Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
046 - Rebuilding Monzo's Data Warehouse from the Ground Up
248 Plays15 days ago
Monzo went from zero rules to one of the most opinionated data architectures in fintech — and cut warehouse costs by 40%+ doing it.
Bruno Campos, Analytics Engineer at Monzo, joins Harry to unpack the rebuild of Monzo's entire data platform — tens of thousands of dbt models, hundreds of teams, one BigQuery warehouse, and a migration that's still only 30-40% done.
Bruno breaks down Monzo's new "OOM" architecture: four fixed layers, data contracts between teams, and an internally-built tool (ModelGen) that turns a YAML file into standardised SQL. The result — faster builds, easier onboarding, and serious cost savings that show up on the GCP bill.
They also get into why "data" as a discipline is still, in Bruno's words, "incredibly immature" — and what it'll take to fix that.
In this episode:
- Why Monzo moved from a free-form warehouse to a heavily opinionated one
- The four-layer OOM model: landing, normalized, logical, presentation
- Interface models — Monzo's version of data contracts between teams
- How ModelGen automates standardisation across 100+ teams
- The real cost and time savings from re-architecting at scale
- Bruno's take on AI's role in analytics engineering — and why the job isn't going anywhere
Recommended
Transcript
Challenges in Managing Large dbt Model Projects
00:00:00
Speaker
It genuinely can get quite complex and I think at this scale there's not many who have solved the problem really and it's so it's very difficult to kind of Google how to run 15,000 dbt model project for hundreds of data practitioners. AI also ah kind of struggles with these because again it was trained on data so how is it going to solve this problem?
Communicating Value of User Experience in Data
00:00:21
Speaker
If there's anyone listening considering kind of putting this forward to their senior leaders and kind of trying to explain that thinking about user experience for data is something that's valuable. The the thing that I would say is essentially that this is the single biggest change that you can make to the speed that the company produces and consumes data.
00:00:43
Speaker
If it's easy for people to be able to develop data and get insights out faster, then everything else is downstream of that,
Shift in Software Developer Roles
00:00:51
Speaker
right? So this is ultimately is a multiplier. I think there's there's been lots of articles talking about how software developers are no longer developers, they're more like software reviewers. So you're reviewing code a lot more than writing. But the truth of the matter is, is that if you are reviewing code more often, then you have a lot more time to think about the what you're doing and how you're going to do it and whether or not that's best for your specific use case. And that ability to detach from the work that you're doing and go, do I need to rethink this? Do I need to pivot here? Would there be a better way to do it? Removing that barrier and friction to starting again or tweaking, ah genuinely, genuinely, genuinely to me feels like a superpower.
Omni AI Analytics Sponsorship
00:01:35
Speaker
Today's episode is brought to you by Omni.
00:01:37
Speaker
Most companies I speak to want AI analytics but are failing to put projects into production. That's where Omni is different. It's the semantic brain that grounds AI in the heart of your business logic, giving you governed answers, whilst also the depth to identify root causes. It's intelligence everywhere, from your spreadsheets to their chat feature, even within your product. Omni moves you beyond the dashboard. Don't just take my word for it. Trust teams like Bluplexity and Synthesia that are already using Omni to deliver intelligence that people trust.
00:02:06
Speaker
Check out them in the show notes or visit omni.co. That's O-M-N-I dot co. Now, back
Introduction to Bruno Campos at Monzo
00:02:14
Speaker
to the show. Hello everyone, welcome to another episode of the Stacks Data Podcast. Today I'm joined by Bruno Campos, the lead or one of the lead analytics engineers at Monzo. um at the point At this point Monzo is definitely fair to say, has scaled. They've got tens of millions of customers and one of the largest analytics engineering teams anywhere, which means the data platform that is built on
00:02:45
Speaker
has to operate a pretty serious scale.
Monzo's Transition to OOM Data Architecture
00:02:48
Speaker
Bruno sits on the analytics platform team within Monzo where he's been deeply involved in helping to roll out and championing a major redesign of Monzo's analytics architecture, what the team now refers to as the OOM data architecture.
00:03:09
Speaker
built around object-oriented modeling principles. In the episode, we're going to unpack why Monzo's data platform is shifting, what it looked like before, what it's moving into, and what actually made them think about this architecture in this way.
Bruno's Role and LLM Application at Monzo
00:03:28
Speaker
at scale. We'll also touch upon Bruno and his AI capabilities and how he's applying LLMs into the workflow of an analytics engineer and and what he really thinks about the role and where it's heading. So Bruno, it's great to have you on the show. Yeah, thanks for joining me.
00:03:47
Speaker
Thank you for having me. chief Excellent. Well, I know we know each other well, Bruno, over the years, but for anyone that doesn't know you, um yeah give an intro to yourself and I suppose where you get where you've got to in your your career.
Bruno Campos' Career Journey
00:04:03
Speaker
Yeah, sure. So, yep, I'm Bruno, working at Monzo at the moment as an analytics engineer. I pretty much have a pretty standard ah data career, I suppose. I started off as a data analyst and then slowly moved further and further towards the engineering side.
00:04:25
Speaker
um i think I've always been interested in tech and have attempted building websites and some apps on my own a few points in time and Data engineering seemed to be a nice marriage of those two concepts and a love numbers, i love data, so that worked out quite nicely. um Yeah, currently working at Monzo and as you described, we have a pretty interesting challenge ahead of us with the current state of the data warehouse at Monzo for such a larger state. So yeah,
00:05:05
Speaker
Excellent, well thanks for the intro
Monzo's Analytics Engineering Team Structure
00:05:07
Speaker
Bruno. Yeah I mean for those again that don't know Monzo has one of the largest analytics engineering teams in the in the industry so help contextualize it let's get a flavor of what the team actually looks like howi and how AE is structured to to solve problems day to day.
00:05:26
Speaker
Sure, yeah. Yeah, so the way that the company works is that we've got these departments, which we call collectives, and they can get pretty large. I think compared to some startups, you could even consider them almost like a separate companies, especially when it comes to the sheer size of people working in them and therefore the data that they produce.
00:05:51
Speaker
um And the way that the data discipline is structured is that You have dedicated data efforts within each collective. um Some distributions vary, of course. So some collectives might have a lot more AEs than others. Others might have just data analysts or just data scientists. It depends.
00:06:13
Speaker
um But that is kind of the organizational structure. um And then from kind of a stack perspective, We work off of one repo, one analytics repo where everybody can contribute to. And of course, one warehouse. So we we use BigQuery. And then there that's where everybody's data lives. And that's where they work on. That's that's where everything is.
00:06:40
Speaker
Excellent. And you've moved from the, oh well, onto the platform
Roles of Embedded vs. Platform Analytics Engineers
00:06:45
Speaker
team. what What's the difference between your embedded AEs, the platform AEs and yeah, what kind of problems sit across the two different types of of roles?
00:06:54
Speaker
Yeah, that's actually a great question. And I think one that not many people get the opportunity to explore as not many data teams are as big or mature as this, right? so I think that's quite interesting. So I started off in operations ops.
00:07:09
Speaker
um and I was indeed one of those embedded AEs. the i suppose the focus there is pretty much on making sure that the data estate for the stakeholders, the users of your data, um are ready to be used.
00:07:28
Speaker
And that can take quite a few different shapes. For the most part, it does involve lots of data modeling and keeping your data models and your your section of the repo very ah tidy and and and with high quality, observable. So you've got monitoring and that sort of thing.
00:07:48
Speaker
but you also naturally sit a lot closer and you're expected to sit a lot closer to the specific product that you're working with. And that might be an actual product within the app. So if you have a Monzo account, if you go into the bank section and you have like, i don't know, a page that the growth page, for example. So it might be that you sit really close to the particular product or as is the case with operations, it might be the purely the customer help functionality that might work.
00:08:18
Speaker
And that is an interesting contrast with the platform side, where of course your main stakeholders are the other technical users of data within the company. So you're almost like an internal product, right? And i think the main difference there, and probably one that I tend to prefer in in every role that i've I've been in, is that Whenever you're creating these more platform shaped products, you have to solve the problem for your specific company. And I think that that to me is a very interesting problem to have because every company is slightly different and data as a discipline industry wide has very unique shapes depending which company you're in. And so when you're in a platform team for data, you're having to kind of doubly
00:09:14
Speaker
solve the problems internally. And the best part is that you can go speak to your customers directly, right? They sit right across from you. And so I suppose in the same way you're you still expected to sit close to the customer, it's just that that happens to be internal.
00:09:32
Speaker
interesting, really good breakdown there, Bruno. So we're going to dive into this whole new re-architecture that Monzo has been doing of the analytics platform.
Scale and Challenges of Monzo's Data Estate
00:09:43
Speaker
And think it'd be good to say first off, if anyone wants to dive deeper into the what me and Bruno are discussing, Monzo have actually gone live with a blog article that we'll put a link to in the in the show notes. But help help set the scene, Bruno. What what is the actual so scale of Monzo's sort of data estate, the number of models, warehouse usage, and what were the types of challenges that you started running into?
00:10:13
Speaker
Yeah, um yeah we we we've essentially had to manage a pretty big estate. We're looking at about tens of thousands of dbt models all within one project um and then to echo back on the description of the org the collectives themselves are between 10 15 i'm not really sure how many exactly but the the independent teams within them can go up to about a hundred if you were to count each different section so that should kind of give you a pretty good view of the complexity and
00:10:53
Speaker
You know, with all of these ah models and all of these teams come so many different micro decisions that the company has made over time that slowly come comp compound and compile into this bigger thing, I guess is the best way to describe it. um Because it genuinely can get quite complex and I think at at this scale there's not many who have solved um the problem really and it's so it's very difficult to kind of Google ah you know how to run a model project
00:11:33
Speaker
for hundreds of of data practitioners and AI also ah it kind of struggles with these because again, it was trained on data. So how is it going to solve this problem?
Efforts to Simplify Monzo's Data Architecture
00:11:43
Speaker
And then of course, there's the human element as well, which is managing all of these these people become become a very difficult thing to do as well, purely on the basis of, okay, like where do you draw the line on rules and what do you let you know happen within the the warehouse? What are the standards that you want to adhere to and so on?
00:12:04
Speaker
So that that problem is still here, right? Even after the re-architecture, but that should hopefully outline the motivation to try to make the the architecture the the simpler part of this complex puzzle.
00:12:20
Speaker
um And that was one of the things that we could do as a business, which is to solve that element and try to make that piece of the puzzle, the architecture, more opinionated, more with more guardrails, with the with the hopes that one, we get standardization across the business within all these different hundreds of teams,
00:12:42
Speaker
But then also so speeding up, hopefully. That's the goal, right? Faster models, faster to build, faster to change. If you get seconded and you go into another team for whatever reason, because it's so standardized, you're not really looking at the completely different world and you have to learn everything from scratch. So that's the the very short before and then hopefully teaser for the after as well.
00:13:05
Speaker
the vision, no I i like that. and and When you joined, you obviously, the team had already yeah been, we're in the weeds of of identifying these problems and starting to design the new architecture.
00:13:18
Speaker
What wast stood out to you most about the the situation when you stepped in and what made you feel sort of ah i think that they're going in the right
Initial Impressions of Monzo's Platform Team
00:13:28
Speaker
direction? Yeah,
00:13:32
Speaker
yeah. yeah Yeah, I think the the the the first impression I had, to be honest, was that the I had a bit of a panic, to to be completely honest, because when i joined, it seemed like they had kind of solved everything. um the The platform team honestly like top class industry-wide.
00:13:55
Speaker
um And when I joined, and I think i've I've spoken to a lot of people who have joined Monzo as well, is that you know from onboarding, everything is very slick. There's a lot of things that are kind of handled for you. But then the other side of that coin, the panic that I alluded to, is that you know, as ah as ah as an analytics engineer, fairly new role, right? This has not been around for very long. The boundaries and and and lines for this role within data are murky at best.
00:14:29
Speaker
For any data role is already quite difficult to to draw specifically. And, you know, from my experience, I've i've i've i've always been quite close to the the tooling, the platform and and and that sort of thing.
00:14:44
Speaker
So when I joined Monzo and I saw the the size of the estate and kind of the scale that we were working at, I had all these ideas. was like, right, I'm going to build this tool to do this and I'm going to automate this bit, I'm going to do that.
00:14:56
Speaker
And within every time that I came up with an idea, i would find a related, already written Notion document on either the solution already being in production or something something that the team was already considering, that sort of thing.
00:15:13
Speaker
So... I felt a little bit, ah I guess, overwhelmed, but then also very impressed with just how much thinking was being done for for for the company. And of course, you'd you'd kind of expect that in hindsight, right, Bruno?
00:15:28
Speaker
Speaking to myself, because, you know, it's ah it's a very successful bank and it is a bank and it's got to have a high standard. And when you know I was looking beyond the tooling and the platform and kind of looking at the the actual data modeling, obviously at the time I joined, we were just talking about the new architecture, but we hadn't actually built anything um within it. ah it was it was quite It was quite stark. Within my career, i've i think most of the time have worked somewhere in the middle.
00:16:04
Speaker
where it wasn't as freeform, if you will, as the previous ones of architecture and ways of working, but it also wasn't as opinionated as the one that we're currently working in now.
00:16:17
Speaker
and This was something that I had to come to grips with and kind of think about quite deeply and understand the why. That was so important when when I joined. I think this is something that we talk a lot about, is trying to pass on that ah part of the the thought process behind the architecture. Like, why did we get here? Like, why did we make the decision to not allow this particular join on this particular type of model. what like what is it If you don't know the why, it seems arbitrary. um
00:16:50
Speaker
But again, like i like I alluded to before, the team has really good a really good culture of of documenting and kind of working through all of their ideas. so that it's easy to access and today with you know your AI friend you can easily search through it and it can come back with really really good answers. um And yeah it was it was honestly quite impressive.
Conveying Architectural Decisions Within Monzo
00:17:15
Speaker
um i was a little bit overwhelmed like I said but very much hopeful to kind of see this new new idea, this is this this new thing, try it out and then you know one day talk about it more publicly and look, here we are. Here we are, no, excellent. You said there obviously about the importance of sharing that why, um how do you go about doing that? Because think it's obviously, I can agree more on how important it it is but yeah, how do you go about sharing the why making sure everyone's living and breathing it?
00:17:47
Speaker
Yeah, and and it's such a good question. i I'm not going to make any claims that I've solved this or we've solved this yet. We're still working on sharing and and and you know making sure everybody's adopting it and and making sure that everybody understands the logic behind it, but also being extremely conscious that this isn't necessarily supposed to be like, all right, we've done. like We're never changing, right? it's it We're supposed to grow and...
00:18:16
Speaker
and learn. But something we're working on a lot now is from kind of looking at the entire, i guess, life cycle of somebody joining the company.
00:18:28
Speaker
ah One question actually that we recently had was, okay, so we've we're technically not building any models in the old architecture anymore. Okay, great. What does onboarding look like?
00:18:40
Speaker
Do we still kind of explain what the old the way the old ways of working is And there's two sides to that, right? You could say yes, because then you have that context. And whenever someone goes, okay, but why can't I do this join in this one model in this one weird you know place? It's like, well, okay, that's because we had all of these other the scenarios. And that's how we got here, right? And that builds a nice picture, a nice narrative.
00:19:07
Speaker
But then similarly, we also want to move away from the old architecture and you know at some point we will entirely. When is that when is when When do we draw that line and say, okay, we in the onboarding even, we don't mention the old architecture? The truth is is that at the moment we still have both living side by side as we migrate, but the process is long and arduous as anyone who's ever migrated from warehouses or architectures would know. um
00:19:39
Speaker
But it involves a lot of presentations, it involves a lot of slack posts, it involves a lot of you almost internal articles and blog posts. um And quite frankly, to be you know very, very candid, I think it involves a lot of patience.
00:19:57
Speaker
um you You need to have the patience to kind of work through the changes with people because they might have not... First of all, not everybody was there in the room when you were going through the coming up with the with the solutions.
00:20:15
Speaker
And that's just natural. That's how big companies function. But you need to make sure that there's enough information out there and that whenever people ultimately meet that information, it comes across with empathy and kind of explains and walks you through the story. um I'm a big narrative
Storytelling in Data Standards Communication
00:20:32
Speaker
enjoyer. So the way that I write, the way that I present is very much with a...
00:20:37
Speaker
you know, it's 2am and you get a pager on your watch and then you get an alert and it's like, okay, this model's broken, but you don't know who owns it. That's why the standard exists to say you cannot merge model unless there's an owner tag, right?
00:20:53
Speaker
So for me, I think that's like an effective way of help sharing that why. But there's a lot of methods and we're still working on trying to get everybody to kind of go through the journey with us as they join.
00:21:08
Speaker
Yeah, no, I think the narratives and stories are a key to to selling, to influencing, to changing opinions so I can see why. and i know John, that's a party who was previously on the the podcast last year.
00:21:24
Speaker
um i believe he's your manager, right? Yes, that's right. he He spoke very highly about you being the lead champion for this new architecture.
00:21:35
Speaker
How have you gone about, you've alluded to some of the points there, but is that yeah like for for other people that are in the midst of a migration or changing an architecture or whatever it might be, how how how do you see yourself being the the an effective champion to two influence?
Promoting and Believing in New Architecture
00:21:53
Speaker
Yeah, yeah it's it's it's a good question. um Which I think the difficulty in answering this question is that everyone has their own ways of approaching things. The way that I like to approach it is if I can understand the why and then if I can then tell that back to someone, then it means that I i understood it quite well. I think there's a quote by someone much, much smarter than me, but which is that if you can teach it, that means that you know you know the that the subject extremely well.
00:22:28
Speaker
And um i often look for opportunities like that of when I could, it's very uncomfortable. It's extremely uncomfortable in the beginning when you yourself aren't fully, um you know, 100% confident in all the different details.
00:22:45
Speaker
But if you put yourself in the in in positions to to help guide others, and that was kind of what I did with this Champions Program thing, um it meant that I, by design, needed to become more of an expert in in it. so you know, it's a lot of reading, it's a lot of going through documentation, but also nowadays we have so many different resources to help you digest information. I'm a big fan of of tools like Notebook.LM that can create a little podcast for you if you. If you're a listener of this, you probably do like podcasts. um So, you know, and you can use it to just put any sort of document or article online and and generate information for you there or create a presentation to help you understand. So for me, that's that's kind of how I um approach this is, you know, digest, try and put yourself in in positions where you'd have to explain that to someone who doesn't know anything. You can do, you don't even have to, you don't really need someone, you can do that on your own and try to explain the topic. um
00:23:51
Speaker
And then, yeah, but I think what's also probably the most important and probably what I should have started with is you know You have to believe that the thing that you're championing is worthy of being championed. And I think in this case it was.
00:24:09
Speaker
Yeah, I think, as you said, the being able to explain something is is so critical into layman's in layman's terms as as well, especially in such, you know, the data being such a technical field.
00:24:23
Speaker
Most data professionals have to deal with non-technical folks on a daily basis. So if you can communicate to them what you're doing in an effective way, then I think you're onto a good good stretch and you can hopefully bring people along on on that that journey.
Overview of Monzo's New Data Architecture
00:24:39
Speaker
So Bruno, um as we said, there is an article online to go that goes into the the nitty gritty of the technical details, but give us a flavour. What does this new data architecture look like?
00:24:55
Speaker
um What are the core ideas behind it? What are you trying to achieve and and why why big yeah yeah Big questions. big questions yeah i I'll go through through some of the core concepts here, but again, yeah as you said, recommend going through the Monzo blogs and you'll see it there for more details. um As I mentioned with the with the before picture, we had a lot of manual processes.
00:25:25
Speaker
So what are our three core principles? It's to be opinionated, formalize the data sharing process, and to automate as much as possible. um And with those three things, we kind of branched into different conclusions, if you will.
00:25:43
Speaker
And with the being opinionated, We came up with these, or the team came up with this, the different layers of of how we model.
00:25:55
Speaker
And these layers are split into four. We've got landing, normalized, logical, and presentation. um Landing uses familiar language. If anyone has done any data modeling before, they probably know what to expect here. Flattening raw events, payloads, making it nice and clean. Here we introduce concepts like event timelines. where you have essentially a full alt audit log of everything that has happened to an event.
00:26:22
Speaker
Very, very long, very good for you to go and audit if something goes wrong. but it's not doing too many transformations here um uh yeah then for the next layer which is the normalized layer we have um the supposed object orientated part of the of the of the logic where we look upstream towards the actual events themselves in the back end and go okay well where these where should these events have been where are they part of one object, right? And I think the best example I can give is always the accounts example. um The way that the backend services work at Monzo is that we have, let's say for accounts, we have one event that publishes account created, and then we have another one for account updated, and another one for account deleted, and so on.
00:27:11
Speaker
So by putting these together and creating one single accounts, object then we basically can trace absolutely everything about the accounts and in the normalized section we have slightly um diminishing i guess uh um grains if you will so we have things like episodes models that basically squish the event timeline so if there's the same um status between accounts for example for several event published uh published events then you can just
00:27:47
Speaker
Squish those together. you have things like your dimensions. So this will be just your your your your information about the specific event and then facts and things like this. And what you'll notice is that traditionally compared to other um data warehouse um architectures, like the traditional DBT one, is that we've not really crossed joined anything yet. So accounts remains accounts between landing and normalized. So these objects stay the same. And then only afterwards within the denormalized layer, which splits into two, there's a logical in the presentation.
00:28:26
Speaker
is where we start to put things together. So that's where we actually get some useful information out there with accounts being joined into you know business accounts and users and lending and and and all these other things. And there in this logical layer, um you can create really, really, really useful concepts.
00:28:47
Speaker
And these concepts are ultimately what go to serve the business. And then finally, the presentation layer is very lightweight, very consumer shaped models that go straight to dashboards or can be used themselves for reporting and can also be used for things like third party integrations and stuff like that.
00:29:10
Speaker
So that's the be opinionated principle. And that's what it where it led led us to.
Introducing Interface Models for Data Contracts
00:29:15
Speaker
Then the formalize the data sharing process led us to creating the concept of interface models.
00:29:23
Speaker
And the idea here is essentially to create a data contract between the team that's producing and maintaining that interface model for the people consuming it.
00:29:36
Speaker
And DBT themselves have some native support um on on on this concept, slightly different here, but essentially the idea is whenever there's someone consuming data from it, you, the the team who owns the the model, can define the terms of when this model is expected to be updated or refreshed or how people should consume the data from it.
00:30:04
Speaker
And then also more importantly, almost the the meta level of this is that if your team has interfaces, you now know to take care of downstream changes if you are making changes to these interface models. So it almost ah works like an exposure, but from model to model. So you're basically setting a contract and saying model B, which relies on model A,
00:30:30
Speaker
is going to be used from by by any team. And and here's the kind of the the contract. And so it works well within a company with about 100 different data teams, because as you can imagine, we're all consuming from each other. And it can be quite difficult when you have 12,000 plus models to make a change.
00:30:52
Speaker
And you know you do your best to try and see if you're going to impact someone downstream or the hope that with this formalized data sharing approach, then we can hopefully reduce these cases.
00:31:06
Speaker
And then finally the Automate principle, and this one's pretty cool. um the The two layers that I mentioned, the landing and the normalized layer, they are all built via an internal tool that
ModelGen Tool for Automation and Standardization
00:31:22
Speaker
we built. So this tool,
00:31:24
Speaker
essentially generate SQL, the tool's called ModelGen, you feed it a nice YAML file. Who doesn't like a good YAML file? With all your different parameters and things, and that can include your you defining your columns, the sources, where they come from, indeed, whether or not this model will be an interface or not. And then you run a very nice simple model gen command on your CLI and then you get your landing and normalized layer layer models.
00:31:58
Speaker
And the beauty of the automation here, there's a lot. There's a lot more to it as well, but the beauty of the automation is that it creates standardization within these layers quite well.
00:32:13
Speaker
And as I mentioned earlier, the idea is that now whenever you go into any domain within Monzo, whether a team you're familiar with or not, you'll see the same shaped objects for that specific team. And that really helps both with communication, but also debugging and I guess cross-pollination of of knowledge.
00:32:33
Speaker
That's really interesting. It sounds like user experience has been such a key sort of part of this design and how you're thinking about ah thinking about this new
User Experience in Data Architecture
00:32:43
Speaker
architecture. I think that's really interesting because definitely think that that's not something that people have thought about in the past or many data teams haven't thought about. but I think as we're moving into our, clearly as you grow to a huge scale, but the the new world that we live in as modern data teams, I'm sure I think user experience from a data platform perspective is becoming so much more important. Was was that intentional?
00:33:09
Speaker
ah hundred percent I think it's inevitable to to try and see user experience as a real lever that you can use to help speed teams up when you have such a large team.
00:33:22
Speaker
you know we we We get professionals of all shapes and sizes from all sorts of backgrounds and we often discuss, even beyond the model gen and the re-architecture, the user experience and and and what that looks like.
00:33:37
Speaker
and you know If there's anyone listening um and if they're considering kind of putting this forward to to their senior leaders and kind of trying to explain that thinking about user experience for data is something that's valuable, the the thing that I would say is essentially that this is the single biggest change that you can make to the the that the speed that the company produces and consumes data. If it's easy for people to be able to develop data and get insights out faster, then everything else is is is ah ah is downstream of that, right? So this is an ultimate, ultimately is a multiplier.
00:34:19
Speaker
um And so, yeah, 100% it is at our forefront. And I would also encourage others to think about it in the same way. Excellent. yeah I think there is, you know, we see convergence ah of roles all the time, but also sort of different departments. And I think that product way of thinking, it yes this it's been a lot long term data as a product has been a um been ah a bit of a buzzword. But I think applying that product thinking and how a product person maybe thinks about the hot the whole life cycle of something is really, really important.
00:34:56
Speaker
What are the some of the the the impacts then that you're seeing? Obviously you've described the architecture, you've described the challenges. ah is Is it working? That's the question, isn't it? That's the, yeah. um it It really is. um i think none of us had any doubts, to be honest, that we would see good results.
00:35:19
Speaker
um Mostly because, you know, a lot of the things that we were we were changing were were were quite obvious in a way, but seeing those results manifest is is really rewarding.
00:35:32
Speaker
Like we mentioned in the beginning, we're about 30 to 40% of the way through, um and we are seeing incredible level of cost reduction. And I want i want to give a small caveat to to the to the cost metric here, because Cost isn't necessarily the thing that you should be kind of too focused on, especially if you're looking at growth and and development. But cost is such a fantastic metric because it it kind of encompasses so many other aspects when it comes to data modeling and data development. um
00:36:08
Speaker
And for us, we've seen more than 40% cost reduction. and you know when you're talking about cost reduction here, are you talking about, like, could you be a bit more
Cost Reductions and Efficiency Improvements
00:36:20
Speaker
explicit? Is this data warehousing cost? Is this time cost? what I think that the cost can be quite a broad yeah in trick.
00:36:29
Speaker
Yeah, it's a it's a great call out. So the the cost specifically here is with the data warehousing costs. So our our Google Cloud Platform bill, bill um but also in terms of of of time, we're seeing a lot more, um I guess, standardization, right? So we're getting a lot less cases of people not really knowing what these models would look like in a different place or whenever you go and help debug a model from a different team member because you have some cross some crossover there a lot faster so these collaborations have the i guess the unaccounted for cost they don't show up on ah on a receipt on the invoice at the end of the day but they're also being seen quite uh quite significantly
00:37:17
Speaker
Excellent. I think one of the things you mentioned there also is that you you you built built a tool internally. You guys must be approaching the scale of the the FANGs, other established banks where your the off-the-shelf tooling maybe isn't quite cutting and you've got such nuances. How have have you found balancing that build versus by discussion as you're scaling and looking ahead?
00:37:42
Speaker
Yeah, that's a tough one. That's a really tough one. um and and And you're right on the money there is again. um We have this discussion very, very often. And the challenge of a company getting to this size is that, again, as I said,
00:38:00
Speaker
it No other company looks exactly like yours.
Building vs. Buying Internal Tools
00:38:04
Speaker
And it's it's so true, especially so for data. So there are so many cases where you know we wish DBT did something slightly differently, or we wish there was a tool that did this for us or whatever.
00:38:17
Speaker
you know Model Gen, you could argue, is us making a very clear decision on the build for for for this particular architecture. um And we're definitely not against the concept.
00:38:31
Speaker
I think we're also very aware of trying to not necessarily adhere to industry standards for the sake of doing so. But we're aware that if we do have industry common tooling and and things like that, there's a lot of transferability right that can happen between those. um But There are cases where we just can't solve the problem with what's with what's out there. And so building almost becomes the only possible option for the for the very specific parameters that we have, right? Which every company has their own.
00:39:11
Speaker
Excellent. Yeah, it's it's always that scale, I think, which comes into it and the nuance. And we can probably touch on that in in a second, actually, as well, when it comes to, i suppose, AI and it's coming for your job. But I think that what you've said is the data and the nuance of data and the data of state, I think we can probably touch on, but it is probably what makes data a very um stable place in the what what whats what's coming with the the revolution.
Lessons from Monzo's Architecture Transformation
00:39:40
Speaker
um Before we move on to AI and workflows and and how you're you're using LLMs as tools, Brudow, what's been the biggest lesson you've learned as part of this sort transformation what surprised you most?
00:39:55
Speaker
Yeah, it's a great question. um i think what' I'll start backwards. I think what surprised me most was just the sheer scale that we were working at and kind of ah reluctantly accepting that the solutions that I've thought of in the past in in previous companies, or much smaller, um don't work here. They're just not gonna work. um And it's for a variety of reasons, but I think just the sheer size and the the amount of data that you' that you're you're you're managing
00:40:34
Speaker
And then also just how many people there are working on the same on the same repo. I think we we get something crazy like 50 or more PRs merged a day. you know like it's it's ah it's It's a level, it's it's it's really, really, really crazy to think about. So for me,
00:40:52
Speaker
the the every to be honest to this day um i i still get surprised um with just kind of how big the ship is that we're trying to steer and then following on from that i think the biggest lesson for me has been really an appreciation for the limits of the tooling that that's out there um and when i say tooling I don't just mean dbt, I mean things like you know google bigquery, the limits that it has. I didn't know it had limits, turns out it does. Just throw some really big petabyte size model models at it and you'll you'll see the limits very quickly. um
00:41:33
Speaker
And so for me it's been it's been like almost ah ah an appreciation again of how interesting the data space is because if if GCP, AWS, Snowflake, you know, all the big ones haven't really solved for this scale and and size that works for everyone out of the box and, you know, then all of our jobs are gone, then really this is a very interesting space to be in. And so for me, the lesson really has been just
00:42:05
Speaker
an appreciation from help for how these tools work and in their limits and that there are limits and it's not all magic. I think there's a lesson in there to thinking about the tools that you're using and understanding how they're working.
00:42:16
Speaker
um yeah The modern data stack has been brilliant but I think there's plenty of people in the industry that just know how to use the tool but not necessarily how the tool works and why it's working in that way and I think the the best engineers that that I've placed and see go on incredible career journeys are the ones that that do understand what's going on behind the scenes in their tooling. So I think yeah just that you've got got an even bigger appreciation, but I think just a lesson in in that as as well.
00:42:46
Speaker
Yeah, that's it. ah So yeah, Bruno, as I have alluded to you, we've withve known each other for quite some time, you were actually kind enough to speak at our meetup, the London Analytics Engineering Meetup, last year and give your two cents on LLMs and AI, particularly when it comes to inside the AI workflow.
LLMs Enhancing Productivity in Data Workflows
00:43:08
Speaker
um I loved your takeaway and and mine was was that an LLM is is literally a tool like we've been discussing behind and there are mechanisms behind it that have its limitations, etc. So yeah, um for those that didn't make the meetup,
00:43:25
Speaker
what you're doing, sign up to our meetup page. and um yeah Could you talk to us a bit about how you personally use AI to in your your workflow and and what are some of the real examples and and benefits you've seen? Because I think it's something that that everyone obviously is is needing to grasp as these tools become more prolific and important.
00:43:46
Speaker
Yeah, 100%. I'll first start by also shouting out the the talk, the meetups. Do go, they're incredible. And it was really an honor to speak at them. So thank you, Cognify, for the opportunity. um Yes. ah How do I use AI, LLMs, MCPs, God, all the buzzwords. um Yeah, I think starting from a place of understanding that they are tools and not magic. That was kind of the title of my of my talk last year. um They really are tools. um They're really good tools and they can do a lot if you are willing to put the time in to actually get the most out of them.
00:44:31
Speaker
I think the example I gave was um somebody using a pencil in my talk, which you know you can use. I'm not a particularly good artist, so I draw very poorly. But someone with the same pencil can create this beautiful work of art. And I think it's the same thing with LLMs.
00:44:51
Speaker
At the moment, I think not just within the data industry, but just probably all of all of the humanity are just learning what this new way of working with this non-deterministic computer program, if you want to call it, which is ah um almost seems like an oxymoron to say. um And we're all still in the discovery phase.
00:45:15
Speaker
trying to understand what the limitations are and what we can do.
Unrealistic Expectations of AI and LLMs
00:45:20
Speaker
But then the other side of that is that people almost expected to do too much. And I think back in 2022, 2023, when GPT came out and we were all amazed and wowed by all the different things that it could do, we almost expected it to continue going down this path of, oh, we won't need to do anything anymore because AI will just figure it all for us and I don't need apps anymore because I'll just like write a prompt and it will go do the thing for me.
00:45:48
Speaker
We've not really seen it get there yet and you know I'm not a a pessimist per se I do think that maybe one day we can get there but I think it's it's slightly further than perhaps some some people might lead you to believe. And if you then subscribe but to that philosophy and you see this as a tool, then you don't feel bad about spending time tweaking um your LLM buddy to to become better. And you can actually create some really, really cool things. um and and And here talking about you know the my day-to-day usage, um I think in platform,
00:46:32
Speaker
I use it a lot less when it comes to actioning, you know, like a data model or something like this. But when I was in ops, it was a superpower that I had, but plainly.
00:46:44
Speaker
um Not only was i was i able to go in and and generate code, high quality quote code, very quickly, but it gave me this ability to, you know, not only do more than one thing at once, which is is the LLM agentic dream, I suppose. But it also gave me this really, really, really powerful ability, which I spoke about in my talk, which is to be able to pivot.
00:47:14
Speaker
I think everyone who has been an IC for a long time or has gone down the IC path from all senior routes has had this feeling when it comes to working on something for a long time. And then you go to your stakeholder and they go, oo actually, can I do this thing instead?
00:47:33
Speaker
And for as much as you're trying to hold that smile during that Zoom call, in internally, you're kind of going, Oh boy, like I don't want to do this. No. Like why are they asking this? It took me so long to get this data together. Don't they know that the thing that they're asking for lives in like four different tables and all different granularities? Like how are we going to get this data in there?
00:48:00
Speaker
Nowadays when that happens, I go, okay, cool. Yeah, yeah, let's do it. Let's pivot. Let's go back and and and and start again. And I think even more proactively is that I can kind of do that myself.
00:48:12
Speaker
um This tool and this new paradigm I think there's there's been lots of articles talking about how software developers are no longer developers, they're more like software reviewers. So you're reviewing code a lot more than writing. But the truth of the matter is, is that if you are reviewing code more often, then you have a lot more time to think about the what you're doing and how you're going to do it and whether or not that's best for your specific use case. And that ability to...
00:48:43
Speaker
you know, detach from the work that you're doing and go, do i need to rethink this? Do I need to pivot here? Do I need to Would there be a better way to do it? And removing that barrier and friction to starting again or tweaking, ah genuinely, genuinely, genuinely to me feels like a superpower. And I think, you know, especially at the beginning of last year when I just joined, um people kind of did talk about me in that way. And I would very quickly go, have you heard of our Lord and Savior Claude Code? or Gemini CLI or Codex, whichever one you're using. And it's like, yeah, there's no secret here. i'm not a superhuman. I just have a really good tool.
00:49:30
Speaker
um And ah yeah, I, again, in my day to day, it is a daily thing. It's like, it's like if I was to ask anyone like, oh do you use the internet on your day-to-day? is like, yes, of course I do. So even within platform, though I'm doing a lot more code code and platformy things, um I still use it every day and it's really, really valuable.
00:49:59
Speaker
Spend some time, build out your your your niche and and how you want your agent to work with you.
Tailoring AI Tools for Individual Workflows
00:50:06
Speaker
Obviously help others as well if you can to to help speed them up, but you can really get some really good use out of it.
00:50:13
Speaker
Amazing. i think it will it sounds like it gives you a bit more of this almost like elasticity to think about how you're building something rather than just getting your head down into the weeds and just sort of focusing in you can take that step back to be a bit more malleable with your thought phrases a bit more philosophical of you know is this the right thing to do and then the execution is um is that lower lower barrier of of entry done by by the agent
00:50:44
Speaker
um any Any tips for anyone that's that should just style is is looking to ah to up how they they use Claude or at any of the LLMs?
00:50:55
Speaker
Yeah, yeah. um I think the the first thing I'll say is that you're not behind. The feel changes so often that whichever day you decide to pick up an article or just open up Claude or whatever it is and start trying things out, you're you're not behind. Things change very often. so go go try things, go play. Don't have that fear of... um not being able to explore, because I think that that's the best way that to to to to get that experience and and try things. The other thing that I always recommend, and this one's a bit meta, is Claude knows Claude very well. Gemini knows Gemini very well, and GPT knows Codex very well.
00:51:41
Speaker
If you don't know, don't open up another tab and go you know Google, a much worse search tool called Google, Ask it there. It knows how to do the thing. What is it that you want to do? Oh, I want this thing that automates how I pull data from this place, this place and puts it into a into a report.
00:52:00
Speaker
Okay, go go go speak with it. Go brainstorm and and see how how it can do it. And it knows its capabilities very well. Although... It won't necessarily guide you and steer you as much as a human would when you're going down the wrong path.
00:52:16
Speaker
But for a starting point, it's pretty good. just Just go play. um And I speak a little bit about this in my in in one of my blog posts at, quick plug, brunofields.com, which is the different levels of the... um agentic ladder, if you will. And one of those levels is specifically this like Ouroboros of knowledge consumption where you're no longer dependent on external sources. You think of a concept and you ask it to tell you if it's possible and how it can be done. And then you just get into this positive feedback loop of just, I can create anything and you can learn so much from that.
00:52:59
Speaker
Yeah, I'd say yeah I've never written a load ah line of code in my in my life, but um yeah, i from I think it was your advice of telling Claude that you're a complete novice and start start me at the beginning, we've been able to build some some pretty cool applications and automations even Cognify. So yeah, for someone that doesn't code to be able to do that, um if you can, then it is powerful.
00:53:29
Speaker
um So, um where does this, how does this fit into what a new AE should look like, Bruno? um You've said like, yeah do you use the internet?
00:53:41
Speaker
um Of course, like what, how is this changing the skills that an AE should be developing today?
Understanding Concepts Over Syntax in AI Industry
00:53:49
Speaker
What matters more in this AI driven world?
00:53:52
Speaker
Yes, yes. um For me, It comes back to a point you mentioned when we were talking about knowing the limits of the tools.
00:54:04
Speaker
I think the the the the most important skill for you to have is understanding the fundamentals. If you no longer need to spend 80% of your career memorizing syntax and remembering how to um you know, do different things with specific languages and and and do all that sort of stuff and like what what syntax I should use for this package and which package I should use for this solution.
00:54:32
Speaker
it frees It frees up a lot of time. And I think where you should be putting that time is in really understanding the fundamentals. um And that can go down to really low level tech understanding of your data warehouse. How does it actually partition data? How does it work?
00:54:53
Speaker
How do you get this information the fastest, depending on what your role is. um To me, that is the most important part. If you're business facing, right? And you have a lot of ah stakeholders that are either the the customers themselves or or PMs and things like this, understand the product better, understand what it is that you're working on better. Like that is what this frees you up for.
00:55:20
Speaker
And that is ultimately to me what I think will create a, I suppose, a a shield, I guess, against this very, very, very rapidly changing tide and and and paradigm, I guess, within not just the analytics engineering role, but all of data.
00:55:46
Speaker
um Yeah, that that that to me is the most important thing, is understanding the fundamentals, understanding your tools very well, where are the limits? And then if you are on the customer-facing side, really, really understanding your product and just kind of, you know, being able to leverage this as a tool, like as you would any other tool.
00:56:07
Speaker
where do you think the industry's going then yeah What's this going to evolve into? And I'm not going to hold you to account on this, as you said, it's changing changing month by month. But yeah, getting a bit philosophical, where do you see the industry going and the part that I suppose analytics engineers and i suppose wider data folks play?
00:56:30
Speaker
Yeah, I guess, yeah, check in on me and in a couple of months and see how how right I am. ah um But that's a question that we all are thinking, right? I ah spoke with my very good friend Toby Henley-Smith, who also did a talk for Cognify at the beginning of the year.
00:56:51
Speaker
Great talk. If you were there, you would know. And one point that he brought brought up in his talk was, what at its core, what it what is an analytics engineer? And he gave a very great analogy, which was, if you removed Figma, do you remove the designer? Is the designer still a designer? What is it that makes a designer a designer?
00:57:10
Speaker
And I think the thing that we both agreed on quite fundamentally is that we are more than our tools. um And if LLMs and agentic CLI tooling are indeed just that tools, then you should be in a position where you're quite resilient to these changes because what you provide is still valuable, which is insights and thought and and architecture and infrastructure, but in in in a very broad sense of understanding what it is that a data platform needs or a data insight is for the business.
Adaptability in Technological Evolutions
00:57:54
Speaker
And that's something that I think won't go away. The other, um I guess not not really even an analogy because this is exactly what did happen, which is that back in the days, 50s, 60s, even before then, part of that the role of an analyst existed then.
00:58:13
Speaker
And part of that role was to do spreadsheet. Now, for us, when we say spreadsheets, we think Excel and a computer that goes and does a lot of things for you. But back then you actually had to draw out on a piece of paper with your your little pen and actually draw out this whole spreadsheet. And you know yes, there were people who were really, really good at doing that and really enjoyed that element. And they were very, very good at producing those those insights.
00:58:43
Speaker
But then with the advent of computers becoming more common and spreadsheets being able to be done digitally, the role of the analyst didn't disappear. It's still around. It's just that the tools that the analyst uses has changed.
00:58:57
Speaker
And I think similarly, i don't know what the future is going to look like, right? I don't know how good LLMs are going to get within the next one, two, five years.
00:59:07
Speaker
But what I do know is that for as long as humans are still the ones mostly running a company, then that humans will still require translation, insights, understanding of of of business goals and and and understanding the mathematical side, like the events and the data points and translating that into into business insights. That's still going to be needed. And so for me, I think the The what you'll be doing will be completely different probably in five years time.
00:59:43
Speaker
um But the the why you're doing it is probably still going to be the same.
00:59:52
Speaker
Fascinating insight and I think a nice optimistic look compared to what you do see out there. And I couldn't agree more. yeah the Every industrial revolution or change in tech technology technology has always disrupted areas, but yeah it's the roles have adapted and persisted. um Everyone's thought there's going to be an extinction extinction of a role whenever there's been a big technological advance, but there that hasn't always been the
Critique of Data Industry's Maturity
01:00:26
Speaker
case. so
01:00:27
Speaker
Bruno, it's been a pleasure to have you on. Before I let you go, um what's were your most would you say your most controversial view on the data industry itself?
01:00:40
Speaker
Ooh, that's a great question. I have so many of these hot takes. I'll give you one ah that I think is a little bit spicy, but I also think if anyone listening can take action from it and and kind of help improve this, it would it would really help.
01:01:02
Speaker
Which is ultimately that I think data is still incredibly immature and um I think we need to start treating it like a first class citizen within companies.
01:01:16
Speaker
There's an old adage, I think every data team I've ever worked for has said, no, date we're not ah we're not a service team, right? It's not like ah we don't get like these tickets and then we do the thing and then we put it back and then we have angry stakeholders, right? We're gonna try and be there close to the product, close to the business. We're gonna try and bring these insights proactively.
01:01:38
Speaker
And then you go and check in on a day to day of what you actually do as a data person. And it looks a lot like a service. um And I think it's because we're we're we're still really, really immature from a discipline point of view. ah the The other point of evidence that I have here is that working in data from company to company changes drastically.
01:02:02
Speaker
I have friends who are called data analysts who do more data engineering than I do. And I've got friends that are called data engineers who do more purely data analytics. Like even the fact that we can't agree as an industry on what these roles are, and I'm not even touching on analytics engineer because, oh Lord. But the fact that we still kind of- We're engineers this month. Right? Like, what what do we call them? the the the the the point is is that we still don't really
01:02:35
Speaker
have ah this this stable foundation um and I think we really need to have it. um I think you know we often compare ourselves to our cousins over in the software engineering side and we often try to replicate by titles and by role style, right? Like data engineer and and and all of these other things. um And I think something that they've nailed and done very, very well is they've created a good set of expectations within the roles. So a back-end engineer does back-end stuff and a front-end engineer does front-end stuff.
01:03:14
Speaker
But within data, there's still a lot of cases of, and if anyone out there is listening can relate to this, okay, yes, I got hired as an analytics engineer. but you're the only data person in the team at the moment, and you are expected to go pull some data from some... Yeah, go pull some data from some dashboard and then pull up some, set up an experiment for us real quick, and like, can you provide some, a report on an analysis on this thing and so on? And again, it's not even...
01:03:46
Speaker
the specific analytics engineer title or data engineer or whatever, it's just because you have data in the name and your company probably doesn't really know what that means. Like, oh, you're a data scientist, i don't really know what that means. Or like, oh, you're an analytics engineer, okay, you've got analytics, cool, that means you do analytic analytics.
01:04:00
Speaker
um For me, I think that's probably the most frustrating part still of being within the data space. But something that I see improving, and again, a huge shout out to Cognify, to you guys, because I think things like the meetups for analytics engineers,
01:04:20
Speaker
Very recently, you also had the inaugural data science and AI meetup. Also recommend everybody sign up to that. um You know, these spaces create a sense of community. And whenever we can go there and talk about these topics, it slowly, slowly creates this middle ground of like, OK, we can all agree we all do this thing. And ah these meetups and and community events that you guys organize help accelerate us to get to a position of a more mature state within the role for the data discipline. So biggest pet peeve, I absolutely despise this. I'll probably never stop talking about this because it won't go away in my lifetime. But I'm very thankful for Cognify and for people that set up these ah community events.
01:05:11
Speaker
Yeah, i I think that is a really good point. and You can imagine i i I also deal with that on a daily and weekki weekly basis. I would agree that I think sometimes the challenge comes in the type of the organisations and where they are in their data maturity. And it's actually sometimes it's I think it very much is to blame with on the data industry because so many people outside of the data industry have no idea. I spoke to someone recently, Chief of Staff, who said, i mean i need a data person.
01:05:45
Speaker
OK, let's break it down what do you actually need here. And I think um there needs to be a bit of education and a bit of accountability on the data. organizations and departments to help help educate, which comes back to the championing, which is maybe a nice ball loop to to leave it on. There we go. Yes, 100%.
01:06:07
Speaker
Bruno, it's been an absolute pleasure. Thank you so much going on and sharing. And yeah, if you've enjoyed hearing what Bruno's had to say today, I can recommend his... Is it Substack?
01:06:21
Speaker
It's just my own website, brunofields.com. That's it. brunofields.com. Head over to to and and follow that. I'll also put a link to his LinkedIn as well, as well as the Monzo blog about the architecture that we've we've discussed today. But yeah, for now, Bruno, um it's been a pleasure.
01:06:39
Speaker
Thank you so much, Harry. And always a pleasure working with Cognify. Brilliant. Thanks, everyone. And we'll see you in a couple of weeks. Hi everyone, just a quick one from me. If you've enjoyed today's episode, I'd be so grateful if you could hit that follow button or leave us a rating.
01:06:55
Speaker
Even better, pass the show on to a friend who might also get some insight from it. It really helps us grow the community and continue to share amazing conversations. I also wanted to take a minute to talk to you about Cognify. For those of you that don't know, Cognify is the leading recruitment partner for modern data teams.
01:07:13
Speaker
We help some of the world's best organizations scale data and drive real value from the hires that they make. If you're thinking about building a team or making a hire and you're struggling with talent or just want some insights on the market, then I'd love to jump on a call with you and tell you a bit more.
01:07:31
Speaker
Equally, if you're looking for a job and want to find your next dream role, then reach out to myself or any other Cognify team. We'd be happy to see if there's anything on our books that we can help you with and give you general advice on the industry.
01:07:44
Speaker
Finally, big thank you to Omni, this season's sponsor. If you'd like to learn more about the AI analytics that Omni can deliver you, then check out the link in the show notes or come speak to me. i can happily point you in the right direction.
01:07:57
Speaker
Again, thanks for listening and look forward to seeing you a few weeks time.