I think attached v41nm is ready for commit. Would anyone like to vote against
back-patching this? It's hard to justify lack of back-patch for a data-loss
bug, but this is atypically invasive. (I'm repeating the question, since some
folks missed my 2020-02-18 question.) Otherwise, I'll push this on Saturday.
On Mon, Mar 23, 2020 at 05:20:27PM +0900, Kyotaro Horiguchi wrote:
> At Sat, 21 Mar 2020 15:49:20 -0700, Noah Misch <noah@leadboat.com> wrote in
> > The proximate cause is the RelFileNodeSkippingWAL() call that we added to
> > MarkBufferDirtyHint(). MarkBufferDirtyHint() runs in parallel workers, but
> > parallel workers have zeroes for pendingSyncHash and rd_*Subid.
> > Kyotaro, can you look through the affected code and propose a strategy for
> > good coexistence of parallel query with the WAL skipping mechanism?
>
> Bi-directional communication between leader and workers is too-much.
> It wouldn't be acceptable to inhibit the problematic operations on
> workers such like heap-prune or btree pin removal. If we do pending
> syncs just before worker start, it won't fix the issue.
>
> The attached patch passes a list of pending-sync relfilenodes at
> worker start.
If you were to issue pending syncs and also cease skipping WAL for affected
relations, that would fix the issue. Your design is better, though. I made
two notable changes:
- The patch was issuing syncs or FPIs every time a parallel worker exited. I
changed it to skip most of smgrDoPendingSyncs() in parallel workers, like
AtEOXact_RelationMap() does.
- PARALLEL_KEY_PENDING_SYNCS is most similar to PARALLEL_KEY_REINDEX_STATE and
PARALLEL_KEY_COMBO_CID. parallel.c, not execParallel.c, owns those. I
moved PARALLEL_KEY_PENDING_SYNCS to parallel.c, which also called for style
changes in the associated storage.c functions.
Since pendingSyncHash is always NULL under XLogIsNeeded(), I also removed some
XLogIsNeeded() tests that immediately preceded !pendingSyncHash tests.
