Lakehouse RT vs Pinot

Summary

• Lakehouse Real-Time (Lakehouse//RT) is a Beta SQL warehouse type on Databricks serverless compute for low-latency, high-concurrency read queries on Unity Catalog managed Delta Lake and Apache Iceberg tables.
• Apache Pinot is an open-source (Apache 2.0) real-time distributed OLAP datastore, originally developed at LinkedIn, aimed at user-facing analytics, internal dashboards, anomaly detection and ad hoc exploration.
• Pinot ingests from streams such as Kafka or from batch sources and stores the data as columnar segments with per-table indexes; Lakehouse//RT reads the tables that Databricks pipelines already write.
• Pinot has its own access control, with table-level permissions on the controller and broker; Lakehouse//RT relies on Unity Catalog and doesn't support row-level security or column masking in the Beta.
• The choice is between serving from tables you already govern in Unity Catalog and running a separate Pinot cluster with its own ingestion and indexes.

Lakehouse RT vs Pinot

Both answer low-latency analytical queries on fresh data, but Pinot is a separate datastore and Lakehouse//RT is a warehouse type inside Databricks. Pinot ingests data into its own segments and indexes, then serves queries through its brokers. Lakehouse//RT, in Beta, reads Delta Lake and Iceberg tables in place under Unity Catalog. Choose based on whether you want a second system tuned table by table, or serving on the governed tables you already have.

What is Apache Pinot?

Apache Pinot is "a real-time distributed online analytical processing (OLAP) datastore". It ingests and queries data from streaming sources such as Kafka and Kinesis, and from batch sources such as HDFS and cloud object storage.
Pinot splits each table into segments. A segment packs its rows in columnar form, with dictionaries and indexes, and tables come in real-time (streaming), offline (batch) and hybrid types. Indexes are configured per table and include inverted, sorted, range, Bloom filter, text, JSON, geospatial, star-tree and vector indexes. Queries run through the broker in SQL, parsed by Apache Calcite, and features such as joins and subqueries use the multi-stage engine. A cluster has controller, broker, server and minion components, and Basic Auth access control can limit each user to specific tables and permissions.

What is Lakehouse//RT?

Lakehouse Real-Time (Lakehouse//RT) is "a serverless compute built for low-latency, high-concurrency use cases", such as serving analytical data to custom applications, running operational analytics or powering BI dashboards. It's a SQL warehouse type: a workspace admin enables the Beta under Previews, in supported regions, and you create one under Compute > SQL Warehouses by selecting Real-Time.
It queries Unity Catalog managed Delta Lake and Iceberg tables, materialized views, streaming tables and metric views, and it runs ANSI read (SELECT) queries only. Applications connect through the Statement Execution API or Databricks drivers such as JDBC. The docs say its performance characteristics and supported features will change before general availability.

How do Lakehouse//RT and Pinot compare?

Dimension	Databricks Lakehouse//RT	Apache Pinot
What it is	Real-Time SQL warehouse type on Databricks serverless compute (Beta)	Open-source distributed OLAP datastore (Apache 2.0)
Data storage	Unity Catalog managed Delta Lake and Iceberg tables in cloud storage	Pinot segments in columnar form, with dictionaries and indexes
Getting data in	Reads tables written by Lakeflow Pipelines, Structured Streaming or other Databricks compute	Real-time tables consume streams such as Kafka; offline tables load batch data
Indexing and layout	Docs recommend managed tables with predictive optimization and liquid clustering	Per-table indexes such as inverted, star-tree, Bloom filter, range and JSON
Updates	Read-only; no INSERT, UPDATE, DELETE or MERGE	Upsert and dedup options during ingestion
Query interface	ANSI read-only SQL via the Statement Execution API and Databricks drivers	SQL through the broker, with single-stage and multi-stage engines
Access control	Unity Catalog; no ABAC, row-level security or column masking in the Beta	Basic Auth with per-table permissions on the controller and broker
Deployment	Managed by Databricks, enabled per workspace in supported regions	You run the cluster: locally, in Docker or on Kubernetes with Helm

How does Databricks approach real-time analytics?

Databricks launched Lakehouse//RT at Data + AI Summit in June 2026 to run real-time analytics directly on governed Delta Lake and Iceberg tables, without copying data into a separate serving system. Lakeflow Pipelines ingest from sources such as Apache Kafka into streaming tables and materialized views, and Lakehouse//RT serves reads on them.
The docs recommend validating queries on a serverless SQL warehouse first, using managed tables with predictive optimization and liquid clustering, keeping queries selective, and pre-aggregating complex joins with materialized views. When a query uses an unsupported feature, Lakehouse//RT returns an error naming that feature, and you run the query on a serverless SQL warehouse instead.

FAQs

Does Apache Pinot keep its own copy of the data?

Yes. Loading data into Pinot means generating segments from raw data and pushing them to the cluster, in batch or streaming mode, and Pinot servers store those segments. Lakehouse//RT reads managed Delta Lake and Iceberg tables in place.

What indexes does Pinot offer?

Inverted indexes for equality and IN filters, star-tree indexes for repeatable aggregation and group-by workloads, and Bloom filters for segment pruning on selective lookups, plus range, text, JSON, geospatial, timestamp, FST and vector indexes. You choose them per table and column.

Which SQL dialect does Pinot use?

Pinot SQL uses the Apache Calcite parser with the MYSQL_ANSI dialect. You pick the single-stage engine or the multi-stage engine, which adds features such as joins and subqueries.

Can Lakehouse//RT write data?

No. It runs read (SELECT) queries only. Writes, DDL, GRANT and REVOKE, and Delta Lake maintenance such as OPTIMIZE and VACUUM aren't supported, so pipelines and other warehouses handle those.