> Indeed.
> It might have already been discussed but other than using a new shmem
> hash for parallel vacuum, I wonder if we can allow workers to change
> the leader’s progress information. It would break the assumption that
> the backend status entry is modified by its own backend, though. But
> it might help for progress updates of other parallel operations too.
> This essentially does the same thing as what the current patch does
> but it doesn't require a new shmem hash.
I experimented with this idea, but it did not work. The idea would have been to create a pgstat_progress_update
functionthat takes the leader pid, however infrastructure does not exist to allow one backend to manipulate another
backendsbackend status array.
pgstat_fetch_stat_beentry returns a local copy only.
> Another idea I come up with is that the parallel vacuum leader checks
> PVIndStats.status and updates how many indexes are processed to its
> progress information. The leader can check it and update the progress
> information before and after index vacuuming. And possibly we can add
> a callback to the main loop of index AM's bulkdelete and vacuumcleanup
> so that the leader can periodically make it up-to-date.
> Regards,
The PVIndStats idea is also one I experimented with but it did not work. The reason being the backend checking the
progressneeds to do a shm_toc_lookup to access the data, but they are not prepared to do so.
I have not considered the callback in the index AM's bulkdelete and vacuumcleanup, but I can imagine this is not
possiblesince a leader could be busy vacuuming rather than updating counters, but I may be misunderstanding the
suggestion.
--
Sami Imseih
Amazon Web Services