DuckDB announced 1.5.6 on September 28, 2026. It is the sixth patch release in the 1.5 (Variegata) line.
The announcement’s Correctness section lists 13 fixes. They include LIMIT pushdown through a volatile projection and wrong results when UNION ALL arms share a join subtree.
It also fixes failed checkpoint marker recovery and renaming over the main WAL handle without closing it during WAL recovery. The WAL is a log file that records changes first.
So a query that runs is not enough to pass a migration. These are announced fixes, not bugs reproduced for this article.
Two fixes that change the meaning of returned and stored data
Two cases in the release notes are worth keeping as regression inputs: silent truncation of very long integer literals into HUGEINT, and reading and writing TIME_NS values in Parquet.
HUGEINT is a 128-bit integer type, and TIME_NS is a time value in nanoseconds.
Both concern the meaning of data returned or written to files, not execution speed. A query that finishes without error can still return a different value.
Caveats: the 2.0.0 date is tentative
The release calendar places 2.0.0 on October 21, 2026 and 2.0.1 on November 18. The same page says these dates are tentative and may be pushed back for stability and quality.
The next major version is a separate path. It is neither a release guarantee nor a schedule that substitutes for this 1.5 maintenance decision.
An upgrade comparison that starts with values, nulls and row counts
Before deploying, the first step is running a small representative dataset through the current version and 1.5.6 and comparing values, null handling and row counts.
Unexpected differences become regression inputs. If performance matters, runtime is measured after that.
This is a suggested check, not a performance benchmark for this article. Evaluating 2.0 development builds is handled separately.
In a patch that carries correctness fixes, the first thing to confirm is not a faster query but whether the same input returns the same value.