Hi,

I had another pass at this and have attached v3, now split into two patches.
The removal and recycling cases have one important difference: only recycling
exposes a new WAL filename to the write path.

0001 is the straightforward bit.  For segments that a checkpoint removes, it
does the unlinks first, fsyncs pg_wal once, and then cleans up the archive
status files.  I left the timeline-switch path alone.  This is not only for
wal_recycle=off; with recycling enabled, surplus segments still take the
removal path.

I reran the removal benchmark with the default 16 MiB WAL segment size.  The
numbers below are for 0001 alone; 0002 was not applied.  This was on ext4 with
fsync enabled and wal_recycle=off, so every eligible segment went through the
removal path.  Each number is the median of five alternating runs after three
warmups:

  files removed    master    0001
              1     22 ms    23 ms
              8     53 ms    32 ms
             32    135 ms    64 ms
             64    201 ms    97 ms
            128    394 ms   154 ms

The one-file case was within run-to-run variation, as expected.  Once there
are several files to remove, the per-file directory fsyncs start to show up;
at 128 files the time went from 394 ms to 154 ms, about 60% lower (2.52x).
This is not a TPS result, just checkpoint cleanup time.  Since 0001 avoids
the write-path changes in 0002, I thought it was worth splitting out for
separate consideration.

0002 is the awkward bit, and I have left it marked WIP.  Going back over v2,
I found a hole in the durable-segment high-water mark.  A foreground process
can find the names in the pending recycle batch already occupied, install its
own segment at a higher number, and move the frontier past lower renames that
are not durable yet.

The current patch plugs that hole with an active-batch marker.  While a batch
is in flight, XLogFileInit() takes the file-and-directory fsync fallback even
if the frontier says the segment is covered.  The new TAP test forces exactly
this interleaving; without the marker check, it times out waiting for the
fallback.

That handles the case I found, but I am still on the fence about 0002.  It
feels like a fair amount of plumbing between checkpoint cleanup and the WAL
write path just to save these fsyncs.  Is there a cleaner invariant that I am
missing, or a less invasive way to make the batched renames safe?

Regards,
Ayush

Attachment: v3-0001-Batch-fsyncs-when-removing-old-WAL-segments.patch
Description: Binary data

Attachment: v3-0002-WIP-Batch-fsyncs-when-recycling-WAL-segments.patch
Description: Binary data

Reply via email to