On Mon, Aug 6, 2018 at 2:29 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Hm. Post-beta3, I think I'd vote for a conservative fix in v11,
> which seems to be "ban for mapped catalogs". Feel free to make
> it work in HEAD, though.
Makes sense. I'm not sure if it's worth pursuing parallel mapped
catalog index builds much further, though. I doubt that ordinary users
care about whether or not this is supported, so this is a matter of
principle. I don't feel strongly on whether or not I should make
mapped builds work on HEAD, so I defer to you, and anyone else that
might have an interest. Does it matter, do you think?
It might be worth teaching heap_beginscan_parallel() to
cross-check each worker's heap relation's rd_smgr.smgr_node to a version
of the same field from the leader, stored in shared memory (in the
ParallelHeapScanDesc). That way, any future recurrence of a similar
bug will be far easier to detect. A "can't happen" error along these
lines seems like it would be worthwhile.
--
Peter Geoghegan
