The database that answers before your users finish clicking

What it does Pinot is a distributed, column-oriented OLAP datastore built for user-facing analytics — dashboards and product features that query the database directly, not analysts poking around after hours. It ingests from streams (Kafka, Pulsar, Kinesis) and batch sources (S3, HDFS, Spark) into the same tables, and answers SQL with P90 latencies in the tens of milliseconds. It also does vector similarity search, which is why it’s now pitching itself at RAG pipelines and “agent-facing analytics” — AI assistants that need fresh numbers fast.

The interesting bit At LinkedIn it reportedly powers 50+ user-facing products, ingesting millions of events per second and serving 100k+ queries per second. The trick is a pluggable indexing zoo — StarTree, inverted, bloom, range, JSON, geospatial — layered on aggressive column compression, so aggregation queries skip most of the data entirely. The vector search bolt-on means one system can serve both your metrics and your embeddings.

Key highlights

• Millisecond-scale SQL over petabyte datasets, with upserts during real-time ingestion
• Hybrid tables: batch and streaming sources merged into one queryable view
• Built-in multitenancy and a Kubernetes Helm chart for clustered deployment
• Standard SQL via a built-in query console and REST API
• Proven at LinkedIn and Uber scale (Who Viewed Profile, UberEats Restaurant Manager)

Caveats

• This is a heavyweight distributed system — a cluster, not a library you drop into an app
• Building from source requires JDK 25+, and the README notes macOS ARM protobuf build quirks
• Best suited to immutable or upserted event data; not a general-purpose transactional store

Verdict If you’re building features where users (or agents) wait on the query, Pinot is one of the most battle-tested options around. If your analytics are internal and batch-y, it’s a lot of machinery for the job.