Avoid endless futile table locks in vacuuming.
If a partially-active table develops a slug of stable all-visible,
non-empty pages at the end of it, then every autovacuum of that table
will skip the end pages on the forward scan, think they might be
truncatable, and take the access exclusive lock to do the truncation.
And then immediately fail when it sees the last page is not empty.
This can persist for an indefinite number of autovac rounds.
This is not generally a problem, as the lock is taken conditionally.
However, the lock is also logged and passed over to any hot standbys,
where it must be replayed unconditionally. This can cause query
cancellations.
The simple solution is to always scan the last page of a table, so it
can be noticed that it is not empty and avoid the truncation attempt.
We could add logic like doing this scan only if wal_level is
hot_standby or higher, or reproducing the REL_TRUNCATE_FRACTION logic
here to scan the last page only if truncation is eminent. But those
seem like needless complications to try to avoid sometimes scanning
one block.
Cheers,
Jeff
Attachments:
always_scan_last_page_v1.patchtext/x-patch; charset=US-ASCII; name=always_scan_last_page_v1.patchDownload+8-2
Jeff Janes <jeff.janes@gmail.com> writes:
If a partially-active table develops a slug of stable all-visible,
non-empty pages at the end of it, then every autovacuum of that table
will skip the end pages on the forward scan, think they might be
truncatable, and take the access exclusive lock to do the truncation.
And then immediately fail when it sees the last page is not empty.
This can persist for an indefinite number of autovac rounds.
This is not generally a problem, as the lock is taken conditionally.
However, the lock is also logged and passed over to any hot standbys,
where it must be replayed unconditionally. This can cause query
cancellations.
The simple solution is to always scan the last page of a table, so it
can be noticed that it is not empty and avoid the truncation attempt.
This seems like a reasonable proposal, but I find your implementation
unconvincing: there are two places in lazy_scan_heap() that pay attention
to scan_all, and you touched only one of them. Thus, if we fail to
acquire cleanup lock on the table's last page on the first try, the
change is for naught. Shouldn't we be insisting on getting that lock?
Or, if you intentionally didn't change that because it seems like too
high a price to pay, why doesn't the comment reflect that?
regards, tom lane
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
I wrote:
Jeff Janes <jeff.janes@gmail.com> writes:
If a partially-active table develops a slug of stable all-visible,
non-empty pages at the end of it, then every autovacuum of that table
will skip the end pages on the forward scan, think they might be
truncatable, and take the access exclusive lock to do the truncation.
And then immediately fail when it sees the last page is not empty.
This can persist for an indefinite number of autovac rounds.
The simple solution is to always scan the last page of a table, so it
can be noticed that it is not empty and avoid the truncation attempt.
This seems like a reasonable proposal, but I find your implementation
unconvincing: there are two places in lazy_scan_heap() that pay attention
to scan_all, and you touched only one of them.
After further investigation, there is another pre-existing bug: the short
circuit path for pages not requiring freezing doesn't bother to update
vacrelstats->nonempty_pages, causing the later logic to think that the
page is potentially truncatable even if we fix the second check of
scan_all! So this is pretty broken, and I almost think we should treat it
as a back-patchable bug fix.
In the attached proposed patch, I added another refinement, which is to
not bother with forcing the last page to be scanned if we already know
that we're not going to attempt truncation, because we already found a
nonempty page too close to the end of the rel. I'm not quite sure this
is worthwhile, but it takes very little added logic, and saving an I/O
is probably worth the trouble.
regards, tom lane
Attachments:
always_scan_last_page_v2.patchtext/x-diff; charset=us-ascii; name=always_scan_last_page_v2.patchDownload+85-49
On Mon, Dec 28, 2015 at 2:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
I wrote:
Jeff Janes <jeff.janes@gmail.com> writes:
If a partially-active table develops a slug of stable all-visible,
non-empty pages at the end of it, then every autovacuum of that table
will skip the end pages on the forward scan, think they might be
truncatable, and take the access exclusive lock to do the truncation.
And then immediately fail when it sees the last page is not empty.
This can persist for an indefinite number of autovac rounds.
The simple solution is to always scan the last page of a table, so it
can be noticed that it is not empty and avoid the truncation attempt.This seems like a reasonable proposal, but I find your implementation
unconvincing: there are two places in lazy_scan_heap() that pay attention
to scan_all, and you touched only one of them.After further investigation, there is another pre-existing bug: the short
circuit path for pages not requiring freezing doesn't bother to update
vacrelstats->nonempty_pages, causing the later logic to think that the
page is potentially truncatable even if we fix the second check of
scan_all! So this is pretty broken, and I almost think we should treat it
as a back-patchable bug fix.In the attached proposed patch, I added another refinement, which is to
not bother with forcing the last page to be scanned if we already know
that we're not going to attempt truncation, because we already found a
nonempty page too close to the end of the rel. I'm not quite sure this
is worthwhile, but it takes very little added logic, and saving an I/O
is probably worth the trouble.
If we are not doing a scan_all and we fail to acquire a clean-up lock on
the last block, and the last block reports that it needs freezing, then we
continue on to wait for the clean-up lock. But there is no need, we don't
really need to freeze the block, and we already know whether it has tuples
or not without the clean up lock. Couldn't we just set the flag based on
hastup, then 'continue'?
Cheers,
Jeff
Jeff Janes <jeff.janes@gmail.com> writes:
If we are not doing a scan_all and we fail to acquire a clean-up lock on
the last block, and the last block reports that it needs freezing, then we
continue on to wait for the clean-up lock. But there is no need, we don't
really need to freeze the block, and we already know whether it has tuples
or not without the clean up lock. Couldn't we just set the flag based on
hastup, then 'continue'?
Uh, isn't that what my patch is doing?
regards, tom lane
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
Forgetting to CC the list is starting to become a bad habit, sorry.
forwarding to list:
---------- Forwarded message ----------
From: Jeff Janes <jeff.janes@gmail.com>
Date: Wed, Dec 30, 2015 at 10:03 AM
Subject: Re: [HACKERS] Avoid endless futile table locks in vacuuming.
To: Tom Lane <tgl@sss.pgh.pa.us>
On Dec 29, 2015 4:47 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:
Jeff Janes <jeff.janes@gmail.com> writes:
If we are not doing a scan_all and we fail to acquire a clean-up lock on
the last block, and the last block reports that it needs freezing, then we
continue on to wait for the clean-up lock. But there is no need, we don't
really need to freeze the block, and we already know whether it has tuples
or not without the clean up lock. Couldn't we just set the flag based on
hastup, then 'continue'?Uh, isn't that what my patch is doing?
My reading was it does that only if there are no tuples that could be
frozen. If there are tuples that could be frozen, it actually does
the freezing, even though that is not necessary unless scan_all is
true.
So like the attached, although it is a bit weird to call
lazy_check_needs_freeze if , under !scan_all, we don't actually care
about whether it needs freezing but only the hastup.
Cheers,
Jeff
Attachments:
always_scan_last_page_v3.patchtext/x-patch; charset=US-ASCII; name=always_scan_last_page_v3.patchDownload+91-49
Import Notes
Reply to msg id not found: CAMkU1w3mQF8fw0N21ARO5KgSedDhVYyZ06f1n9qRyRZyjvHAA@mail.gmail.com
Jeff Janes <jeff.janes@gmail.com> writes:
On Dec 29, 2015 4:47 PM, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:
Uh, isn't that what my patch is doing?
My reading was it does that only if there are no tuples that could be
frozen. If there are tuples that could be frozen, it actually does
the freezing, even though that is not necessary unless scan_all is
true.
Ah, now I see.
So like the attached, although it is a bit weird to call
lazy_check_needs_freeze if , under !scan_all, we don't actually care
about whether it needs freezing but only the hastup.
True, but this is such a corner case that it doesn't seem worth expending
additional code to have a special-purpose page scan for it.
regards, tom lane
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
Jeff Janes <jeff.janes@gmail.com> writes:
So like the attached, although it is a bit weird to call
lazy_check_needs_freeze if , under !scan_all, we don't actually care
about whether it needs freezing but only the hastup.
I think this misses unpinning the buffer in the added code path.
I rearranged to avoid that, did some other cosmetic work, and committed.
regards, tom lane
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers