talk about the upcoming one with Snowflake. So, but actually maybe before we do, I'm wondering, for a company as young as yours to be partnering up with someone like Snowflake is not entirely expectable, let's say. So I wonder if you'd like to say a couple of words about how that came along. Yeah, I think Snowflake is obviously unique. I guess that goes without saying. They beat all sorts of expectations.

And they have a history of being what I could say is a kingmaker with startups. They've actually partnered with a lot of early stage companies and turned them into super fast growing companies on the path to being unicorns or turning into unicorns. So you have Matillion, Fivetran, Heap Analytics. There's a huge list of companies that have partnered with Snowflake early in their lifecycle and have since just become wildly successful. And part of that, I think, is integrating products with the largest cloud data company, as well as the go-to-market effort that you can do collaboratively with them. A lot of large companies move at a pace that is hard for startups.

Snowflake is actually founded in, what, 2012? So they still have a lot of that startup DNA there. A lot of people are still there from that, and so they're still capable. Whereas an IBM or another company might not operate at that pace and might have a sort of different dynamic. So I think that sort of explains why Snowflake would partner with small companies is they're looking for frontier technology that's going to last a long time. And so the origins of our partnership was, I would describe it as sort of market pull, where we started seeing our users and their users talking about our respective tooling. We started seeing a blog post surface around the tools that we offer, integrating with

Snowflake. And that naturally led to a discussion with us and their strategic partnerships team saying, hey, it looks like the market's really pulling on this direction. They want to see 100x faster data science, but they want to also do the querying through Snowflake. We should talk and make this happen. And so it was sort of organic that came from that, which was great validation early on. Thank you.

Go ahead. And then, you know, that's sort of your initial customer discovery. And then we started figuring out the roadmap, how to collaborate on all this, working with mutual customers, and then actually putting the rubber to the road. Okay. Okay. Well, since you mentioned, well, this being user-driven, basically, in a way, I wonder if you would be able to share basically what kind of user base you have if not necessarily sharing numbers. I don't know.

If you can share growth or qualitative characteristics maybe. What kind of segment, user segments you have. Yeah. Anyone with a data problem is theoretically a fit, of course. We have to pick and choose as a startup because some companies are going to have bigger demands than others. Today, our customer base is spread across tech companies that are doing research in AI or deployments of AI. We have Fortune 500 companies, and we also have researchers who are fighting coping so

we work with the COVID-19 Alliance it's a organization a volunteer organization of data science machine learning professionals from all kinds of different companies that are joining forces on their volunteer time to use their data analytics provided more insight into how COVID is happening epidemiologically as well as medically. So they're using Saturn Cloud to do that. Then we have these Fortune 500 companies that are doing their customer analytics. Some of them are doing financial trading and so on.

And then the tech companies are doing innovation in AI, in computer vision, in these sort of frontier fields. So it's a pretty big mix. Every day, we're having inbound leads, and we're like, oh, that company is a great fit. We didn't even think that mining would be relevant. But yeah, mining is geospatially intense data. And so they have to be able to parallelize out and scale out their computing. Yeah, I'm just curious because, you know, it sounds like you've been experiencing rapid organic growth, basically. And I guess it's probably because, you know,

typically for companies your size, your age, it happens when, you know, it happens when you're basically addressing something that's like a gap, let's say, and you're also a member of the community yourself. So it kind of spreads. That's what seems to have happened here. Yeah. I mean, we had published a case study,

I want to say a couple months ago, with a startup where they had a model that took over 60 days to run. So they could iterate on this machine learning model four times a year. That is really hard to have that operate. And using our tooling, we've collaborated with them and we brought that down to just 11 hours. And so that's like a huge breakthrough. And I think when companies see 100x faster data science is on the table, and it's just a few lines of code away, two things tell them, wow, we could do so much better. And at the same time, if our competition does that,

they're going to pass us right away because it's so much faster. So there's a lot of urgency, I think, to get that speed, especially when they know the entry barrier is just, you know, pen lines of code or something away. It's not a new language. It's not a new field. Python is everything they've known before. Well, you know, the more I hear about it, the more it all makes sense. And it also kind of begs the question, actually, so how come Matthew Rocklin wasn't involved in that or was he in some capacity? Did he, I don't know, did you work with him in some way? Did he consult with you?

Yeah, so Matthew Rocklin was really working on GASC for the past six years, I think, at Anaconda and NVIDIA. And so building the ecosystem, the community, obviously maintaining the project. And so he's made, obviously, invaluable contributions to it, along with a huge team of other open source maintainers and contributors that we have some of our team, some are still at NVIDIA, Anaconda.

And very recently, Matt Rocklin also launched a company called Coils, which focuses on desk services, consulting, training, as well as having their own desk products. Okay, I see. So, yeah, it sounds like he's obviously part of the ecosystem. He just chose a different path, basically. It's slightly different from ours, but also with startups,

we're converging on paths and then we're diverging and everyone's sort of experimenting and exploring. So we'll see over the next five years. I think companies like Snowflake and Databricks are the ones that are really shaking the market. And so we'll see where they go and the market goes. We're sort of operating on those tectonic plates a little bit. And so we'll see where they go and the market goes. We're sort of

operating on those tectonic plates a little bit. And so we get shaken around by how they decide to execute. Okay. So since we're back on the Snowflake topic again, how would you describe what you bring to the table for Snowflake and its users in a way that someone who's never heard of you before would get it? Yeah. So, I mean, a lot ofabyte data set, which in 2020 is becoming not abnormal as it would be 20 years ago, you want to go beyond that to really do the advanced analytics. You want to see four terabytes and you want to not have to downsample any data. You want the fullest insight. And you also don't want to wait a week. And so our integration works very seamlessly with them, where you can just use a Python connector natively in Snowflake

or on Saturn Cloud to either do the SQL querying or actually do the advanced analytics. And so it's meant to offer a very seamless experience where you're not using multiple platforms and having different screens. It's very integrated in that way. And that enables them to basically go to the level beyond sql uh to do those analytics that um sql is not you know purpose built for as much as it is it's querying and other kinds of analytics python is is really the power engine for

that okay okay well great you you answered the follow-up question that was beginning to form before it was actually expressed, and that was okay. So does that mean that those people, Snowflake users, will be able to somehow magically do that advanced analysis in Python without actually writing Python?

But I guess no. It's good, but it's not that good. If they could do it without accelerating Python, they could use just vanilla Python as well if they wanted to. I would say because they have cloud data, they've probably pursued a very big data strategy already. So I think most of them would want to accelerate it to get that extra performance,

especially since it's not a lot of work. Thank you. Yeah, Makes sense. I think you mentioned that it's not exactly certain at this, at this time when the partnership will be announced, right? Sometime later in the, in the next week, you said. Next week is the target. Yes. Okay. Okay. And well, what happens basically after the partnership?

I mean, okay, so you get a nice expansion to your user base and your revenue stream and all of that apparently and some good publicity as well. But I was more wondering about your own roadmap basically. So first, does that change your roadmap at all? And then what is your roadmap actually? Yeah, I think the partnership is the public veneer of what's happening. And it's a point in time where maybe people notice.

But under the hood, there's obviously constant customer and user discovery that'll keep being pushed and informing a product roadmap. So we'll roll out this integrated product feature and then another one, which will then add a new slew of customer questions and demands around, well, I like the data masking thing you guys just did. Can you also add a component that does this?

And then we socialize that with other users and we realize, wow, that's a huge opportunity right there to go into that. And so nothing will change in terms of the product development and iteration cycles that are continuing to happen between our customers and us. We'll continue to pursue that as well as innovate on different sort of go-to-market motions, offering joint Saturn Cloud Snowflake workshops, webinars where we show people the tooling, how to easily achieve this. We'll target different industries.

This is for fintech and this one's for computer vision. And so I think the partnership is maybe something a lot of people might hear about and notice. But under the hood, there's just constant product innovation happening as well as different sort of kind of customer user events okay you mentioned something about verticals basically so what I guess you're going to be marketing vertical solutions and you mentioned a couple of specific markets like financials I think you mentioned that maybe one or two more. So what's going to be specific about those? So for example, with hedge funds,

they're obviously dealing with all kinds of data, alternative data, standard fundamentals data. And particularly if they're trading and doing stuff in more real time, the demand of performance is huge. And so when we're talking to individuals who are quants and trading, they're looking at really performance is everything and that this can't break. They need streaming real time analysis coming through. And so that's a very sort of special market to speak to. Then we see oil and gas or mining or companies that are working with geospatial analytics.

Dask is perfect for that. It's capable of multidimensional arrays. And that also is a segment that has a tremendous amount of data. They may not need it in real time, but when they do it, they have to get it right because mining in the wrong place at the high level is just a very expensive mistake. And so the better predictive power they have, the more value there is. So that's an accuracy, like hyper accuracy focus. So there's many different ways of approaching how we talk about it in tailoring solutions

to different protocols. Okay. That makes sense. I was just wondering whether... Yes, you basically described how different market segments may have different specific needs. I was wondering if you're going to be tailoring your offering to offer to offer to them something specifically I don't know fine-tuned or maybe potentially then give them some some

data sets that they may be useful to them absolutely I mean we have numerous discussions going on right now with potential future partners where we would have integrated you know data querying tools. So finance obviously is the big one, but there's a lot of companies that offer and sell financial data that our customers are using. And then they'll say, port that from here to here to here and come into Saturn Cloud. Those partnerships would basically look like a direct line to that

through Saturn Cloud, where you almost have a marketplace or some ability to tap into those data sets easier. Snowflake actually has a marketplace for something like that as well. And so it's all themed around just closer, closer integration to the different data products that they're consuming. And so making it less painful to ETL all of that in. Okay. Okay. So it sounds more like, well, integration, let's say, on the organizational levelally connecting with relevant data products. Tailoring the solution itself, I think we're still some way away from that. Although what we've zeroed in on now is tailoring product and documentation to huge data areas like computer vision, gradient boosting within machine learning and particular workflows that we want to be done very painlessly

on Saturn Cloud. You can get in, you can look at our docs, and within a few minutes, start building a model and have conviction that this will actually accelerate no problem. There will be no roadblocks or potholes. In a sense, I guess we're thinking more about workflow

specific ways of tailoring the product that can span across multiple industries, but doesn't necessarily In a sense, I guess we're thinking more about workflow-specific ways of tailoring the product that can span across multiple industries, but doesn't necessarily have an industry taxonomy or categorization scheme as we think about it. What about the organization itself? I mean the company, basically. I think you said earlier you have now something of 20 people, which is actually quite a lot for, again, for a company your age, let's say.

So is that going, are you planning to grow your headcount to be able to keep up, basically? Yes, we are looking for a head of data science and we are also looking for a senior data scientist to join the team. So we're still growing. We have an insatiable desire to add more data science people to the team just because they're capable of working with our customers. They understand them well. They can work with the product and they use the product in our own analytics.

And so it is sort of a, you know, partly dogfooding the product is the phrase, as well as working with customers who want to get to this level of performance. And then, you know, in 10 minutes on the call or 30 minutes, we can kind of show them, all right, you presented a very complicated problem and now it's done this way. Yeah, we're still hiring a lot.

Okay. And obviously you're a remote team, that goes without saying. And so I was wondering a little bit about how does the company culture, let's say, develop and work out? And I was also wondering, because all this kind of rapid organic growth, it makes me wonder whether you have previous business experience, basically. Have you been a founder before? Yes, actually, this is my second startup,

and it's also my co-founder, Hugo Shih's second startup. So we've been in the startup space, both of us, I think since our mid twenties, I guess we got bored of corporate life very quickly. It is, you know, the whole remote culture now, it is very different from the in office kind of culture where you can walk and take coffee breaks together, and there's all that kind of stuff. And so it's a lot of Zoom calls and finding ways to get those long tail of questions asked that you could normally just walk over to someone's desk and ask. On Slack or Zoom, things can get buried or they can get derailed. And so these things can

happen. And so we've come up with different sort of coping mechanisms of saying, let's just have these meetings scattered throughout the week, which if we don't have an agenda for, we can show up and we can just discuss anything and you can jump off anytime you want. And that's intended to capture the long tail of questions around like, I had a question about this one minor thing I asked on Slack. It got buried because of XYZ. So we've done that. And then, you know, obviously culture is trickier on Slack. You have emojis.

That's not a substitute, right? And so we've experimented with different sites of online gatherings and found this tool called gather.town. I'm not sure if you're familiar with it no um it basically looks like your classical 1980s or 1990s game boy uh ui where you have characters walking around on the screen and as you become into proximity with each other um their screen will pop up and you can see their faces and so it's almost as if you're at a bar or at some place together and then there's different games and so that it was that was actually really fun i think we scheduled it for two hours and i think we went over about like three hours or

something like that we also sent you know gift cards where people can order drinks to their house and food and so to kind of make it social, we're sharing food and drink together in a remote online world. Okay. Yeah. Yeah. I mean, I was just curious because, well, since you're, you've been a founder before, you know, that it's always, actually to me personally, I tend to think that getting the culture, the right group of people, you know, in the right mindset and all of that, getting that part right is much, much trickier because it's

intangible that actually than actually getting you know your task or your your superconductor or whatever it is you're building right yeah absolutely and culture is one of those things that once it's set into motion it would be hard to change if you want to make it different yeah you have to be very transparent about your values up front. And the team that you're building, especially early on, the first three, four, five colleagues that are joining you will have a disproportionate impact on that culture. And so things like diversity, inclusion,

this is so important to do early on because that will, first of all, attract a broader sort of world of talent, as well as influence your thinking and the way your culture develops. So we sort of see that as hugely important and something we were very intentional about when we started. Yeah, you know, it just kind of amazes me, to be honest, because even in face-to-face settings,

there's always a huge amount of assumption and projection. And if it's remote, you know, it multiplies by a factor of, I don't know, God knows how many. So the fact that it even works, you know, just amazes me. It is. We have to figure out how to make it work. And then when the vaccine comes, we can forget about it maybe. But for now, we have to make it work.

Okay. I hope you enjoyed the podcast. If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook.