On Thu, May 27, 2021 at 10:22 PM Julien Rouhaud <rjuju...@gmail.com> wrote: > > On Thu, May 27, 2021 at 10:05 PM Paul Guo <paul...@gmail.com> wrote: > > > > Also note that ForwardSyncRequest() does wake up the checkpointer if > > it thinks the requests in shared memory are "too full", but does not > > wake up when the request is actually full. This does not seem to be > > reasonable. > > See below code in ForwardSyncRequest > > > > /* If queue is more than half full, nudge the checkpointer to empty it > > */ > > too_full = (CheckpointerShmem->num_requests >= > > CheckpointerShmem->max_requests / 2); > > > > /* ... but not till after we release the lock */ > > if (too_full && ProcGlobal->checkpointerLatch) > > SetLatch(ProcGlobal->checkpointerLatch); > > Ah indeed. Well it means that the checkpointer it woken up early > enough to avoid reaching that point. I'm not sure that it's actually > possible to reach a point where the list if full and the checkpointer > is sitting idle.
In theory this is possible (when the system is under heavy parallel write) else we could remove that part code (CompactCheckpointerRequestQueue()) :-), though the chance is not high. If we encounter this issue those affected queries would suddenly hang until the next checkpointer wakeup.