KISS-ing Your Data+AI Platform: Rapid Multi-Cloud Migration to Databricks

Summary

Adevinta migrated its global data and AI workloads from multi-cloud DIY stacks on GCP and AWS to a unified Databricks Data and AI platform in under four months.
The team moved AI/ML models from SageMaker, GCP, and Java applications into Databricks using MLflow for serving and monitoring, reusing only 20% of existing code for planning models.
Unity Catalog became the central artifact hub for models and dashboards, and Genie-driven AI consumption grew to 12.5% of total compute across the organization.

KISS-ing Your Data+AI Platform: Rapid Multi-Cloud Migration to Databricks

Faced with a strategic shift toward marketplace autonomy, Adevinta transformed its global data+AI strategy by migrating from complex DIY stacks on GCP and AWS to a unified Databricks platform in under 4 months. this video, led by our Principal Architect and Product Director, provides a deep dive into how we overcame technical and organisational "old ways of working" to achieve this high-velocity transition. Beyond data engineering and data warehousing capabilities, we will focus on how we migrated our AI/ML platform to Databricks, demonstrating how a unified foundation in Unity Catalog and the Databricks AI stack has accelerated our migration and time to production, and how this foundation is now helping in our ambition to unlock even more value from data+AI across the company.
Talk By: Iker Martinez de Apellaniz, Head of Data, Adevinta

Chapters

FAQs

How did Adevinta migrate its multi-cloud data platform to Databricks so quickly?

Adevinta used a lift-and-shift approach leveraging MLflow to migrate AI/ML models from SageMaker and GCP environments. By applying the KISS (Keep It Simple) philosophy, the team focused on consolidating rather than rebuilding, completing the core migration in under four months.

What role does Unity Catalog play in Adevinta's Databricks deployment?

Unity Catalog serves as Adevinta's central artifact hub, governing models, dashboards, and datasets across the organization. This unified governance layer accelerated migration timelines and provided the foundation for certified Genie space adoption at scale.

How did Adevinta handle migrating AI planning models to Databricks?

Adevinta's planning models required only 20% copy-paste code reuse from their original implementations, significantly reducing migration effort. The team adapted model logic to the Databricks environment rather than rewriting entirely from scratch.

How is Genie being used at Adevinta after the Databricks migration?

After migration, AI consumption driven by Genie grew to 12.5% of total compute at Adevinta. The team implemented certified Genie spaces with defined scope, quality standards, and CI/CD promotion processes to govern self-service adoption at scale.

Full transcript

[00:08] Hello everyone. Welcome to this to this talk. Um I can now say that I'm doing a a bingo speaking in conferences. I've been the first session in the morning. It was awful. Everyone was sleeping.
[00:20] Just as I'm after lunch, everyone was sleeping. This is kind of the last one or previous to last one. So hopefully half of the audience will be sleeping only. >> No, we're awake.
[00:32] >> So again, thanks for joining. Uh this comfort this talk is around transformation that we happen in our company about the choices we made and
[00:44] how we decided what is the best platform for our needs, which is actually multiple times this decision as as you will learn, right? It's not about kissing. It's not a sexy platform, at least not in the literal expression. It
[00:56] is sexy to hear. So before starting, my name is Iker. I have an engineering background. I was a data engineer and an enabler. Then I was a data foundations product leads. I was a
[01:09] product manager of of um how we were building our own data platform for well, multiple teams, multiple people. The last official role is principal architect and product director for data.
[01:23] It's a very long role. And actually I should have scratched this and now I'm a head of data for Kleinanzeigen. So to add more roles to the title. Like um I work in Adevinta, which is a multinational marketplace company. So if
[01:37] you can think of Craigslist, it's like owns multiple Craigslist in Europe and I will explain what. I was working while this presentation on the focus of this in Marktplaats, which is one of these marketplaces in the Netherlands. And now I'm doing the same in
[01:49] Kleinanzeigen. So Adevinta, what is Adevinta? It's again uh European focused company of marketplaces, multiple brands, very strong big parts of the population are
[02:01] using our brands everyday, right? And we have, for example, those brands like Leboncoin in France. I don't know if there is anyone from Leboncoin. No, they would they are around in other talks. Have a Milanuncios Fotocasa in Spain,
[02:13] also very popular. Kleinanziegen Immobilienscout24, really big companies in Germany, right? And the closest one geographically would be Kijiji in Canada. And I don't know if there is one from Canada, but it's a
[02:25] video one. We used to have also Segunda Mano in Mexico, which may be more geographically close, but we sold that one. Historically, Adevinta has been buying and selling marketplaces, right? So,
[02:37] what happened actually, this is the picture in 2024. What happened last year is that this is my company, Blackstone and Permira, they joined and paid $30 billion for dollars for Adevinta. So,
[02:49] they own the They basically bought the company and as you can see, our direction, which was like, "Okay, we are going for a big marketplace, big corporate platform." It changed a little bit
[03:02] drastically. And I was like, "Okay, do completely the opposite of what you were doing." What does this mean? We were, back in the time, 500 data professionals working as one,
[03:15] like one unit, using one platform, working as one system, and consolidating everything into one. We have 100 Kafka clusters, 100 unicorn clusters, which are Spark,
[03:28] Kubernetes, things, right? We have our own team running this, our own team running Airflow on top, managing all of this. But with this strategy change, suddenly we have eight marketplaces, which are the focus of this strategic
[03:41] change, that need to learn how to operate in isolation. So, stop talking with your French friends. Actually, saw them after 1 year here. Um They need to run their systems they need
[03:53] to learn how to govern and leverage the data, right? So, eight times last year, we did this the same decision that I will explain here of where do we put the data?
[04:07] Where do we put our ML models? How I do these things, right? How do I develop agents? Which tooling I use? And even more like, do I keep a slack or not? Do I keep conference or not? Do I
[04:20] keep GitHub or not? So, even more. But that's a different story, maybe for next year. So, let's focus on Benelux. I was head of data or something something Benelux interim helping us, right?
[04:32] In that moment, because on top of this, you are restructuring your organization, suddenly you don't need as many people. So, the target operating model completely changed. So, you okay, so we don't have a data platform team, we
[04:45] don't have an analytics team, we don't have a head of data, we need to hire this person. So, we need to select how we are going to work now.
[04:57] By the way, we have three, four, I don't know how many clouds. We have some people were doing this with GCP, and they were migrating to AWS, but now they are doing something else. Some people are doing AWS, some people I'm not sure, some people are using Snowflake, whatever.
[05:11] And everyone is was working on the different way. So, because there was a big in the same country in in Benelux, when in in Amsterdam in this case, we have some people that came from a global perspective. So, I work for a multinational company, and some people
[05:24] that were working on the marketplace. I work for Mark Plats. And it's not the same culture. Basically, something that was happening is like, "Hey, you are the head of data
[05:37] science, help me migrate this information on these data models, and you own them now." It's like, "No, I don't own this. Head of data science, I'm not doing ML. You are the closest person. There's no one in ML in this in the whole company. You should own this, right? So, it was a
[05:50] fun moment. Specifically, speaking out of ML, we had Unicorn. This is our fancy logo for the 100 cluster we were managing.
[06:03] We had some Databricks, mostly in data warehouse, building analytics, some models in Vertex AI, and some models in SageMaker. Is there any other solution because probably we also had it, right? I don't
[06:15] know. Any other ML solution I'm missing the puzzle? So, we said, "Let's keep it simple." KISS is from Keep It Simple, Stupid, right? Let's keep it simple. Let's
[06:28] reduce and let's try to do it with Databricks. If Databricks works, why overcomplicate our lives? Why am I imagining we came from a team of 20 people managing ML platforms to
[06:40] five that are doing ML platforms and data lakes and access control and governance? You cannot run all of these, right? So, keep it simple. Just try Databricks. If it fails, you
[06:53] can go to try Vertex or whomever. And that was because, of course, we were under ongoing term changes, data operating model, collective dismissal processes. So, the team that was running previously
[07:06] the recommender and some of other all of these ML models, basically, they were laid off. And they were going to be laid off in a couple of months. They were handing over in everything they had to say like, "I don't I don't
[07:19] want this. You run it for me." Right? So, we needed to find an alternative. Like, where do we put this? This was in in Vertex AI. It was in GCP. Like, no one knew how to run this. This
[07:32] was written in Java and everyone was using Python in that company or the opposite. And again, we don't want to run anything in GCP. We need to remove our footprint in GCP. So, we said let's move it from Vertex to Databricks.
[07:47] Right? And if this works, well but let's move it in Databricks, but there was a team say like, we are going to do this very urgently, so I will get
[08:00] it and put it in my random Java application, random random I want to run because it's what I know. I will say like, but this is not the correct way. Cuz if you do this, 6 months from now
[08:12] we need to migrate it again somewhere. So, why don't we try to do it already in Databricks? It was a little bit of a either you to do it faster than myself
[08:24] moment. Right? You know the results, I don't need to spoil it, but we won.
[08:38] have time to retrain it or do anything like this. We It needs to work for the current workload. This is again many many people using this recommender every day. And we need to learn because whatever we
[08:50] do now needs to become the blueprint for 24 models in the line, right? And of course, we need to deliver the value with ML. So, it's not like we do ML for the sake of ML. We need to prove that this is actually giving us a good a
[09:03] good result and a positive ROI. And again, I was saying we want this to become the blueprint for many other things. So, if it works for our models in Vertex, can it work for
[09:17] our models in Unicorn or our models in Sagemaker? Will this become the golden path of I don't know, 20-something machine learning models live in production at the time?
[09:33] Okay. So, we have to do couple of checks first. The first one is we need to serve exactly the same model without changing. What's the benefit?
[09:45] That we are using an a standard an industry standard open source for creating the models. Guess which one? MLflow. So, we are doing these things with MLflow. Oh, looks like Databricks is using
[09:58] basically MLflow. So, it's kind of a plug and play of the model. We don't need to put it move it from here and put it there and see if it works. Right. And this is what we did. Using the Mosaic API for serving MLflow.
[10:12] So, basically we only change the model interface to say like now this is the new API, this is the old API. Does it work? Does it Does it hold the load, right? So, we could handle the migration the
[10:25] urgency of the migration is just lift and shift the same thing in a different place, but it's the same thing. And we don't need to rebuild the model because no one wanted to rebuild the model, retrain, the info, etc.
[10:38] So, okay, now it's in production. Now we have it. What happens? Will it hold the the the load? Because there's many people querying this model, so will it work? Will it sustain, I
[10:52] don't know how many requests per second. What's a big question, right? The answer is Yeah, actually it did, right. So, yeah, well, I forgot about how many requests per second.
[11:05] Will it scale up and down? Cuz we don't want to pay more than we want. Well, yeah, it does. Right. The answer to everything is yes. Uh How do we observe this? How do we
[11:18] monitor this? Like, okay, it's compatible with what we are doing. It's compatible with Datadog, so it actually works. It It also works with Prometheus and with Grafana, so yeah, we can plug and play it with our systems and don't need to use a a third-party
[11:31] option or a second a new thing that we don't want to use, right? So, suddenly we have three models deployed without any migration of data or the models, just they are running
[11:43] and awaiting queue of 10. So, 10 more. And so far, so good. We have the models plug and play. We are using MLflow. In Unicron, we were also using MLflow as the standard, so
[11:56] this is amazing. We can just put the same artifact. It's like USB-C. USB-C, everyone is happy, right?
[12:08] But then, we enter the moment of this model is not something that you have it statically for 1 year. It's something that we need to retrain. So, how do we now retrain one model, put it again in the same system,
[12:20] and check that it's doing equally or better, if possible, better. Right? >>
[12:34] use this. Let's create a template to reproduce the same the same model and the retraining. Right? And where is the data? Unity Catalog. So, okay, the data is in catalog. We have the DBX asset bundle.
[12:48] Perfect. Looks like it could work. Right? And then we started using the experiments and the models from Databricks to to do this. And again, it was a fast iteration
[13:00] that we could reproduce the same results from the from the previous stack, so perfect. And basically, we are not adding more tools. We were not adding anything else. It was quite simple. And again, this is the goal. You move from a
[13:12] very complex situation. You don't want to put problems in people's lap. You want by basically them to be happy, be easy, and go home at 5:00 if possible or before.
[13:28] Okay. So, the result. We did this, we migrated many problem models. But, does this work, right? So, couple of things. Everything is in in the Unity catalog, so
[13:40] good. And before that, 1 year ago, we were not using that much the Unity catalog, we were using it just for tables and analytics and yeah. And now we are starting to see, okay, this Unity catalog actually it can hold
[13:53] more than just tables, right? It's more than the Glue schemas we used to have for a list of tables, so yeah, it is starting to look nice. Um, we actually have to expand the catalog to hold ML models as a new
[14:05] artifact in the catalog. We didn't have a folder for them, so it made us think of of this. It was a fun moment, very interesting, when we discovered the inference tables. People were saying like, "Ah, this is better, this is worse, the
[14:17] results are perceived as worse." Like, okay, where is your accuracy numbers from the previous model? How was it working? Give me some observability, some monitoring, something, something. Well, we don't
[14:30] have. Then how do you know that it's worse? It feels worse. And we discovered the inference table. Look, inference tables for free, because we are publishing this with Mosaic in
[14:42] the December flow models. We know if it works or if it doesn't work, and I can tell you and point with my finger to what is working and what is not working. If you cannot demonstrate something
[14:54] similar or drastically different, I will trust the data, not your opinion. So, nice moment, nice discovery, and I think they are really enjoying this, because before that, basically, we were blind
[15:06] with with many of the ML things we are doing. The This is one is which is also like we were implementing jobs, we were implementing dashboards, auto fit, we
[15:19] could show the dashboards to the people, and this was again in the same kind of package. Before that well, good luck trying to find a dashboard that shows you if if if it's working or not, right?
[15:31] Um we could evaluate the models with the tables in a spark very easily also like something that before it was impossible you have to create a notebook connect and basically it it was not ready.
[15:43] Not only because of the tooling is because the organization was not using the time. Now I have this this phrase of if it's a 5-minute task there are no excuses not to do it. So because now it's a 5-minute task maybe it's a little
[15:55] bit longer everyone was trying to do it was doing this kind of looking at the inference table, looking at the dashboards and looking at the retraining and trying to retrain and improve and change some parameters more frequently.
[16:11] Okay. The result as you can see is like it worked, right? So now there was a question because I told you like we have to do this eight times. So okay, Mark Blat check.
[16:24] So now planning second. Will it work for planning second? Will it be the same? Can we reproduce this? Right? And this is a moment where when I was in the giving this talk in Madrid this is this was the last slide because
[16:37] we were in the process of creating it. The short answer is like yeah, we change some things. We are not serving it in in data bricks because that part because of company policies they want to save it somewhere else.
[16:49] Or that's the official message because we are actually serving it with data bricks. Just don't tell the people in in the DevOps team. Right? So it worked and it's the same like we could migrate the same kind of reproduce
[17:01] the models, migrate them and what it was pain before with manual looks now it becomes quite simple. Now the consolidating of GCP models into vertex
[17:13] and and some others into data bricks which was a migration process and heavy. We didn't even notice this time. It was something that the teams did in last BAU.
[17:25] And quite autonomously, too, because we have the previous bundle, copy-pasted 20% of it. Same person did it in both in both times. And basically, yeah, it it's the same. It's copy-paste of the
[17:37] thing. Also, because we were using previously the standards. So, I mean, if you're not using MLops MLflow, you're using something custom-made, good luck, right? So, start with your standards. And that's what I like from Databricks
[17:49] and Sergio in the background can connect to see this. So, again, spoilers, it can, right? We copy-pasted the majority of it. We deploy the same models or maybe
[18:01] similar models in Klein as I am, but on the Klein side infra. And the good thing is that you can point where you can um deploy it, so it's kind of a switch or an if case in the code to do this.
[18:14] And what I was mentioning before, the teams did it alone. We were enjoying other problems that we still have in other parts of Databricks.
[18:27] the thing that is happening with ML is how many talks about pure machine learning MLflow are in the agenda of this conference versus AI or agents.
[18:40] MLflow is from the past. No one wants MLflow or ML anymore, like a minority of us, right? Compared to how much we can do with AI and sunny AI. So, now the next step and the next um
[18:53] challenge for us in Klein as I am is actually to can we create the equivalent for these uh standards, reusability, self-seriousness into the AI world, right?
[19:06] And we took one of the problems, which is probably or similar in many of the companies. I was attending many of the talks is like Genie and the analyst querying Genie and getting the insights, right? So,
[19:22] I took a picture of how we are consuming data. This is an my own interpretation of consumer consuming patterns and the outputs of our data platform in January. So, the
[19:34] AI consumption, which is people reading with cloth or reading with Genies or equivalents, in January it was 3%. Okay.
[19:46] In April it was 7%. So, okay, 3 months, yeah, duplicated. And when I was doing the slides, it was 9% in May. This is the cost of AI. This afternoon
[19:58] was 12%. I checked quickly before joining. 12.5% of our cost is already because of AI consumption. And part of this, to be honest, is ML ML
[20:11] consuming it also. So, not purely AI as as Genies. So, inside the company already, like, when I speak with the people is like, "No, we need to do more dashboards and migrate." And it's like, "You think you need to do more
[20:23] dashboards, but what they are asking you is to have Genies." We we say Genies. I don't know which kind of Genies is now. Is agents or is it spaces or Genie agents? We need Genie agents. Yohoo! And Genie one.
[20:37] Right? So, it's a change of culture. And this is data that I demonstrate is like, "You are not consuming the data in the same way that you were consuming it before. You are not looking at dashboards up, down, left, right. No.
[20:50] You are looking at, "Why is this going up? Explain it to me. What is the trigger?" So, it's a a little bit of a paradigm change. >>
[21:03] it's so shiny, it's so nice that everyone is using it because it just get a genius 1,000 questions. It replies every single time. Sometimes it's good, sometimes well, sometimes bad.
[21:15] It replies. And what happens if you have a free bar of sweets or of of potatoes or of Imagine eating like in a conference for I don't know one month in a row. Well, you can get the result, right?
[21:27] So, we need to control this, right? The challenge that we have in the team now that we are that we recently faced it on is going to is still going to be ongoing is how
[21:39] do we fight this carrot that is people using AI super fast, super easy, giving me something like dopamine of of of the reply. How do we fight it so it gives
[21:52] you the nice sustainable healthy part of it, right? So, we need to govern this. And we decided to do couple of things.
[22:04] Apart from the very fundamental good data, which is our colleagues in the other side working is we need the standard interfaces. So, you cannot use Gemini and Cloud and Codex and any other
[22:18] random tool to go into Databricks and query. Like I mean, you can, but in our company is like, "Please don't." Right? Because we have a better method for this, which is Genie spaces. With Genie agents, sorry.
[22:33] And also we need to have certified data sets, data assets, sorry, data assets. And now, before what was before a data asset in like one year ago, what was one data asset? What would you say is is a data asset?
[22:46] A table? Who thinks a table is a data asset? Yeah, raise your hand. Who thinks a graph dashboard, Tableau dashboard, Looker dashboard is a data asset?
[22:58] ML model? >>
[23:10] So, the first thing and and this is when I say like Ali got his stole the image from me. I have this prepared before his talk, right? My design to put the logo there.
[23:22] So, is let's go to DBX one or to Genie one. And this is our interface for every business user, right? So, data practitioners, they have the lake house is still, but data consumers just go here and use this, right?
[23:35] You can use this is still through the MCP, but use DBX one. Right? And the second one is the certification. So, we are going to certify everything. We started with Genies.
[23:48] And we have two certified Genies already. Custom Actually, three. Customer effort score, marketing campaign, and something very similar to the previous talk, which is the cost of one is certified, right? What if it's not we
[24:00] certified? >>
[24:13] So, when a Genie space lands in pro and everything that is in pro, right? And if it has a certified tag, you can trust it. It's good. Whatever it tells you, if it
[24:26] tells you like I don't know, 2 + 2 is 5, you trust it. It needs to be good. So, it needs to tell you 2 + 2 is 4, right? And again, this is a culture change also for the this AI consumer super fast,
[24:38] whatever give me like production means high quality. So, you can draft your things and and and pilot your things in pre. Production, no. Production needs to be good.
[24:50] Right? A Genie space or a Genie agent, right? Needs to be have a a scope. Like it's not the all the world. It's not just one day but it's like a scope of we
[25:03] are covering with this a domain part, a problem. Like I think today was better explained by other people. Like an agent has a purpose and it's for a a thing, right? So, it needs to have a well-defined scope. It needs to have
[25:15] instructions. It needs to have example SQLs. In our case, we were facing a problem of cost because of going to pick more data than needed and more columns than needed. So, it needs to be starting to be focused into this. And with
[25:28] genies, you can focus like don't cherry-pick all these big day tables. Just go to these ones which are optimized for cost. And how to join columns, how to how to join tables, etc. And it needs to have benchmarks. Like is
[25:41] this giving me always a good result or not? So, well, I'm because of the DBX one, no, sorry, genie one.
[25:53] This naming changes I I go with Because of genie one, it needs to be easy to find and discover what are my certified data assets, right? Right. So, this is also very easy because you have the tags and you have the certified
[26:05] tag which is a system tag with a check, blue check, amazing. But we need to put it there which we don't we didn't have it 1 month and a half ago. So, how? This is again technically
[26:18] easy, right? Conceptually to agree on this, not so easy. First, who creates a genie? A data practitioner. Well, who is a data practitioner? Well,
[26:30] that's a bigger question. But anyone that is able to work with data and knows the tables, knows how a query works, knows how to join two tables, so a data practitioner is not a title. It's more like a hat that you wear sometimes. So,
[26:43] data practitioner will create a genie space and it will do it where? >> Pre-production. >> In pre. Thank you for for the anonymous voice. So, yeah, in pre-production, right? Not in production, right? And
[26:55] this is a limitation that we have that everyone can do it in pro, but this is bad. We cannot computationally govern not to create it in pro, but we are finger-pointing like you did it in pro, you should do it in pre. Once it's in that part,
[27:09] some people will test the genius and like this is good enough. I am adding my questions and I am adding my expected results, and this is going fine. And then, you say, "Now it's ready to go to
[27:21] production. Now it has high quality." And you click it. How you click it, I will show in the same next slide, but it's basically with a GitHub action. Why? It's what we have. It works. It's simple enough. It's a little bit of copy-pasting and clicking buttons,
[27:34] but it works. And then, another very tricky point like who reviews this? How many people in the room are engineers or have engineering background?
[27:49] You know what is a pull request, a CI/CD kind of methodology, right? So, we want different pair of eyes checking that this is good. Because if I say that I'm good, I I can say I'm the best football player in the room. And according to me, I am the best
[28:01] football player in the room, but maybe someone disagrees, right? So, we need to do this. Then the product manager, the data product manager, which is a figure now in in our teams, will review and say like, "Yes, everything that is in pro is my
[28:13] responsibility to make sure that it's okay and to validate it." It's a little bit of a bottleneck, but we don't have so many genies yet. So, we will see when we scale what's if that's a problem.
[28:28] And once this happens, basically everyone can see the genie certified in pro, so happy happy. Right? And this is what we are doing. So, this is the demo. You can see a genie space,
[28:40] basically. You have instructions, you have joins, you have SQL expressions, right? And some of the functions, and you have the monitoring with the benchmarks and the benchmark that run.
[28:52] Before that, what we see in Genie spaces were like you have the tables, maybe some documentation, cross fingers to have it. Now have benchmarks, if it fails, the CICD process will fail
[29:06] and it will not go to pro. If you need to change it, this immediate next question that we got is how do I evolve it? How do I change this? You go back to pre, change something, clear the same process, and it updates. And for our users, this became an
[29:19] a standard interface. I go to this Genie. You can change whatever you want in the background. I will not notice. So, we can add more tables, create more context, improve the efficiency of the queries, no one will notice, right?
[29:36] This has been live for 2 weeks. So, that's why we have only three Genies. Maybe when next year we have 25 or or more. Uh and it's working really well. Now, what's next? We need to do the same for dashboards.
[29:48] We need to do the same for what tables we have it. We need to do the same
[30:02] is reviewed not by one one pair of eyes, but but one by many. And basically, this will elevate the quality of everything that is in production and our data operations.
[30:15] If this is high quality, high reliable, and it works, any decision that people can take by querying Genie will be better. That's the assumption. Don't ask me to qualify or to demonstrate. That's
[30:28] the assumption. Again, next year we can talk about benefits and value of it. Um some final words, this is the example when we were doing the migration and so on. I think the
[30:40] the the the picture is speaks about when we were coming back from from Amsterdam, Mariana and myself. Uh he was coding in the in the VIP room of the airport. He made a rake agent
[30:52] kind of a genie for his luck while we were waiting for the plane. I was chatting and drinking coffee. But it was kind of the a nice representation of how the future is looking a little bit
[31:04] like we will have some data practitioners that are working hard on creating genies, creating task force, creating tables, data products. And we will have some business users or
[31:16] lazy people like me chatting with the phone or with the laptop with the data like, "Hey, where are we losing money? Where are we gaining money? What can we do to improve our products, features, etc."
[31:28] So. And that's it. Any questions? Do we have time for questions? >>
[31:46] >> What does your branching strategy look like? Just go back to your slide on CD. With that, did you have like a monolith repo? >> So we have a Today we separated the repo of genies
[31:58] completely. >> Okay. >> Right? And basically because it's the same genie with with JSON files, it will overlap if you if you don't change the name of the space and the idea, I think it will basically overlap the first version of the same one. If
[32:11] you change the ID of the space or the name, it will look like a new one. So that this is not perfect. This is working. That's why it's new. We want to move it to the repo where that we have the DBT modeling and the rest.
[32:27] And then make it closer to the DBT modeling. So it's like if it's the genie for this part of the business, it goes into this folder and it's more grouped. And we have I had yesterday with the I forgot the name. Uh
[32:39] and I said it today the docs team to bring it closer also to the definition of a dashboard, to the definition of a table, and and again create like a bundle of everything around one domain or one specific
[32:51] feature. That's like the the intention, but again, next year they can tell you.
[33:24] >> We have some of them models are very simple. Have like 20 to 30 models migrated, right? Some models were like very simple. Once per year kind of almost open source models that you don't need
[33:36] to look into the data to to train or to retrain. And some other were like a little bit more complex, right? I cannot tell you how many features. It's It's true that they we could detect like maybe this feature is different or
[33:48] is is like the categories are slightly different. And then looking into it is like basically you change the category tree difference you train it 1 year ago to now. And it's impossible to reproduce, right?
[34:00] Even with the same data. And the same data was missing. Like what is the same data? That table, yeah, good luck. So it was easy to do it with the other table,
[34:12] but it was it was not exactly the same data. It was very similar. And again, what I was saying like we could not compare with the previous one objectively because
[34:24] all of this pre-info was not possible to to execute. But one version with the next one in the system that was easy to see the improvement and the changes and more recently in client side and we were
[34:36] having like different evaluations week after week of the same lifetime value models and you could go to the to the feature of what's different. The fun anecdote is like the same Argentinian guy here said like this is
[34:50] too complex for me to analyze. It's too many features, too many things. Genie, find the difference of these two evaluations of the same model. And it created a report, three pages, exported to PDF, sent to the data scientist, leave me alone. Ciao.
[35:03] And it was a he was super happy with this. He came jumping like, "Hey, look what I did with I didn't do the analysis but Genie did it for me and it was super fine. So
[35:20] And thank you and go for the Oracle Park. >>

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.