Latest / MBA Training Data / How Uber built its ML data supply chain: lessons from the Michelangelo feature store
Transcript
- 0:00Welcome to the Leaders Insights Podcast.
- 0:08Today's episode, how Uber built its ML data supply chain, lessons from the Michelangelo
- 0:13feature store.
- 0:14Uber built a feature store in 2017, and nine years later, everyone's still copying the
- 0:19homework.
- 0:20Why did that one decision age so well?
- 0:22Because they solved a boring problem before it became an expensive one.
- 0:26A feature, that's just an input a model learns from, like how many trips did this rider
- 0:31take last week?
- 0:33Was being rebuilt by five different teams, five different ways, all slightly wrong.
- 0:38The fraud teams last week, and the pricing teams last week, didn't agree.
- 0:43So it's a duplication problem.
- 0:45That doesn't sound like it deserves a whole platform.
- 0:47It doesn't, until you count.
- 0:50Uber had teams spending most of their model building time on data plumbing, not modeling.
- 0:55When two teams compute the same number two different ways, one of your models is silently
- 0:59wrong and nobody knows which.
- 1:02That's not waste, that's a liability.
- 1:04Give me the concrete failure.
- 1:06What actually breaks?
- 1:07The classic one is training serving skew.
- 1:10A fancy way of saying your model was trained on one version of a number, and shown a different
- 1:14version in production.
- 1:16You calculate average trips per week, in a big overnight batch job for training.
- 1:21Then some engineer re-implements it live for real-time predictions, and the two drift
- 1:26apart.
- 1:27Your model degrades, and every dashboard says it's fine.
- 1:31And the feature store fixes that?
- 1:32How exactly?
- 1:33One definition, computed once, served two ways.
- 1:37Michelangelo, that's Uber's internal machine learning platform, named after the painter,
- 1:43let's you write the feature logic in one place.
- 1:45The same pipeline fed both the training data and the live prediction.
- 1:50So the number a model learns on is provably the number it sees when it's deployed.
- 1:54That's the fairy tale.
- 1:55What did they get wrong first?
- 1:57They underestimated ownership.
- 1:59A shared feature is great until it breaks and nobody's on call.
- 2:04Early on you get this tragedy of the commons.
- 2:06Everyone uses the writer trip count feature.
- 2:09Nobody maintains it.
- 2:11And when the upstream data changes, 40 models degrade at once.
- 2:15Uber had to bolt on governance.
- 2:17Who owns this?
- 2:18Who gets paged?
- 2:19What's the freshness guarantee?
- 2:21So the technology was the easy part, and the org chart was the hard part.
- 2:25It always is.
- 2:27The feature store is 30% infrastructure and 70% a contract between teams.
- 2:33You're telling the pricing team, trust the fraud team's definition of an active writer.
- 2:38That's a political act dressed up as an engineering one.
- 2:41Let me push on the fashion angle.
- 2:43In 2026, half the industry is obsessed with large language models.
- 2:48Do feature stores even matter anymore?
- 2:50More than ever, and I'll tell you why the hype crowd is wrong.
- 2:53Language models handle unstructured text, but the decisions that move money.
- 2:58Should we approve this loan, price this ride, flag this transaction, still run on structured
- 3:03features about your customer.
- 3:06Tekton and Databricks, both vendors.
- 3:09So take the enthusiasm with salt.
- 3:11Report feature store demand growing right through the language model boom.
- 3:15The plumbing didn't go away because a chatbot showed up.
- 3:19You mentioned vendors.
- 3:20Should a company buy a feature store or build one like Uber did?
- 3:24Build it only if your Uber scale and feature reuse is your bottleneck.
- 3:28Almost nobody is.
- 3:30Uber built theirs because the market was empty in 2017.
- 3:34Today you'd buy Tekton or use what's inside Databricks and spend your energy on the
- 3:38definitions, not the pipes.
- 3:41Firing your own feature store in 2026 is like forging your own bolts to build a shed.
- 3:47So the durable lesson isn't build a feature store.
- 3:50No.
- 3:51The lesson is, define your important numbers once, assign an owner, and make everyone
- 3:56consume the same version.
- 3:59You can do that in a spreadsheet governance doc before you buy a single tool.
- 4:03The infrastructure just enforces a discipline you should already have.
- 4:07Give me the one thing a data leader does Monday morning.
- 4:10Pick your five most used features, the numbers that show up in the most models or reports.
- 4:16And ask two questions.
- 4:18Who owns each one?
- 4:19And is it computed the same way everywhere it's used?
- 4:22If you can't answer either, you've found your training serving SKU before it finds you.
- 4:27That's your five.
- 4:28The full CDO track, structured no fluff, is at MBA-training.com.