| From: | Greg Lamberson <greg(at)lamco(dot)io> |
|---|---|
| To: | Yash Jadhav <yash(dot)jadhav8008(at)gmail(dot)com> |
| Cc: | "pgsql-hackers(at)lists(dot)postgresql(dot)org" <pgsql-hackers(at)lists(dot)postgresql(dot)org> |
| Subject: | Re: Extensible sync handler registration (register_sync_handler) |
| Date: | 2026-11-16 19:36:00 |
| Message-ID: | CH0PR07MB9822B995B1146427EF1AF0DFA97C2@CH0PR07MB9822.namprd07.prod.outlook.com |
| Views: | Whole Thread | Raw Message | Download mbox | Resend email |
| Thread: | |
| Lists: | pgsql-hackers |
Hi Yash,
Thank you for taking the time to study and test the patch, and for
reproducing the crash with test_sync_handler. You're right: Before this
patch every handler lived in core, so a NULL optional callback could
never be reached, and now an extension can register one. A NULL check
before each optional callback is worth adding, and I'd gladly take your
incremental patch.
Since the patch turns syncsw[] from a static array into one an extension
can extend, I also measured what that costs at run time. I compared
master at 009ea1b08d7 with and without 0001, both built with meson's
release settings (-O3, GCC 14.2.0) on a 6-vCPU KVM guest:
- Checkpoints with 2001 fsync dispatches each, on tmpfs: Timed with a
uprobe, ProcessSyncRequests() took 6868.4 us per call stock and
6882.2 us patched (1.0020), with a 1.9% spread between runs of the
same build. The checkpointer's instruction count under perf stat
differed by 0.02%, against a 0.07% spread.
- Checkpoints after 500 tables were dropped and recreated:
SyncPostCheckpoint() took 1931.3 us and 1932.2 us per call (1.0005),
with a 1.6% spread.
- pgbench TPC-B on tmpfs: 11122 and 11166 tps, inside a 6.7% spread.
Nothing was measurable in either direction. The smallest difference
this setup could have detected was about 65 ns per dispatched file, and
the measured one was about 7 ns. With the client on the same vCPUs of a
shared VM, I'd read that as "not measurable here" rather than zero.
The generated code agrees, and it sharpens what I said about it before.
The per-file dispatch sequence is unchanged and each of the three
functions keeps its instruction count, at -O3 and at -O2. In
ProcessSyncRequests() the load of the table pointer replaces an address
computation rather than adding an instruction. SyncPostCheckpoint() runs
2 more instructions per call when unlinks are pending. In
RememberSyncRequest() the change sits only in the SYNC_FILTER_REQUEST
branch that DROP DATABASE uses, where the patched loop is slightly
shorter; that path wasn't measured at run time.
I also owe the thread a correction. My first message said Lamstore's
storage manager works today against stock PostgreSQL and Percona Server
for PostgreSQL 18. That holds on Percona, which provides smgr_register(),
but not on stock PostgreSQL, which gives an extension no way to register
a storage manager.
Regards,
Greg
| From | Date | Subject | |
|---|---|---|---|
| Next Message | Zsolt Parragi | 2026-11-16 19:36:32 | Re: [PATCH] Preserve replication origin OIDs in pg_upgrade |
| Previous Message | Zsolt Parragi | 2026-11-16 19:20:54 | Re: Use correct collation in pg_trgm |