- See the Release notes archive for earlier-version release notes.
The rollout of version 4.33 is expected to be completed by the end of November 2026.
Firebolt might roll out releases in phases. New features and changes may not yet be available to all accounts.
Firebolt Release Notes - Version 4.33
Version 4.33 is primarily a stability release. It fixes several issues that could return incorrect results, crash an engine, or slow it down over long uptimes, and it hardens the engine against malformed input.New Features
Compatibility with the upcoming major release Over the past months, we’ve been working on some of the most ambitious projects in Firebolt history to make the engine even faster, double down on open formats, enable more data sources, and improve developer experience. These improvements will ship in an upcoming version. The 4.33 release contains various improvements that allow it to work with objects created in the upcoming major release. AsynchronousSELECT queries with the Null output format
Asynchronous queries can now run SELECT statements, including EXPLAIN, when the request sets output_format=Null. The query runs to completion and its result is discarded. This is particularly useful for warmup queries.
Performance Improvements
Late materialization forUNION queries
Queries with ORDER BY ... LIMIT over a UNION can now apply late materialization to each branch of the union, reading fewer columns for rows that do not make it into the result.
Faster planning of complex queries
Queries that join many tables on unique keys and group by several columns plan significantly faster; queries that repeat many identical subqueries also plan significantly faster.
Bug Fixes
Incorrect results withNULL key values
Fixed IS NULL filters on a nullable column of a composite primary index skipping matching rows, which incorrectly skipped NULLs. Fixed IS NULL filters on nullable partition columns skipping NULL partitions, with the same effect. Fixed window functions with PARTITION BY not grouping NULL keys together.
Engine stability
Fixed a crash during data shuffles between nodes under heavy load. Fixed a crash while spilling to disk under memory pressure. Fixed a crash when cleaning up spilled data after running out of memory. Fixed a crash in REGEXP_REPLACE when the strings in one batch of rows it processes add up to more than 2 GiB, which only happens with very large string values. Fixed a crash when running out of memory while handling metadata. Fixed a crash during planning of certain subqueries. Health checks of heavily loaded engines no longer time out, which previously caused unnecessary ngine restarts.
Long-running engines
Fixed auto-VACUUM failing with “Out of memory” on every run on engines with significant uptime and unusually high DML rates. Fixed query latency growing with engine uptime on tables whose partitions are frequently created and dropped.
Auto-stop and cancelled queries
A query that is being cancelled no longer keeps an engine from auto-stopping.
Tables in custom schemas and table statistics
ALTER TABLE ... OWNER TO now works for tables in user-defined schemas. Tables using STATISTICS on DATE, DECIMAL, or TIMESTAMP columns are now supported as expected. Concurrent TRUNCATE and INSERT on a table in a user-defined schema now correctly conflict with each other.
External data and locations
Fixed reading Apache Iceberg tables that are partitioned by a DECIMAL column. Listing files in Amazon S3 now retries additional classes of transient errors. When listing does fail, the error now more precisely reports the underlying cause.
Statements in a closed transaction fail immediately
A statement or COMMIT in a transaction that was already rolled back now fails right away with a user error stating that the transaction is closed, instead of retrying for up to 30 seconds or returning an internal error.
Security Fixes
Hardening against malformed input More robust handling of crafted compressed data (LZ4, T64, Brotli, and LZO- or ZSTD-compressed ORC) and malformed Parquet files.HLL_COUNT_MERGE now accepts only sketches built by HLL_COUNT_BUILD. Databricks AI functions now accept only plain serving endpoint names.