Horton

RE: Is this a problem in GenericXLogFinish()?

Hayato Kuroda (Fujitsu) <kuroda.hayato@fujitsu.com>

View thread View on postgr.es
From: "Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>
To: 'Michael Paquier' <michael@paquier.xyz>
Cc: Amit Kapila <amit.kapila16@gmail.com>, Alexander Lakhin <exclusion@gmail.com>, Robert Haas <robertmhaas@gmail.com>, Jeff Davis <pgsql@j-davis.com>, Heikki Linnakangas <hlinnaka@iki.fi>, "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Date: 2024-04-09T10:25:36Z
Lists: pgsql-hackers

Attachments

Dear Michael,

> On Fri, Apr 05, 2024 at 06:22:58AM +0000, Hayato Kuroda (Fujitsu) wrote:
> > Thanks for pointing out. Yes, we fixed a behavior by the commit aa5edbe37,
> > but we missed the redo case. I made a fix patch based on the suggestion [1].
> 
> +       bool        mod_buf = false;
> 
> Perhaps you could name that mod_wbuf, to be consistent with the WAL
> insert path.

Right, fixed.

> I'm slightly annoyed by the fact that there is no check on
> is_prim_bucket_same_wrt for buffer 1 in the BLK_NEEDS_REDO case to
> show the symmetry between the insert and replay paths.  Shouldn't
> there be at least an assert for that in the branch when there are no
> tuples (imagine an else to cover xldata->ntups == 0)?  I mean between
> just before updating the hash page's opaque area when
> is_prev_bucket_same_wrt.

Indeed, added an Assert() in else part. Was it same as your expectation?

> I've been thinking about ways to make such cases detectable in an
> automated fashion.  The best choice would be
> verifyBackupPageConsistency(), just after RestoreBlockImage() on the
> copy of the block image stored in WAL before applying the page masking
> which would mask the LSN.  A problem, unfortunately, is that we would
> need to transfer the knowledge of REGBUF_NO_CHANGE as a new BKPIMAGE_*
> flag so we would be able to track if the block rebuilt from the record
> has the *same* LSN as the copy used for the consistency check.  So
> this edge consistency case would come at a cost, I am afraid, and the
> 8 bits of bimg_info are precious :/

I could not decide that the change has more benefit than its cost, so I did
nothing for it.

Best Regards,
Hayato Kuroda
FUJITSU LIMITED
https://www.fujitsu.com/

Commits

  1. Fix inconsistency with replay of hash squeeze record for clean buffers

  2. Set LSN for wbuf in _hash_freeovflpage() iff wbuf is modified.

  3. Fix an uninitialized access in hash_xlog_squeeze_page().

  4. Use REGBUF_NO_CHANGE at one more place in the hash index.

  5. Assert that buffers are marked dirty before XLogRegisterBuffer().

  6. Fix bug in GenericXLogFinish().

  7. Improve hash index bucket split behavior.