Trino

Trino is an open-source, distributed SQL query engine for fast, interactive analytics over data where it already lives: Parquet and ORC files on S3 or HDFS, Iceberg, Delta Lake and Hive tables, and also PostgreSQL, MySQL, Kafka, Elasticsearch and dozens of other systems through connectors. It has no storage of its own. A coordinator plans each query and a fleet of workers executes it in parallel, pipelining data through memory, so queries over terabytes return in seconds. Trino was called PrestoSQL until December 2020; it is the fork created by Presto’s original authors.

Commander Bun Bun, the Trino mascot

How Trino runs a query

The general flow (coordinator, splits, parallel scans, partial aggregation, exchange, final aggregation) is animated on the Query Engine page. Trino-specific pieces:

  • Coordinator: accepts queries over HTTP (CLI, JDBC, BI tools), parses and optimizes them with a cost-based optimizer, breaks the plan into stages, and schedules tasks on workers. It also returns results to the client.
  • Workers: run tasks, read splits through connectors, and exchange data with each other over HTTP. Data flows between operators in columnar pages held in memory.
  • Connectors and catalogs: a connector is a plugin that knows how to talk to one kind of system. A catalog is a configured instance of a connector (etc/catalog/lake.properties, etc/catalog/crm.properties). Tables are addressed as catalog.schema.table.

Federation: one query, several systems

-- Join lake data (Iceberg on S3) with live CRM data (PostgreSQL)
SELECT c.segment, SUM(o.amount) AS revenue
FROM lake.sales.orders AS o
JOIN crm.public.customers AS c ON o.customer_id = c.id
WHERE o.dt >= DATE '2026-11-01'
GROUP BY c.segment;

Trino pushes what it can into each source (filters, projections, sometimes aggregates and limits into PostgreSQL) and does the join itself. Federation is great for exploration and small dimension tables; for heavy recurring joins, copy the data into the lake.

Fault-tolerant execution

By default Trino is pipelined: if a worker dies, the whole query fails and must be rerun. That is fine for 10-second dashboard queries, painful for 3-hour ETL. Since 2022, fault-tolerant execution (retry-policy=TASK) writes intermediate exchange data to a spooling store such as S3, so only failed tasks retry. It makes Trino usable for batch ETL at the cost of extra I/O.

History: Presto, PrestoSQL, Trino

Year Event
2012 Martin Traverso, Dain Sundstrom, David Phillips and Eric Hwang start Presto at Facebook to replace slow Hive queries.
2013 Presto is open-sourced.
2019 (Jan) The founders leave Facebook and continue development as PrestoSQL, under the Presto Software Foundation.
2019 (Sep) Facebook donates the original repository (PrestoDB) and the Presto trademark to the Linux Foundation’s Presto Foundation.
2020 (Dec) PrestoSQL is renamed Trino because of the trademark. The foundation becomes the Trino Software Foundation.
2022 Amazon Athena engine version 3 moves to Trino. Fault-tolerant execution lands.

Since then the two projects have diverged; see Presto for how they compare today.

Where you meet Trino

  • Self-hosted on Kubernetes (official Helm chart) or VMs.
  • Starburst: the main commercial company behind Trino (Starburst Enterprise and Starburst Galaxy).
  • Amazon Athena (serverless, engine v3 is Trino-based) and Amazon EMR.
  • Inside many in-house lakehouse platforms as the interactive SQL layer, often next to Spark for ETL.

Common mistakes

  • Undersized memory. Joins and aggregations happen in memory; set query.max-memory-per-node deliberately and use fault-tolerant execution for big queries (the older spill-to-disk feature is documented as legacy).
  • Letting the coordinator also act as a worker in production clusters. It slows planning for everyone.
  • Over-relying on federation. Pulling a 500 GB PostgreSQL table through JDBC on every query will hurt both Trino and the database.
  • Ignoring table statistics. Without ANALYZE (or Iceberg/Delta stats), the cost-based optimizer guesses join order and can pick a terrible plan.

Learning Resources