+ +/*
+ + * vacuum_worker_init --- initialize this module's shared memory hash
+ + * to track the progress of a vacuum worker
+ + */
+ +void
+ +vacuum_worker_init(void)
+ +{
+ + HASHCTL info;
+ + long max_table_size = GetMaxBackends();
+ +
+ + VacuumWorkerProgressHash = NULL;
+ +
+ + info.keysize = sizeof(pid_t);
+ + info.entrysize = sizeof(VacProgressEntry);
+ +
+ + VacuumWorkerProgressHash = ShmemInitHash("Vacuum Progress Hash",
+ +
+ max_table_size,
+ +
+ max_table_size,
+ +
+ &info,
+ +
+ HASH_ELEM | HASH_BLOBS);
+ +}
+ It seems to me that creating a shmem hash with max_table_size entries
+ for parallel vacuum process tracking is too much. IIRC an old patch
+ had parallel vacuum workers advertise its progress and changed the
+ pg_stat_progress_vacuum view so that it aggregates the results
+ including workers' stats. I think it’s better than the current one.
+ Why did you change that?
+ Regards,
I was trying to avoid a shared memory to track completed indexes, but
aggregating stats does not work with parallel vacuums. This is because a
parallel worker will exit before the vacuum completes causing the aggregated
total to be wrong.
For example
Leader_pid advertises it completed 2 indexes
Parallel worker advertises it completed 2 indexes
When aggregating we see 4 indexes completed.
After the parallel worker exits, the aggregation will show only 2 indexes
completed.
--
Sami Imseih
Amazon Web Services
