Tecton Review
VisitFeature store platform for operational machine learning
Tecton is an enterprise feature store platform for building and serving ML features in production.
AI Panel Score
7.8/10
6 AI reviews
Reviewed
Feature store platform for operational machine learning
Tecton is an enterprise feature store platform for building and serving ML features in production.
AI Panel Score
6 AI reviews
Reviewed
Tecton is a managed feature platform designed to operationalize machine learning by centralizing the infrastructure needed to build, store, and serve ML features. It addresses one of the core challenges in production ML: ensuring that the features used during model training are identical to those served at inference time, eliminating training-serving skew.
The platform supports three types of feature computation: batch features derived from historical data, streaming features updated in near real-time from event streams, and on-demand features computed at request time. Teams can define feature logic in Python using Tecton's SDK, and the platform handles orchestration, storage, and serving automatically.
Tecton is primarily aimed at data scientists, ML engineers, and data platform teams at mid-to-large enterprises building customer-facing or operational ML applications. Common use cases include fraud detection, personalization, recommendation systems, and risk scoring where low-latency, accurate feature retrieval is critical.
A central component of Tecton is its feature registry, which acts as a shared catalog where teams can discover, reuse, and monitor features across projects. This reduces duplicated engineering effort and promotes consistency across an organization's ML initiatives. Features can be monitored for drift and freshness through built-in observability tooling.
Tecton integrates with major data warehouses and lakehouse platforms including Snowflake, Databricks, and AWS, and operates within existing cloud infrastructure rather than requiring data migration. It competes in the MLOps and feature store space alongside tools like Feast, Hopsworks, and cloud-native offerings from AWS SageMaker and Google Vertex AI.
Sub-10ms latency online store backed by DynamoDB or Redis for real-time inference serving at production scale.
Single feature definition produces both training data and online serving values, eliminating the training-serving skew that breaks production ML accuracy.
Catalog of registered features with lineage, owners, and usage stats so teams can find and reuse features across model projects.
Compute features from historical data using SQL and Python on Spark or Snowflake, with backfill support for large historical training sets.
Centralized repository for ML features with versioning, lineage, and reuse across teams — eliminates feature duplication across model projects.
Compute features at request time from raw inputs or external API calls, supported in the same Python framework as batch and streaming definitions.
Update features in near real-time from event streams via Kafka, Kinesis, or Pub/Sub for low-latency model serving.
Define features as code in Python with declarative transformations, version-controlled in Git and reviewable like application code.
Native integration with major cloud platforms; deploys into customer VPC with cloud-native data sources (Snowflake, BigQuery, Redshift, Databricks).
First-class integration with Snowflake and Databricks as compute layers, allowing teams to use existing data warehouse infrastructure.
Custom pricing based on feature-platform usage (online and offline feature serving volume), team size, and deployment options. Contact Tecton sales for a quote.
Databricks closed the Tecton acquisition on August 22, 2025 — vendor-survival question, settled.
“Tecton is the 2019-founded feature store from ex-Uber Michelangelo engineers, acquired by Databricks in August 2025 after $160M raised. The buying lens now is fit with your data stack, not whether the vendor lasts the contract.”
Databricks closed the Tecton acquisition on August 22, 2025. The founders built Uber's Michelangelo before they built this. The vendor-survival question is gone — your feature store is now a Databricks line item.
On-Demand Features and the sub-10ms Online Feature Serving on DynamoDB are what Snowflake-shop teams were paying $160M-funded Tecton to handle. Feast is open-source and free; Hopsworks competes on the enterprise tier. The reuse-the-feature-registry pitch is real, but only if your peers are already on Databricks Lakehouse.
However, the catch is the same one every acquisition brings: roadmap pressure toward AgentBricks. If you run on AWS SageMaker or Vertex AI, watch the changelog for parity. Pilot one fraud-detection model for 90 days before signing the mid-five to six-figure annual contract.
Peer ML teams in fraud, personalization, and risk scoring already run managed feature stores.
Databricks-backed feature store is a defensible board-level choice in MLOps.
Two-to-four week platform install plus per-feature migration delays first-model payback.
Real-time feature serving advances production ML, but value depends on existing data warehouse alignment.
Databricks acquisition closed August 22, 2025; founders built Uber Michelangelo, $160M raised across three rounds.
Databricks shops running real-time ML who need a managed feature store.
Teams on AWS SageMaker or Vertex AI without a Databricks footprint.
Databricks closed Tecton in August 2025, folding Uber-Michelangelo lineage into the lakehouse it already runs on.
“Databricks announced the Tecton acquisition on August 22, 2025 with terms undisclosed, three years after Tecton's $100M Series C set a roughly $900M valuation. The Feature Repository ships sub-10ms online serving from declarative Python, but the three-year roadmap now lives on Mosaic AI's priorities rather than Tecton's.”
Databricks closed the Tecton acquisition on August 22, 2025, terms undisclosed — a $900M-valuation feature platform folded into the lakehouse it already runs on. For a head of ML platform, the question shifts from 'pick Tecton or Feast' to 'how long does Tecton stay first-class versus a feature inside Mosaic AI'.
The craft holds up. Mike Del Balso and Kevin Stumpf shipped Uber Michelangelo before founding Tecton in 2018, and that lineage shows in the Feature Repository — declarative Python definitions producing both training data and sub-10ms online serving from a single spec. Point-in-time correctness is enforced by construction, not by convention.
However, the three-year bet now depends on Databricks' priorities, not Tecton's roadmap. Snowflake shops lose a neutral vendor; Feast stays the open-source hedge for substrate independence. For a Databricks-committed team, this is a buy — for everyone else, the integration surface narrowed overnight.
Tecton was the managed feature-store category leader before Databricks consolidated it.
Sub-10ms online serving from a single Python spec matches how production ML teams actually work.
Native Snowflake, Databricks, BigQuery, Kafka, and Kinesis coverage spans the standard data plane.
The August 2025 Databricks acquisition narrows independence and tilts the three-year roadmap.
Michelangelo lineage and point-in-time correctness are real platform craft, not surface features.
ML platform teams already running on Databricks who need production feature serving.
Snowflake-first teams who want a vendor-neutral feature platform.
Databricks closed the acquisition in August 2025 — pricing was contact-sales then, still is now.
“Databricks acquired Tecton on August 26, 2025 at an undisclosed price, against a prior $900M mark and $160M total raised. No published tiers, no per-feature anchor — procurement walks in blind.”
Tecton's strategic partner became its parent. Databricks closed the acquisition August 26, 2025 at an undisclosed price, against a $900M prior mark and $160M total raised since 2019. Kleiner Perkins led the 2022 Series C. Snowflake Ventures was on the cap table — that line just got awkward.
Pricing was contact-sales before the deal and remains so. Online Feature Serving at sub-10ms p99, 99.99% uptime SLA. Category norm for managed feature platforms lands $50K-$150K annually for a mid-size ML team. Feast is open-source and free if you can staff the platform engineering.
Hopsworks publishes per-tier pricing — rare in this category. AWS SageMaker Feature Store charges on storage and read units, predictable but capped at AWS. The catch is the new owner. If you're on Snowflake or BigQuery, you're now buying real-time features from your warehouse vendor's direct competitor.
Custom MSA, no published overage rate, vendor onboarding requires VPC deployment and 2-4 week platform install.
Enterprise-only motion means standard auto-renewal and term lengths; post-Databricks renewal terms are an open question.
Single Enterprise tier, contact-sales only, no per-feature or per-seat anchor on the pricing page.
Training-serving skew elimination and sub-10ms Online Feature Serving give measurable production-ML value versus rolling your own.
Category norm $50K-$150K annually but real cost depends on online store choice and Weave-equivalent ingestion not disclosed.
ML platform teams who serve real-time features at sub-10ms latency in production.
Teams who need published pricing before procurement engagement.
Tecton SDK and sub-10ms online store run the inference path, but contact-sales pricing kills the side experiment.
“The Python SDK plus the sub-10ms online store handle the daily ML serving loop in a way Feast self-hosted won't match without a platform team. But the contact-sales-only pricing means no Snowflake user spins up a Tecton workspace on a Tuesday afternoon.”
The Python SDK defines a feature once and the same code path materializes training data and serves online — that's the registry guarantee Feast offline-only deployments don't make on their own. Batch lands on Spark or Snowflake, streaming hooks Kafka or Kinesis, on-demand runs at request time. Three pipeline types, one decorator vocabulary.
Online Feature Serving claims p99 under 10 milliseconds for in-region lookups against DynamoDB or Redis. For fraud scoring that latency budget is the difference between approving a transaction and timing out. Point-in-time joins prevent the training data leakage that breaks production accuracy six months in.
But there's no list price and no self-service tier — every evaluation routes through Tecton sales, and enterprise contracts land mid-five to six figures annually. SageMaker Feature Store ships in the AWS console for teams already there. Databricks acquired Tecton August 2025; the standalone roadmap depends on Agent Bricks integration now.
Single Python definition for batch, streaming, and on-demand features holds up past the demo; ops gravity is real but predictable.
SDK reference reads like it was written by the engineers who use it; the contact-sales pricing page is the friction.
Initial install runs 2-4 weeks and every evaluation routes through sales; no Tuesday-afternoon spin-up.
Point-in-time joins, on-demand transforms, and DynamoDB/Redis online store choice give serious teams real knobs to turn.
Native hooks into Snowflake, Databricks, Kafka, and Kinesis mean no data migration — features compute where the data already lives.
Data teams who run real-time inference on Snowflake or Databricks.
Solo ML engineers who need a free tier for a side project.
Tecton is Databricks now — the August 2025 acquisition closed the loop on the Uber Michelangelo founders' bet
“Online Feature Serving hits sub-10ms p99 on DynamoDB or Redis, the Feature Registry kills duplicated fraud-signal work across teams, and the Python SDK keeps definitions in Git. The catch is contact-sales pricing and a Databricks acquisition in August 2025 that puts the standalone product's future on the clock.”
The website redirects to Databricks. That's the first thing. Tecton, the feature platform built by the three Uber Michelangelo founders in 2018, was acquired by Databricks in August 2025 after raising $160M across three rounds. Pricing is still contact-sales — on day one that tells you who this is for.
The Online Feature Serving store hits sub-10ms p99 latency on DynamoDB or Redis. The Feature Registry earns its keep month three — a catalog where teams find and reuse features instead of rebuilding the same fraud signal in four notebooks.
But Feast covers the basics open-source for free, and now that Tecton lives inside Databricks, the next two years will say whether the standalone product survives or becomes a Databricks SKU. The 2-4 week install isn't homework — it's a commitment.
Solid Python SDK and docs, but the .ai homepage already redirects to Databricks mid-rebrand.
Declarative Python features feel familiar fast; the platform complexity behind them takes months to master.
Backend ML infrastructure — mobile is not the relevant surface, scored neutral.
Contact-sales gate plus a 2-4 week platform install means the first 10 minutes is a sales call.
Sub-10ms p99 online serving and point-in-time correctness by construction eliminate training-serving skew.
ML platform teams who need real-time feature serving at production scale.
Solo data scientists who can run a notebook on a laptop.
Databricks bought Tecton's assets in August 2025 — the strategic investor became the owner.
“Databricks acquired Tecton's assets on August 26, 2025, folding the Uber Michelangelo team's feature platform into Agent Bricks. The standalone vendor question is moot — the open question is how long the standalone product survives the integration.”
Asset acquisition, not company acquisition. Fenwick's filing language matters — Databricks bought Tecton's assets August 26, 2025, after years as a strategic investor and Series C participant. The investor became the buyer. That pattern usually ends one way.
The bones are legitimate. Mike Del Balso, Kevin Stumpf, and Jeremy Hermann built Uber's Michelangelo before launching Tecton in 2020. Series C was $100M led by Kleiner Perkins, $160M total raised. Sub-10ms p99 on the Online Feature Store backed by DynamoDB or Redis is real engineering, not slide-deck latency.
But the roadmap belongs to Agent Bricks now. Feast is the open-source escape hatch with Tecton-compatible APIs — that's the exit story, and it's better than most. Contact-sales pricing was always opaque; under Databricks it'll get bundled. Standalone-SKU survival is the watch item, not product quality.
Michelangelo pedigree plus managed online serving plus point-in-time correctness is a real gap vs Feast's DIY ops.
Feast offers Tecton-compatible APIs and Python definitions travel, but Online Feature Store data and registry lineage don't.
As a standalone SKU the runway is unclear; under Databricks the technology lives on inside Agent Bricks.
Sub-10ms latency and training-serving consistency claims match the docs; contact-sales pricing is the one opacity tell.
Asset-acquired by a strategic investor in 2025 — the classic absorption pattern, but the tech and team survive intact.
ML teams who serve real-time features at production scale on Databricks.
Solo data scientists who need a free open-source feature store.
Common questions answered by our AI research team
Tecton supports batch, streaming, and real-time feature pipelines.
Yes, Tecton ensures consistent feature computation between training and inference, eliminating training-serving skew.
Tecton uses contact-sales pricing rather than published tiers. Pricing is based on number of features served, request volume, infrastructure footprint (online store size, compute), and support tier. Enterprise contracts typically land in the mid-five to six figure annual range, with the open-source Tecton SDK available for evaluation before commitment.
Native integrations include Snowflake, BigQuery, Redshift, Databricks, S3, Kafka, Kinesis, and Pub/Sub. Deployment runs in the customer VPC on AWS, GCP, or Azure. Initial setup typically takes 2-4 weeks for the platform install plus team onboarding; migrating existing feature pipelines into Tecton is a per-feature engineering project.
Tecton supports major cloud environments, though specific cloud provider names are not detailed in the available content.
Yes, Tecton connects data pipelines to ML models, bridging data engineering and ML workflows.
Each feature is defined once in Python; Tecton uses the same definition to materialize training data from historical sources and to serve inference requests in real time. The training-serving skew problem (different code paths producing different feature values) is solved by construction. Point-in-time correctness is enforced for time-series features, preventing data leakage.
The online store typically serves features at p99 latency under 10 milliseconds for in-region requests. Latency depends on online store choice (DynamoDB, Redis), feature complexity (single-row lookup vs aggregation), and request size. On-demand features (computed at request time) add the cost of the transformation logic itself.
Tecton is a managed feature platform, meaning it is provided as a managed service rather than self-hosted.
Feast is open-source and free; Tecton offers Feast-compatible APIs as a starting point but adds managed infrastructure, online serving, point-in-time correctness, and enterprise support. Compared to in-house builds, Tecton trades license cost for engineering time — most teams underestimate the 12-24 months of platform engineering required to match Tecton's feature parity. The right choice depends on team size and ML maturity.
Company
TectonFounded
2019Pricing
Contact for pricingFree Trial
Available
Tecton is a San Francisco-based feature store and AI data platform company that helps teams build real-time ML and generative AI applications on production data.
