Re: remove more archiving overhead

On Mon, Feb 21, 2022 at 05:19:48PM -0800, Nathan Bossart wrote:
> I also spent some time investigating whether durably renaming the archive
> status files was even necessary. In theory, it shouldn't be. If a crash
> happens before the rename is persisted, we might try to re-archive files,
> but your archive_command/archive_library should handle that. If the file
> was already recycled or removed, the archiver will skip it (thanks to
> 6d8727f). But when digging further, I found that WAL file recyling uses
> durable_rename_excl(), which has the following note:
>
> * Note that a crash in an unfortunate moment can leave you with two links to
> * the target file.
>
> IIUC this means that in theory, a crash at an unfortunate moment could
> leave you with a .ready file, the file to archive, and another link to that
> file with a "future" WAL filename. If you re-archive the file after it has
> been reused, you could end up with corrupt WAL archives. I think this
> might already be possible today. Syncing the directory after every rename
> probably reduces the likelihood quite substantially, but if
> RemoveOldXlogFiles() quickly picks up the .done file and attempts
> recycling before durable_rename() calls fsync() on the directory,
> presumably the same problem could occur.

In my testing, I found that when I killed the server just before unlink()
during WAL recyling, I ended up with links to the same file in pg_wal after
restarting. My latest test produced links to the same file for the current
WAL file and the next one. Maybe WAL recyling should use durable_rename()
instead of durable_rename_excl().

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com