Version 6.4.7

Released on 2026-11-04.

Note

If you are upgrading a cluster, you must be running CrateDB 5.0.0 or higher before you upgrade to 6.4.7.

We recommend that you upgrade to the latest 6.3 release before moving to 6.4.7.

A rolling upgrade from >= 6.3.0 to 6.4.7 is supported. Before upgrading, you should back up your data.

Warning

Tables that were created before CrateDB 5.x will not function with 6.x and must be recreated before moving to 6.x.x.

You can recreate tables using COPY TO and COPY FROM or by inserting the data into a new table.

Table of contents

See the Version 6.4.0 release notes for a full list of changes in the 6.4 series.

Fixes

  • Fixed some issues in the PostgreSQL wire protocol implementation that:

    • Disallowed pipelining select queries after write queries

    • Caused failures if pipelining statements using new objects immediately after DDL statements created those objects, if there was no intermediate sync message between those statements.

    • Could cause clients to appear stuck if a statement within a pipeline had an execution error.

  • Fixed an issue that caused CrateDB to send the wrong transaction state within a ReadyForQuery message to clients using the PostgreSQL Wire Protocol in response to queries sent in the simple query mode after a BEGIN. This could cause issues with clients like asyncpg. Causing them to send an additional ROLLBACK which then failed,

  • Fixed an issue that caused queries on partitioned tables to exclude matching partitions if it had a partitioned column derived via a generated expression, and if that generated expression was included within the query via range filter.

  • Added missing support for the UUID type to the hyperloglog_distinct aggregation.

  • Fixed an issue where a logical replication subscription’s active connection could be unexpectedly closed when an earlier connection attempt for a subscription with the same name failed afterwards. As a consequence, both data replication and metadata tracking could stop for the affected subscription.

  • Fixed an issue where logical replication metadata tracking could stop after a subscriber restart if the connection to the publisher was not ready. As a consequence, schema changes and new partitions could fail to replicate, while data replication on existing partitions could continue.

  • Fixed an issue where data replication on existing partitions could stall after a publisher restart when using pg_tunnel mode for logical replication.

  • Fixed an issue where a logical replication subscription could enter a failed state and stop replicating after a publisher restart if the subscriber sent requests before the publisher’s transport service was ready to handle them.

  • Fixed an issue that could leave logical replication subscriptions stuck in the restoring state after the initial restore had completed, preventing schema updates and the creation of new partitions on the subscriber.

  • Fixed an issue which caused the optimizer_eliminate_cross_join rule to drop non-equality filters (e.g., a greater-than condition) that are combined with an equality filter over the same two tables. The issue also caused NOT conditions to be negated. This led to wrong results. In this example:

    SELECT t1.x, t2.y, t3.z
FROM t1, t2, t3
WHERE t1.x = t2.y
  AND t1.i > t2.i
  AND t2.y = t3.z;

    The query would behave like:

    SELECT t1.x, t2.y, t3.z
FROM t1, t2, t3
WHERE t1.x = t2.y
  AND t2.y = t3.z;

    A NOT condition like t1.x = t2.y AND NOT (t1.a = t2.b) was turned into t1.x = t2.y AND (t1.a = t2.b).

  • Disabled the optimizer rule optimizer_reorder_nested_loop_join by default to fix a performance regression of cross and outer joins that could happen if a query was run on multiple nodes and table statistics were available.

  • Fixed an issue preventing creation of a publication if it included a table that is included into another publication.

  • Fixed an issue where the optimizer_eliminate_cross_join rule wouldn’t eliminate a cross join even though there’s a matching equality condition in the WHERE clause that can transform it into an inner join. Here’s an example:

    SELECT t1.x, t2.y, t3.z, t4.id
FROM t1
CROSS JOIN t2
CROSS JOIN t3
JOIN t4 ON t2.y = t4.id AND t3.z = t4.id
WHERE t1.x = t3.z
  AND t1.i > 0;

    Previously, the resulting plan would contain a cross join. With the fix, the plan contains inner joins only and hence the query performs better.

  • Fixed an issue that caused queries with filtered nested CROSS JOIN``s to return incorrect results when ``optimizer_eliminate_cross_join was set to false and a filter above the join tree referenced only one side of the current cross join, e.g.:

    SET optimizer_eliminate_cross_join = false;

SELECT t1.x, t2.y, t3.z
FROM t1
CROSS JOIN t2
CROSS JOIN t3
WHERE t1.x = t2.y
  AND t1.x = t3.z
  AND t2.y = t3.z;

  • Fixed an issue where double quotes and backslashes in array elements (e.g. TEXT[]) weren’t escaped when sent over the PostgreSQL wire protocol in text format. This caused clients to parse the elements incorrectly.

  • Fixed an issue that caused inner join queries to fail when a join condition referenced multiple tables that are all on one side of the join, e.g.:

    SELECT t1.x, t2.y, t3.z
FROM t1
JOIN t2 ON t1.a = t2.b
JOIN t3 ON t1.x = t2.y AND t1.x = t3.z;

    A query like this used to fail with Index 1 out of bounds for length 1.

  • Fixed an issue causing CREATE SUBSCRIPTION to get stuck if pg_tunnel mode was used together with sslmode set to prefer or required and subscriber cluster didn’t trust publisher cluster’s certificate.

  • Fixed an issue where dropping a table on the publisher cluster left it in publications that explicitly listed it. Dropping a table now also removes it from those publications. References to non-existent tables in publications are also removed when a node restarts.

  • Fixed an issue causing COPY TO to ignore the generated file name and write to the path preceding the ? or # character, if the output URI or the table name contained them, thus overwriting existing files.

  • Fixed a regression, introduced in Version 6.3.0, allowing creation of functions with the same signature with CREATE FUNCTION instead of throwing an error.

  • Fixed a regression, introduced in Version 6.3.0, which caused CREATE OR REPLACE FUNCTION to add a duplicate function with the same signature instead of replacing the existing one.

  • Fixed a regression, introduced in Version 6.3.0, which caused RESTORE SNAPSHOT to not overwrite existing only on cluster UDF functions (as documented) and instead merge snapshot and cluster metadata.

  • Fixed an issue that caused Logical Replication to stop when primary shards on the subscriber cluster were relocated to another node or when a replica shard was promoted to primary.

  • Subscribing user used in CREATE SUBSCRIPTION with pg_tunnel mode now requires to have AL privilege to avoid a security issue when authenticated user can can switch to CrateDB’s unauthenticated internal transport protocol.