Re: .ready and .done files considered harmful

On 9/20/21, 1:42 PM, "Alvaro Herrera" <alvherre@alvh.no-ip.org> wrote:
> On 2021-Sep-20, Robert Haas wrote:
>
>> I was thinking that this might increase the number of directory scans
>> by a pretty large amount when we repeatedly catch up, then 1 new file
>> gets added, then we catch up, etc.
>
> I was going to say that perhaps we can avoid repeated scans by having a
> bitmap of future files that were found by a scan; so if we need to do
> one scan, we keep track of the presence of the next (say) 64 files in
> our timeline, and then we only have to do another scan when we need to
> archive a file that wasn't present the last time we scanned.  However:

This sounds a bit like the other approach discussed earlier in this
thread [0].

>> But I guess your thought process is that such directory scans, even if
>> they happen many times per second, can't really be that expensive,
>> since the directory can't have much in it. Which seems like a fair
>> point. I wonder if there are any situations in which there's not much
>> to archive but the archive_status directory still contains tons of
>> files.
>
> (If we take this stance, which seems reasonable to me, then we don't
> need to optimize.)  But perhaps we should complain if we find extraneous
> files in archive_status -- Then it'd be on the users' heads not to leave
> tons of files that would slow down the scan.

The simplest situation I can think of that might be a problem is when
checkpointing is stuck and the .done files are adding up.  However,
after the lengthy directory scan, you should still be able to archive
several files without a scan of archive_status.  And if you are
repeatedly catching up, the extra directory scans probably aren't
hurting anything.  At the very least, this patch doesn't make things
any worse in this area.

BTW I attached a new version of the patch with a couple of small
changes.  Specifically, I adjusted some of the comments and moved the
assignment of last_dir_scan to after the directory scan completes.
Before, we were resetting it before the directory scan, so if the
directory scan took too long, you'd still end up scanning
archive_status for every file.  I think that's still possible if your
archive_command is especially slow, but archiving isn't going to keep
up anyway in that case.

Nathan

[0] https://www.postgresql.org/message-id/attachment/125980/0001-Improve-performance-of-pgarch_readyXlog-with-many-st.patch