I reviewed your latest code, and it's nearly complete. mdimmedsync() syncs only "active segments" (as defined in md.c), but smgrDoPendingSyncs() must sync active and inactive segments. This matters when mdtruncate() truncated the relation after the last checkpoint, causing active segments to become inactive. In such cases, syncs of the inactive segments will have been queued for execution during the next checkpoint. Since we skipped the XLOG_SMGR_TRUNCATE record, we must complete those syncs before commit. Let's just modify smgrimmedsync() to always sync active and inactive segments; that's fine to do in other smgrimmedsync() callers, even though they operate on relations that can't have inactive segments.
On Tue, Dec 03, 2019 at 08:51:46PM +0900, Kyotaro Horiguchi wrote: > At Thu, 28 Nov 2019 17:23:19 -0500, Noah Misch <n...@leadboat.com> wrote in > > On Thu, Nov 28, 2019 at 09:35:08PM +0900, Kyotaro Horiguchi wrote: > > > I measured the performance with the latest patch set. > > > > > > > 1. Determine $DDL_COUNT, a number of DDL transactions that take about > > > > one > > > > minute when done via syncs. > > > > 2. Start "pgbench -rP1 --progress-timestamp -T180 -c10 -j10". > > > > 3. Wait 10s. > > > > 4. Start one DDL backend that runs $DDL_COUNT transactions. > > > > 5. Save DDL start timestamp, DDL end timestamp, and pgbench output. > wal_buffers | 512 This value (4 MiB) is lower than a tuned production system would have. In future benchmarks (if any) use wal_buffers=2048 (16 MiB).