RE: Parallel Apply

On Wednesday, July 15, 2026 1:03 PM Kuroda, Hayato <kuroda(dot)hayato(at)fujitsu(dot)com> wrote:
> I found cfbot cannot accept the v21 patch set, please see the updated
> version.

Here is the updated patch set.

In this version, I've re-split the patches to make each one implement a more
specific, self-contained functionality that is a necessary part of supporting
parallel apply. I also rewrote the commit messages for each patch and moved the
overall design summary to the 0001 patch, so reviewers can understand the
architecture by reading just the first commit message.

In addition, I did quite a bit of refactoring and simplification to make the
code easier to read, and added more comments where needed.

For dependency tracking, I replaced the old change-count-based cleanup with a
change-size-based memory limit control. If the tuples saved for dependency
tracking exceed the limit, parallel apply is suspended and subsequent
transactions are applied by the leader until the preceding transaction commits
and the cached dependency changes are cleaned up.

In 0010, fixed a bug where the DML operation was not marked as parallel unsafe.

For local (subscriber-side) unique key and foreign key dependency tracking,
Kuroda-san and I worked on supporting partitioned tables. For partitioned
tables, the leader now collects unique indexes from the leaf partitions rather
than the parent table. Patches 0011 and 0012 implement this support.

Best Regards,
Zhijie Hou