#140 Taming BigQuery costs with Alvin.ai (with Martin Sahlen)

"We identify, find, and execute the savings automatically. It's not a tool that gives recommendations or endless lists of tasks to complete." Martin

"It would be crazy not to try [Alvin]. We designed the entire product and platform in a way that it becomes almost like a no-brainer for everyone involved in the process of buying it." Martin

Show full (AI-generated) transcript

[00:00:00] Lizzie: Hello, and welcome to The Measure Pod by Measurelab, the podcast dedicated to the ever-changing world of data and analytics, with your hosts, Dara Fitzgerald and Matthew Hooson. Between them, they've spent more years than they'd like to admit wrestling with dashboards, data quality, and the occasional Google curveball.

[00:00:32] Lizzie: So join us as we share stories about how analytics really works today and where it might be headed tomorrow. Let's get into it.

[00:00:40] Dara: So joining us on the show today, we have Martin Sahlén from Alvin AI. Martin, firstly, welcome to the podcast. Have I pronounced your surname correctly? That's the big question I have to kick things off.

[00:00:53] Martin: I don't know. I think you, you broke it a bit there. But I... It's it's actually a Sw- Swedish. I'm [00:01:00] Norwegian. The name is Swedish. It has an, actually an apostrophe in it, which I normally don't use because data quality and inputs so generally ill-advised to use. And I actually have a good friend from Norway that has this very long French-sounding name, and he has made his life's mission to basically break forms and then basically shame them online about this validation.

[00:01:22] Martin: Anyways, a bit of a side note, but my name is Sahlén is usually how I say it, or Sahlén. It's a bit hard. Yeah, it's S- Sahlér or with an R or something, or maybe it's like this I don't know if it's Scottish or Irish. I'm bad at this English dialects British dialects. Sorry.

[00:01:38] Dara: You've given me a get out there.

[00:01:40] Dara: I can just claim that's the Irish pronunciation of your name.

[00:01:43] Martin: Yeah, exactly. Yeah.

[00:01:44] Dara: Exactly. An- anyway, listen, it's great to have you on the show. Appreciate you taking the time to talk to us. We kick things off by just basically getting our guests to introduce themselves, so it's over to you really.

[00:01:54] Dara: You can go into, you can go back as far as you like, or you can keep it quite brief. It's entirely up to you. But just to give our listeners a [00:02:00] sense of obviously, we're gonna talk about Alvin quite a lot, but if you could give us just a little bit of your background as well, just so our listeners can get a flavor of who you are and what's led you to the point that you're at today with Alvin.

[00:02:10] Martin: That's a very profound,

[00:02:12] Dara: It's a

[00:02:12] Martin: big one ... very profound question. I think, maybe I can try to cover briefly, the big lines of, I think, what has shaped me and my experiences. So I studied computer science at university. This is now a long time ago. I started there in 2009.

[00:02:30] Martin: And during the discourse of the, in, in Norway, it's a bit different. Like in the US, you have a lot of these like different subjects, and you You know, mix and match a little bit of... I don't know how it is in, in, in the UK, but if you go to the proper computer science, degree in Norway, it's basically like a five-year pretty set curriculum where you may...

[00:02:51] Martin: A little bit in the later years you can choose some additional subjects, but it's a bit fixed. And then throughout the course I guess at around year [00:03:00] three I don't know I had the feeling that I know enough now to... I know enough of the tech stuff and coding to teach myself to acquire new...

[00:03:09] Martin: the goal of university, I think, is to be critical and acquire knowledge, and I was like, "I really don't wanna write a master's thesis on some very narrow ML or AI." It's a bit different now. Maybe I should have done that, but I did-- I don't wanna write about some obscure silicon substrate application in whatever and just leave that to, to gather dust.

[00:03:30] Martin: And then I found that there is this very esteemed entrepreneurship program at the university where I can basically still get the computer science degree, but finish it more like in a cross, cross type background study with a lot of interesting people, very selective, very small classes.

[00:03:49] Martin: So I was like, "I'm gonna do that." A bit like an MBA, and so I decided to do that and that was also more like maybe I can learn a little bit more about the other aspects of of building a business on [00:04:00] sales, marketing. Generally how do you actually build a business and make it scale and work.

[00:04:04] Martin: And that's what drew me into this this I call NTNU School of Entrepre- Ent- Entrepreneurship. Great stuff, great alumni, great network. It's been yeah, profound, to use that word again, I think for me and my career to just be introduced to that mindset and not just do the technical stuff.

[00:04:20] Martin: And that's what led me into entrepreneurship and, company building, and been doing that ever since. That led me to join start another company, join a company which I was part of for many years, led me to move to New York for a bit, and then eventually to Estonia, where I've now been for the last seven years building, building Alvin.

[00:04:43] Martin: So yeah, it's it's like this topic where you can go on and on, but I'll, try to leave it there. And but yeah, always been, like, a startuper, always been a founder or part of founding teams. I think for better or worse, I sometimes wish I'd been like, in a big four doing the boring process stuff because, [00:05:00] it's I'm very dependent on having more systematic people around me that, you know- I can do the R&D, but then you need this really professional people that have the guardrails, that have done the the, this the last 5% that takes 95% of the time.

[00:05:16] Martin: Yeah.

[00:05:17] Dara: So we're gonna get a... Yeah you, a bit like you said that th- this is the kind of thing that could keep going on similarly with the follow-up questions. I'm immediately interested in why, and maybe this is gonna take us off topic completely, but why Estonia? I would guess that was a personal decision, but maybe then if you want to keep this away from your personal life, the question could instead be how have you found running a business in, in, in a country that you've relatively recently moved to?

[00:05:41] Dara: I know you said it's not that recent now, but presumably it was when you set up Alven in the first place.

[00:05:46] Martin: This is, I think, where this serendipity comes into play. When I was in my then company, my previous company, we had someone who applied for an internship. This [00:06:00] is a guy who was in Estonia, and he sent us a very well-crafted YouTube video as his application for like in, s- just like an op- ops person, just someone that could just deal with a lot of the ops, op stuff.

[00:06:12] Martin: We needed that. We, he, we were like, "This is cool. Let's hire him. Let's move him to Norway from Estonia." And throughout this course of the next couple of years, he, excelled. He did really well. He was running all the sales ops. He al- actually moved to New York also with me, so then we were in New York together and then at one point we were like, "Let's start a company.

[00:06:35] Martin: Let's do this ourselves. Let's..." We grew to 40 people, 50 people. It's not the same. You miss that excitement and this early stage type, there's no process. We're just gonna do stuff. We decided to do that while we still were in, in in New York, and then I think it really hit us that we were super excited and then okay, but we're seeing mon- New York is pretty expensive and we realized that we need, [00:07:00] you need time to really flesh out and ideate and you can't rush everything in in, in the early stage.

[00:07:05] Martin: And we managed to scrape together some friends friends and fools friend, family, fools' cash and realized that this is three month of rent in New York with ramen. And we also got, So we put up spreadsheets actually in different places and okay, Oslo. At that point, we got another guy who's my current co-founder actually.

[00:07:24] Martin: That's a different story. He's from the UK, actually. Dan is his name. So we looked at, okay, there's London, there's Tallinn, there's Oslo, they're staying in New York, and it's well- We have never been there, but if we move to Tallinn, we have a year of runway. If we stay in New York, we have three months of runway.

[00:07:40] Martin: So we looked at the spreadsheet and we're like, "Okay, let's just book the tickets to Tallinn and we'll go there." And this was more than seven years ago, and we're still here. Now I have a family here and Dan is married, and it's it's been a pretty, transformative experience. So yeah, that's, that, that's the story of how we ended up in Tallinn.

[00:07:59] Dara: Love it. Yeah, it [00:08:00] works. And one other question I have before we maybe do start leading into the, a more technical discussion. Where did the name come from, Alvin AI?

[00:08:07] Martin: It was a bit of rebellion against what we found at the time to be very boring and technical names in the data tooling space in general.

[00:08:19] Martin: So we were like, "Yeah, let's just make something human and random." I don't know if it hurts us or helps us in terms of, enterprise sales and these things. I think at this point the product is so good that it doesn't really matter. But yeah, it was just like, "Ah, let's just be a bit..."

[00:08:35] Martin: Just a bit different and a bit like chirpy. So

[00:08:39] Matthew: yeah. Probably a good... i've got two questions. I'm not sure which order, so I'll ask them in, but one is, what... You say you've done a lot, a bit of serial entrepreneurship, and you've come up with a lot of different businesses, ideas along the way.

[00:08:52] Matthew: What, how did you identify this specific problem? 'Cause Alvin is very, like a very specific solution to a very specific problem, and I guess, [00:09:00] what is it? And answer those in whatever order you want. So just for the listeners who, who've never come across Alvin before.

[00:09:06] Martin: It's a bit like the cart and the horse in terms of explaining Alvin.

[00:09:10] Martin: But for people who know what data warehousing is which I assume maybe the listeners would know, what the da- data platforms are. They know what Databricks, BigQuery, and Snowflake, and, dbt Analytics is. BigQuery is Google's or Google Cloud Platform's offering in the data warehousing space.

[00:09:31] Martin: We decided- How would you- ... simply a tool that plugs into any BigQuery environment and automates cost optimization. And that's the very short. So the product itself is engineered in a way to provide completely transparent and automated cost savings. So it's not a tool that gives recommendations or endless lists of tasks to complete to, to obtain the savings.

[00:09:57] Martin: We identify, we [00:10:00] find, identify, and execute the savings automatically

[00:10:04] Dara: Am I right in thinking that it was originally and the reason I say am I right in thinking is 'cause everybody now, I, you use AI to help you out with research and and everything really. But my understanding is that the, that Alvin started out slightly differently, and it's over time has narrowed down to that focus.

[00:10:20] Dara: Is that right, that it was more around lineage and observability, and then you've narrowed the message down? Is that correct, or have I just got the wrong end of the stick?

[00:10:29] Martin: No it's absolutely correct. It's we really got traction on the fundraising and investor side, I think, around, when COVID started.

[00:10:40] Martin: And at that point, we ... As I said, that's what I meant initially when I said that we were raising money, things take time. We actually had a pretty scientific approach to get started in terms of identifying problems, talking to companies, and I think we, what we identified is that, especially in [00:11:00] this 2020, '21 COVID, post-COVID madness with tooling and funding and, a bit craz- crazy times, like a bubble almost around this modern data stack.

[00:11:11] Martin: And we found that the companies are, at that point, not really that cost sensitive. It's really more of a understanding what the hell is going on, which was the premise of all of the observability and lineage, and it cost was an afterthought really. It was more about we need to move fast, but we need to have the right data.

[00:11:31] Martin: So that's what we were, we're tackling. And we were all about ingesting logs, creating a very complex refined data model for analyzing all types of logs from data warehousing systems. Actually, for Databricks, Snowflake, and BigQuery and beyond. And I think as we were building that out we obviously saw that there was a lot of competition.

[00:11:57] Martin: We're not the only ... We were, I guess we're not the [00:12:00] smart people, the only smart people in the room. Lots of other people see the same as we do. And we were winning deals against, on BigQuery, we were winning deals against Dataplex and, all of the built-in stuff, so we knew that we probably had by far the best lineage solution.

[00:12:17] Martin: We were using a combination of Google's own Zeta SQL, which is what they use to power BigQuery. So we had basically decoded that horrible C\+\+ library and were using it at scale in combination with ... So we had something really good. We could do- Types of lineage that no other company could do in terms of struct and, all of the good stuff.

[00:12:37] Martin: But we found that it was increasingly, just a, an added feature in, a data catalog, and there were lots of open source that is I think we were, like, over-indexing on this academic completeness versus, 80/20, you can do the manual stuff and the rest is fine. And I think then as the space was changing, maturing the entire kind of investment [00:13:00] hypothesis when we raised our seed round, which was also very large in, Estonian and probably European kind of measure, especially for the stage we were at with amount of customer and revenue and these things.

[00:13:12] Martin: It was best of breed tools. Alvin is by far the best lineage tool out there, and this will sit alongside a host of other tools that connect to the data stack. And of course, now we can say that, that, that investment thesis didn't age super well that, that is what it is.

[00:13:29] Martin: You have to adapt. So then what we understood is that we have very good understanding of metadata, of logs, of, what's going on in the warehouse. And then we thought about now cost and, there's there's a lot of problems around cost.

[00:13:48] Martin: Data teams is struggling to prove their value. There's all of these trends that we, now think we can-- can, we are in a much better position to handle than we were before. And we [00:14:00] also realized that I have about 11 years of pretty deep hands-on professional experience with BigQuery.

[00:14:07] Martin: Marcelo, our founding engineer, has about the same and is a Google developer expert. And we're like our entire, life's work is around GCP and BigQuery. This is what we know." And then we decided to ask some of our initial customers whether just whether they would be interested in, in, in testing some automated cost optimization.

[00:14:28] Martin: And that was the start of what this is Alvin now, really. Just a step away from generic observ- this red ocean observability and kind of lineage market and into this more, I would say, market-leading BigQuery optimization platform. And we have a lot of plans beyond that that, that is coming, of course.

[00:14:47] Martin: But definitely BigQuery and GCP is where we have the, the most customers and and where we have, I would say, very, very- Yeah. A-as with Linage I would say it's the leading capabilities in terms of how it [00:15:00] works.

[00:15:00] Dara: And what are the, i-i-in fact, actually, sorry, let me get my, back to what I called Matthew today, and I've don't know which way around to ask my questions.

[00:15:07] Dara: Maybe the s- obvious first one is to say, when you're speaking to, a potential new customer and they say okay, I get it. How does it work?" What's your kind of what's your simple answer to how it works?

[00:15:20] Martin: It all depends really on how your current situation is.

[00:15:26] Martin: And we try to not have this one-size-fits-all sales approach. I would say even with... it doesn't matter if it's a small customer, if it's a large customer, everyone gets the same kind of white glove treatment. It's almost the idea that in any encounter with Alvin even if we find that we cannot work with you at the moment for, different reasons, we really aim to give them value as if you would have a really good consultant that, gives a initial statement of work that you decide to not use it, but you still have a great analysis and [00:16:00] be-better understanding of things done than later than before.

[00:16:03] Martin: So if a customer is currently running all of their BigQuery spend on demand, and this is like even when it gets a bit, technical because BigQuery has multiple different billing models. It's on demand, where it bills you on how much data that's being scanned. Pretty simple, easy to understand.

[00:16:22] Martin: It can have some nasty surprises if you don't know what you're doing. And then there's capacity, which is more compute. Think like a Snowflake warehouse or, Databricks, where you just have compute, but it's a bit more fine-grained in how it can scale. So if a customer is coming to us and they say that we are currently paying, $10,000 per month on BigQuery, we would either then try to understand, okay, what SKUs are this this based on?

[00:16:50] Martin: Is it on demand? Is it reservation? May-maybe it's even a mix of those. And then what we'll, what we'll tell them is that based on an a- analysis of your environment, [00:17:00] we can set up a service that can route all your queries between the different billing models that BigQuery has to offer, and thereby gain significant cost reduction without any, like any analysis or any work on on your end.

[00:17:17] Martin: And this is normally achieved through setting an environment variable or changing a database connection parameter. That's how the pitch goes. So it's a very, easy, most companies understand very quickly Kind of how it works and, how easy it is to integrate.

[00:17:33] Martin: And that's I think, the good thing for both parts because it's not this two-month POC, like a lot of investment. In, in some companies we might not even be known to a data team because there's some ops or DevOps person that just configures their Kubernetes clusters that all the containers have this parameter set, and then the network the BigQuery client would then use our services instead of going directly to the BigQuery API.

[00:17:59] Martin: [00:18:00] And kind of everyone's happy, it's quick, and then, it's run a POC like a week, may- maybe two weeks, and then we conclude whether, we... whether we were successful or not. So it's, Yeah.

[00:18:12] Matthew: And I g- I guess I really like the the cost model. It seems it seems just so intuitive and tell me if I've got this correct, but essentially you charge a percentage of whatever is saved on what you've just described.

[00:18:26] Matthew: So if they are... they start proxying through Alvin and you're sending it in, don't know, you're sending it to committed use rather than to on-demand or all any other mechanisms you've got, and you save them £8 on a query, then you'll take percentage of that eight, £8.

[00:18:43] Martin: It's very important though to note that I noticed that you mentioned send it to committed or to, to a commitment.

[00:18:49] Martin: This is extremely important to note and something where we have had multiple cases of of a deal basically being lost or very [00:19:00] close to being lost based on the interaction with with the other party. And then we'd like to I don't understand this. Here a customer could save 25K per month.

[00:19:09] Martin: We have been very helpful. We have been super we feel that we have done a great job in in, in throughout the process. And then we understand that, oh, they actually think that we are now they actually think that the savings are generated from buying commitments on their behalf and that we're somehow creating financial liabilities to, to generate the savings.

[00:19:29] Martin: A bit like, platform like ProsperOps do this with this micro commitment, and they create this portfolio of micro commitments on VMs and database. It's very refined. That obviously creates a liability, but it's such a small- it's so like small increments and it, it's actually proven to, to give the, good rate optimization.

[00:19:50] Martin: So I think, maybe they think that we do it like that, and they're like I've, I'm not ready for that. I don't want things might change." So this is purely based on SKUs that [00:20:00] are, on demand or pay as you go. They are essentially what's the word to cover that?

[00:20:05] Martin: They are pay as you go as comms. So if if if someone decides that they don't wanna work with us for whatever reason- ... then from that moment on, there is no, let's say, residual cost or anything like that in, in their environment. Because it is a little bit, too good to be true in some extent that that, what is the catch here?

[00:20:23] Martin: And that, that can sometimes be hard to to explain. But I think it, in, in some cases we go through, this process, which is I think what customers appreciate or non-customers after the process, of course, that we basically get all of their metadata information about the jobs, their setup the tools they have connected and, cost drivers and these things.

[00:20:46] Martin: And as I said, it means that, if they decide to not work with us, they would still leave with a much better understanding of their environment. May- m- sometimes we might say that, "Maybe if you just do this, you don't have to work with us for a while." Or, if [00:21:00] you're, if sometimes we see that spend might be too low for us to be able to actually have a meaningful impact, like in theory we, we could maybe have some impact, but then we find that, there's this like increments that Autoscaler works with.

[00:21:14] Martin: And, if the overhead of that dis- is too big, then maybe we have to wait a little bit. So then we could give some advice. As I said, we have now 25, 30 years of combined experience on BigQuery, maybe more. So there's normally always some nuggets to be had from working with with our team.

[00:21:30] Dara: Are they the main reasons then? 'Cause you said earlier, you said, "Oh, if someone's not ready to work with you." And I was thinking that at the time. I was thinking hang on. What is the what reason would somebody have not to use it?" And then you've named two. One is this maybe fear or paranoia that there's some kind of liability being baked in, and the other one then is maybe that there is a minimum threshold.

[00:21:49] Dara: Are they the kind of two main... Are there other reasons why people will take a proposal from you or speak to you and then just not actually proceed?

[00:21:58] Martin: I think so [00:22:00] far those are the main kind of main kind of reasons. I think when it comes to the, just the spend itself, it's also a little bit like, you know- We decided to have a company with certain budgets and certain team setup and configuration.

[00:22:14] Martin: Maybe... Oh we do see, and that's fair enough, some companies might have more of a not invented here syndrome. But I'm not saying that to, to share we see that some companies have indeed built similar stuff or the dynamics and mechanics of what we're doing is not necessarily rocket science.

[00:22:32] Martin: We have also written blog posts about this. We have obviously refined it to a very large kind of turnkey solution. But certainly we also know that, companies might have some ideas how they want to do it themselves. They might have security or, concerns about letting some type of black box into their environment.

[00:22:53] Martin: And, we have to respect that. It's not necessarily for everyone. And then, as I said, we always try to just be helpful and [00:23:00] useful. And in... What we see is that in some cases, maybe not now, but six months later we find some way of working together. And I think the good thing for us about this case that it's often once you ask the right way, and if you're generally try to be a nice and understanding person, you understand where the objections come from and who have them.

[00:23:22] Martin: And this has also been a significant factor in improving our product to, to offer things like ability to host certain critical parts in their own platform. There can be... We can do certain guarantees about flows of data and, what goes where, which is, critical for for a lot of companies when it comes to compliance and security.

[00:23:42] Martin: So it's all of these things where, you know, okay, not now, and then we're like, "Okay let's actually build this and get back to them." And then, that kind of move moves the needle.

[00:23:51] Matthew: What about longer term? So say you've got someone comes in and starts working with Alvin. Is there ever a case where, you know, over a long period of time they forget about that [00:24:00] initial saving that they saw when they first started working with you and just forget the value that's being added because it's built in, and then a couple of years down the line they're like, "What's this we're paying for?

[00:24:09] Matthew: Get rid of it," and then have a rude awakening when all the cloud bills go back up again? Do you see that?

[00:24:14] Martin: I'm not in a position to give numbers, but I can say that the churn numbers are looking very well, I think, for, comparable companies. So there, there is, of course, a big element of, is, seeing is believing, and there are cases where- Where we, it's, I guess the onus is on us to prove our numbers. Like we have this UI report on the generated savings and I can, of course, we can get more into the details about how that attribution works. But i- in these cases we have also learned a lot.

[00:24:47] Martin: So normally we, just provide a company with a full, run this on their information schema, you would see exactly the same so that you can, do like a cross between RUI, your [00:25:00] BigQuery information BigQuery editor information schema queries, and also the billing.

[00:25:05] Martin: So we haven't had, directly this like rude awakening in terms of a churn experience, but we've had a couple of this I would say deals where they don't maybe fully believe, or like they, they believe it, they just were like, "Okay, if we just turn it off what happens?"

[00:25:21] Martin: And then it's "Okay let's sign the contract and continue." So I think it's, and but that's why like I understand it if I would engage with a vendor that, that makes claims cer- certain claims, I want to validate those. And I'm a data...

[00:25:35] Martin: I think that is, to me, like the most motivating thing as a founder and a technical founder is, of course, that we work with ve- extremely serious and diligent people. They would never buy or use our product if it wasn't working. Like that's to me, like the thing that is extremely motivating, that we have built a product and a platform that is working and is clearly delivering value.

[00:25:57] Martin: So that makes it fun and [00:26:00] motivating, I think for everyone in the team that you can see the impact that, that you're making. And of course, we have made it easy for ourself, by, I think, choosing a product where the to build a product where you can set the North Star metric of the product is, very financial.

[00:26:15] Martin: And I think this is it. This is, this was generally a massive issue, and I think is still an issue for almost any company that, that does something within data observability. It's y- you always see this case study. You always see this like time saved. You look at it and if you think about your own work and workload, and you're like a little bit are you sure they still don't try to run some queries directly on the information schema?

[00:26:41] Martin: Do they actually use this and trust this?" Like it always especially knowing that people still might do this with our product and we at least we, our metrics is okay we are s- we are- I think this is something that we don't really talk about much though, but that in almost every case, we're able to also improve [00:27:00] performance quite a bit.

[00:27:01] Martin: So it's has that aspect to it also. But as I said, like we, we have pretty hard metrics in our product which, you know for obvious reasons. But I think that, yeah, it just, I know f- from, knowing it deeply in my soul how hard it is to try to build a product around even cost optimization that that just leaves, it leaves basically the user hanging at now you must go and implement this.

[00:27:26] Martin: And then I'm like, how do you quantify the value of that product? Like, how do you tribute what you do to the cost of the product or all of these things? So it's almost like we went backwards and we designed the entire product and platform in a way that can we automate? Is it can we make this so easy that it becomes almost like a no-brainer for everyone that's involved in, in, in the process of buying it?

[00:27:54] Martin: Is that it would be crazy not to try it. And I think we, I'm not not [00:28:00] being delusional here. There's a lot of work left and it's certainly not perfect yet. But I think we, in terms of seeing how the process works and how the pricing works and all of that, I think those elements we are pretty happy with at least that it's it seems to hit the right nerve with people we talk to also.

[00:28:16] Dara: Is that I was gonna ask you about the the performance aspect. Is that a byproduct or is it something separate that you've worked on? Is that just by optimizing cost, you're gonna optimize performance or are they-

[00:28:27] Martin: It's ... So I would say this is this is a twofold thing.

[00:28:31] Martin: I would say more, more recently, we have actually had a lot of ... not a lot, but we had enough that it's interesting as a signal, enough POCs and customers that are not primarily motivated by by cost reduction, but because in order of, in order to gain cost reductions, you may actually have to accept performance degradation.

[00:28:58] Martin: But this is a very [00:29:00] interesting topic that we approach in a very systematic way, where we realize that if a company runs a workload they run dbt, it runs every night at midnight, and it runs in 10, 10 minutes, and it uses a ton of slots. It is like they, this is like idling quite a lot.

[00:29:20] Martin: And then the observation is that, okay, now- The tools, the team, the people, everyone might have this idea that it's very important that this finishes in 10 minutes. That's very important because reasons. And then you look at it and you say what we observe is that the actual usage of these assets happens more when people come into work in the morning, maybe eight hours later.

[00:29:44] Martin: So actually now by instead of optimizing for pure throughput and performance, we're optimizing for the resource allocation to those jobs, then we can massively reduce the cost. Okay, we can accept that maybe the [00:30:00] job is three times slower, but 10 minutes or half an hour for a full pipeline like that is...

[00:30:06] Martin: So those conversations have been very interesting to have with customers. And once, once you make them understand the trade-offs and how our platform works and how we approach this and we create this I guess comfort comfort around that is not about your dashboards being, loading in 10 minutes.

[00:30:22] Martin: It's about figuring out the right SLAs and the right SLOs for all the workloads, because that is a massive part of cost optimization and resource optimization. So it's like the performance performance is getting it right, not overdoing it and not underdoing it.

[00:30:38] Martin: And that's what brings me into performance as maybe more like a feature, is that now we have built out a lot of this very intricate SLAs and SLOs and, this we have this... We strive to keep the platform very functional and simple for customers, but we have a pretty crazy backend that we call the control plane, where we can basically look at [00:31:00] everything that goes on in real time with customers and, tweak and all of that when when needed.

[00:31:05] Martin: And this is where we also have this concept of SLAs and SLOs, where indeed for some customers it's okay, now that we have something that can, work with reservations in real time to manage the, scaling and the limits more effectively, something we do to actually reduce the burn, but can we also use this to actually pile on more resources than maybe even BigQuery would at some point because we need more throughput and these things.

[00:31:33] Martin: So it's like an interesting... There is this duality with cost and performance, and we approached it purely from the cost side, I think first, but then you realize that it's, yeah, it is a duality, right? You move it in one direction and something else pops out, and that kind of goes both ways.

[00:31:50] Matthew: For y- you've mentioned a couple of numbers as we've been chatting, like $10,000 or euros or whatever it was you mentioned, as in maybe a floor where you can get the most utility out of things. [00:32:00] But for anyone for any, listeners who are thinking like maybe I...

[00:32:03] Matthew: Maybe this could be super useful for me," are there any other sort of... What other signals would you be looking for within a BigQuery organization that may mean Olive is gonna be really good? And conversely the opposite. Like I can imagine, for example, someone's got a 10,000, 15,000 pound monthly bill, but it's all coming from Looker Studio.

[00:32:21] Matthew: And that might not be something you can get in between and help with.

[00:32:24] Martin: So as, as- We decided- Typically, I would say it all depends. We generally just we generally just recommend every, any company to like to do this savings estimate process. Just connect the data and we'll do a review.

[00:32:37] Martin: Because often we find that someone would say something, make some claims, and then they, we realize that they don't really have the full picture. Once we actually access everything, we can find stuff that they would never think about. And actually we have a lot of clever ways that we have developed to integrate where we can actually work pretty well with tools like Looker Studio even that doesn't necessarily [00:33:00] require a direct, proxy connection.

[00:33:01] Martin: And so we over time we... I think this is part of our, value proposition is that, and this automated aspect, is that it's not like purely one thing at this point. It's just this big arsenal of tools that we can pull out and, configure on a specific customer account that, that would make it work.

[00:33:18] Martin: But when it comes to the patterns and anti-patterns I guess 10,000 is, is like a more like this limit that we often see. At 10K, this is where any BigQuery company normally would start to look at the cost and be like, "This is pretty high." Maybe it's like at the point where it starts to become like a, an FTE type FTE type cost and then they're, like, thinking about, "Okay what do we do about this?"

[00:33:47] Martin: And then it might actually also make sense to start looking at, can we do something with capacity and reservations and all of that. However, we actually work... We have great success with companies that [00:34:00] are much, much lower than that, even down to, 5K below. It depends, I would say, a bit also on the patterns that, that occur.

[00:34:08] Martin: Because if you look at the reservation in BigQuery, and this is I actually wrote I wrote a pretty extensive article about the history of BigQuery pricing models, and- And reservations and commitments which goes into fairly minute details on this. But a reservation is essentially something that executes a lot of really small micro-commitments on behalf of the customer.

[00:34:36] Martin: So this is why a reservation says it has a minimum scaling threshold of 50 slots, and it has a minimum lease time or minimum billing time of 60 second. And if you look at the old flat rate model that BigQuery had, you had these fla- flex flex slots, which had the exact same characteristics. [00:35:00] So BigQuery's, additions and autoscaler, at least...

[00:35:04] Martin: This is not official, but at least my understanding of it based on everything I see and unders- and know, is that it's simply an automation around this old or this API for buying this flex kind of micro, micro-commitments. And so once you understand this you don't have to, you don't have to understand all the history, but once you understand how the autoscaler works, then it also means that you can understand the situations in which it may not work super well.

[00:35:33] Martin: So imagine that you every second you have a job that is using one slot, and then you do that, repeatedly, then you basically are, like, for every, second, you are then wasting 14... So you basically have 98% waste simply because BigQuery's scaling increments doesn't really suit that.

[00:35:56] Martin: So that is the characteristic [00:36:00] that we see where we are like, there isn't enough bin packing, in the, how the jobs are executed in the environment for us to effectively benefit from this routing or this billing model arbitraging. But it, so it means that you can have...

[00:36:13] Martin: Let's say you have a very low spend but a lot of it is concentrated in, one time of day, for instance, then we might be able to work, work better, better with that because the impact of this continuous kind of waste is is much less. Does that make sense? It's we're so into, we- we're so into this so it's may- maybe hard to realize if you're just speaking to deaf ears or if it makes sense,

[00:36:38] Matthew: no, that makes sense. Yeah. Yeah, that's what, that's what I was getting at. Yeah, there's different patterns of how things are using and the different ways that you could spot that signal and help. Absolutely.

[00:36:46] Martin: Yeah, I know we're quite we're quite, we, we-- This is also something that a little bit I think is part of this white glove onboarding or as I said, a lot of customers don't or companies don't even know about these things and, maybe we can [00:37:00] even get to a point where, you know, okay, we can take five minutes to change the cron schedule, and suddenly, we can actually work with them much, much easier, so yeah.

[00:37:10] Matthew: And I can't I'm flabbergasted that we've got 45 minutes in and the word AI hasn't been mentioned yet, so I'm gonna break that right now. Do, have you noticed a pattern of change, obviously with LLMs generating SQL and then sitting and passing that, that data back into warehouses?

[00:37:26] Matthew: Is that... I assume you have noticed that trend, but, how, how-- Is that affecting things in any way? Is it pretty much same as or?

[00:37:32] Martin: We decided, if I will be honest maybe no investor would ever want to touch us, but there, there isn't a, there isn't a single ounce of AI or LLM in our optimization paths or algorithms or, general technology.

[00:37:49] Martin: It's purely deterministic and purely based on, actual signals and stuff like that, which is, y- be- because we do this [00:38:00] attribution at a very deep level, where it's every decision must be logged, and every- everything is like, "Why did we do this?" It's a pretty critical question to answer, both for, SOC 2 and compliance stuff but also for, billing and, actually the savings calculation themselves.

[00:38:17] Martin: So we know there are other companies in the space that are talking about this and saying this, and maybe they have found a way to make it deterministic. I'm not smart enough to... As I said, I canceled my AI and ML degree to do more business. But we find it, that to be a bit risky, and it also makes audits and general, this massive security forms with enterprise customers easier to deal with.

[00:38:41] Martin: Of course, we use a lot of Antigravity and, whatever tools to speed up the coding, idea generation, the product management side. It's super, super good, but not not in the product itself. And we haven't I don't know I would say our tool and product [00:39:00] works more in a, I would say that it just chugs whatever it's thrown at it, and it optimizes it.

[00:39:05] Martin: So we don't necessarily go too deep into- for different reasons, we don't really go too deep into looking at exactly the SQL that our customers are run, is... as I said, for different reasons, it's not something we should do or that we have an interest in. It's more the things we care about is more like are we able to, are we able to produce stable hashing and fingerprinting and, these type of things that we need for the product to work reliably, which, which is, yes I think that's more like customer specific, I would say.

[00:39:36] Martin: Not something that we see as a trend across the customer base that that happens. So I would say maybe it's like an, yeah, un-undefined or like I, I don't have a clear answer to, to-

[00:39:48] Matthew: I suppose that y- you answered what I was getting at. I was wondering essentially, does your s- does your solution here, Alvin, that sits in the middle and does all that deterministic [00:40:00] savings of and routing of queries, et cetera, work on any type of query?

[00:40:04] Matthew: 'Cause you can imagine the queries are going, the volume of queries are going to be increasing because more people are empowered to send SQL queries to a warehouse than there were before. 'Cause before you had to know SQL or you had to be an analyst, whereas now you may have, Joe Bloggs in, in the marketing team who's sending SQL queries not knowing how to write a word of SQL.

[00:40:25] Matthew: But it sounds as though, the way it's built, it can just handle that and help reduce the cost of the SQL passing through Alvin.

[00:40:32] Martin: Yeah. This has been a very key one of the, I would say, most key design criterias that we set out to Alvin does not make any assumption about the query sources or- Yeah

[00:40:46] Martin: Or really anything for that matter that, that, that goes on in the customer environment. I think we, again, a bit based on, the hard-earned experiences that tying yourself hard to certain [00:41:00] tools, certain assumptions about things may be a bit easy to get started, but but it's it would be very hard to break, break out of that.

[00:41:08] Martin: So that's why to, to any tool that works with Alvin, it simply looks and feels like the BigQuery API, and that kind of makes it easier for us also because we simply see, A, we get the query. Maybe there's some labels or comments that identify, what app or sources comes from but we don't really have to make any other assumptions beyond that.

[00:41:31] Martin: So it's almost like the, simplicity by stupidity or, just like work with the simplest kind of level- of input that, that you can that, that works pretty well for us.

[00:41:41] Dara: Do you find, and I know this is gonna differ case by case and maybe I'm not, maybe there's a nuance of the product I'm not understanding here, but I could imagine in theory somebody would use Alvin and they would immediately benefit from the cost savings, but they could also then learn from where those inefficiencies are.[00:42:00]

[00:42:00] Dara: Because what they probably wouldn't wanna do is have Alvin optimizing the same query repeatedly when they could actually do something on their side to change that workflow. But then in reality, I know that not everyone's gonna do that, and they might be just happy to pay knowing that it's gonna be, that they're making a saving compared to what they were doing before.

[00:42:19] Dara: So I know it's gonna be different for different cases, but in theory, would somebody see their co- would somebody see the need for Alvin drop over time, or is it al- is there always gonna be something that Alvin is offering because it's looking across the whole thing and it's got this, all this deterministic decision-making built in that there's

[00:42:38] Dara: Do you see what I'm getting at? Is there ever a case where somebody would basically-

[00:42:43] Martin: Yeah. We generally just try to be ahead of that that problem. So just to to, it's important to note that there are cases where this is more or less relevant. Let's say if you connect Looker or some other BI tool, then [00:43:00] Alvin will pass every single query from e- every single dashboard element.

[00:43:05] Martin: There's simply no way that you can even know what dashboard element is, representing those queries. The benefit of Alvin there is undisputed. It's you just can't do that yourself. Then we have stuff like DBT or Dataform or we know where you have more specific models.

[00:43:21] Martin: So generally what we- how we approach this is that Alvin has a built-in, attribution model that we use to to actually generate, an invoice customers based on the savings. So at the very simple level, it works like like Matthew said earlier, that, we, we generate the savings, we take a cut, cut of it, and then so now let's say that a customer looks at their logs and they, might decide to change the DBT model maybe in accordance to what we did to it or they might change, choose to add partitioning, do something that, changes the cost.

[00:43:58] Martin: Then either, the [00:44:00] cost delta goes to zero because then Alvin sees that, A, this query is is running on a reservation already or it's running on demand or, you know- Decision on the specific specific model or, what- whatever is that it should run on demand or reservation.

[00:44:17] Martin: So it's almost again, the system is designed where we don't really make any assumptions. We simply know that this is the optimized, thing for this thing, and we simply, execute on that. If it should change, we execute that. If not, we just leave it as is, and we leave we leave a record of us just leaving it as is.

[00:44:40] Martin: So basically by having it like that it automatically just detects that. So we just tell customers that, if you have a process where you would like to manually optimize your models or do stuff, then, every two weeks you do the top 10 dbt models or, whatever you have in [00:45:00] your environment, and you would like to do that, then by nature of our attribution model, we would never claim savings or kind of claim anything that we actually haven't actively done.

[00:45:11] Martin: So wh- while we though jokingly say, "Of course, we don't want you to do that," but I don't know we find that as you have, tons of dashboards not uncommon, 5,000 dbt models, there's gonna be this massive long tail. There's gonna be new models. There's gonna be changes over time to models.

[00:45:29] Martin: So it just sits there and always just takes care of it. So we've found that to be, less of a problem. We, we've been worried about it, and we're like, "Oh, what if they do that? What if what... maybe it doesn't work." And it certainly is a relevant thing that we must always think about what is the added value, how can we, make it even more valuable and these things.

[00:45:50] Martin: But in practice, it hasn't been that much of a, much of an

[00:45:54] Dara: issue, I think simply because it's just like this turnkey- Yeah ... turnkey service. And then on [00:46:00] the... to go to the other, as a bit of a follow-up, and to go to the other end of the other end of things, so that's the risk of, your customers figuring out they can do a lot of this for themselves and it's just-

[00:46:08] Martin: Just to give a final comment.

[00:46:09] Martin: So this is also I think, I think part of the product and part of the appeal is that, for almost I think everything that Alvin does if we set up a reservation, if we manage reservation, you can actually see in the reservation changes logs in the information schema everything that we did.

[00:46:29] Martin: So it's almost there's no- There's no hiding. Everything that we do is there. But I think it's as always, the devil is in the detail.

[00:46:36] Dara: Of course, yeah. Yeah. No, absolutely. The follow-up I was gonna ask you, and I would hit you from one end and now I'm gonna hit you from the other end.

[00:46:42] Dara: So do you, with BigQuery and GCP in general just getting more and more automated and intelligent as time goes on, do you have ... you must have some concern, but what are your thoughts on whether Bi- whether BigQuery itself is gonna end up just baking a lot of this in to the system itself?

[00:46:58] Martin: It's always a concern, it's [00:47:00] it's ... the only thing I can say is that we were very concerned about this previously when we were doing data lineage and observability, and then seeing where that went, and now Snowflake, Databricks, BigQuery has all of this built in.

[00:47:14] Martin: But, we also know that this ... how do you say this? It's the quality of these products and services are, especially on the GCP and Dataplex side, it's mostly, I think, a checkbox thing where some enterprise just, "Oh, we, we'll get that," but it's not ... if you ever tried to use the lineage in Dataplex, it's still as crap as it was, ye- years ago.

[00:47:35] Martin: And but that doesn't matter, right? And that's what we have learned, of course, is that it's not, in some cases a bit like this Betamax and VHS, right? That as long as they have it, they have distribution, they offer it to some VP that goes to a dinner on Cloud Next and, gets a lot of promises, and then you still lose the deal even if the product is better.

[00:47:54] Martin: So that we un- have understood. But I think I think what we have done repeatedly now [00:48:00] with with, especially with BigQuery, is that even larger customers that are now considering using commitment, using capacity, we can offer them a much more appealing suite of of a solution where basically you have zero lock-in it's dynamic, it scales.

[00:48:18] Martin: Maybe you even want to use commitment, but just use commit- just use Alvin first. We can streamline the spend. We can ensure that you have the right buckets of the SKUs in your BigQuery spend, and we tame the autoscaler so it doesn't, you don't have an inflated perspective of your use when you go to your GCP rep and start negotiating on your potential discount and commitment.

[00:48:41] Martin: So it's like a, we just see it as like a, a supplement really. It's not by nature of how it works and, at least how we have seen Big- BigQuery for men- like a decade and more now- it doesn't really strike me as something where, it's just so much competition in [00:49:00] the data warehouse space also, and cost is a massive thing.

[00:49:03] Martin: So it would be it's almost like it's an advantage to, to have this have this platform on GCP. And we're pretty friendly with the Googlers. We are a valid- validated solution partner. We're on the marketplace, we, so I don't know. It's it's a bit like this, if a VC asks you what if Google does this?"

[00:49:23] Martin: Or, like they... No, not exactly the same because it, they're, they obviously have an offering and tooling and stuff around it. But generally, I think what is a threat which I think is more genuine for, us than most companies in the data space is that, as I said, someone goes to Cloud Next, someone gets wined and dined, someone gets 500,000 in credits and freebies, and then they're like, "Oh we don't have to think about this.

[00:49:49] Martin: Yeah." I think that, that is the, I think the killer, especially when you move a bit upmarket, is that someone's talking in, in, in the back room, and they get these amazing [00:50:00] deals that, that, they would never talk about this publicly. But it makes it just very hard for an up-and-coming vendor to squeeze in because, they just know that once you get them on the hook, and I can tell you that 500K of credits, if you're a large company, that can go very quickly.

[00:50:16] Martin: And then they got you with all the commitments and what not. And this is actually one thing that we have experienced which we have entered a deal or a process and realizing that, you know, and even telling a company that it's, it sucks that you signed this agreement because, based on actually what we can show you of your data, had you worked with us, we could, provide you zero lock-in and 50% cheaper.

[00:50:39] Martin: But this is the world of politics. And, it's a very, I'm sure you're familiar with that. It doesn't follow logic and principles.

[00:50:46] Matthew: So I, yeah, I ask the same question every week, and essentially it is, it doesn't have to be about AI, it can be about anything, and inevitably it ends up being about AI 'cause that's all anyone talks about nowadays.

[00:50:57] Matthew: But where do you see the industry [00:51:00] at large or the world at large going and big changes that are coming over the sort of next two years? Have you got any predictions you would put a stake in the ground on?

[00:51:08] Martin: That's that's a tough one

[00:51:09] Martin: the world is going to shit.

[00:51:11] Martin: That, that's for sure. So we can talk about the data industry. And I've been a bit ... I knew it was coming. Everyone was talking about it, and based on this best of breed to more large tools, the consolidation, I think is like one of these predictions that everyone talks about, but it's just happening like crazy.

[00:51:34] Martin: Like a lot of, friends or, frenemies as we call them, different companies are being acquired east, east and west. And it's interesting. This is like the trend that I think will just continue in, in the space. It's gonna be ... We decided to some extent this AI and LMS makes it easy for everyone to create their own company.

[00:51:51] Martin: But but I think especially in the data and cloud, there's been so much over-funding and now there's a lot of companies that are just [00:52:00] finding like a safe harbor. And I think that's a trend we will see see more and more. It'll be interesting to see now how all of these like founder modes really talented, good people now come into these companies and, how they will hopefully transform them into being competitive against all of this, AI will eat SaaS and all of these other predictions.

[00:52:22] Martin: I think that will be like an interesting because interesting trend to watch that a lot of companies now have acquired and are growing like this in- inorganically like that how they will fare against all of the upstarts of the AI wave.

[00:52:36] Dara: All right. Martin, thank you again for joining us on The Measure Pod today.

[00:52:40] Dara: It's been really interesting. Appreciate your time and yeah, thanks for your thoughts. No worries.

[00:52:46] Dara: That's it for this week's episode of The Measure Pod. We hope you enjoyed it and picked up something useful along the way. If you haven't already, make sure to subscribe on whatever platform you're listening on so you don't miss future episodes.

[00:52:59] Matthew: And if you're enjoying the [00:53:00] show, we'd really appreciate it if you left us a quick review. It really helps more people discover the pod and keeps us motivated to bring back more. So thanks for listening, and we'll catch you next time.