Re: Making WAL archiving faster — multi-file support and async ideas

Hi,
I have some concerns about the parallel archiver due to the requirement for
sequential WAL archiving. The fundamental rule is that WAL segments must be
archived in strict sequential order for a successful restore. Consider a
scenario where PostgreSQL has segments 1, 2, 3, and 4 ready, and the
parallel archiver successfully copies segments 1, 2, and 4, but misses
segment 3. The user might be unaware of the gap and could attempt a restore
using the incomplete archive. While we hope this would cause a clear error
during recovery, there is a risk that partial application of non-sequential
segments might lead to silent corruption or other unforeseen issues.

As far as I know, tools like cp naturally process files in the order they
are received. This ensures that files are processed correctly, and it seems
easier and more reliable to build on this concept than to introduce a new
parallel paradigm. Given that the archiver already uses a priority queue,
this should not be difficult to implement.

Thanks for considering my feedback.
--
Best regards,
Alyona Vinter

On Mon, 25 Aug 2025 at 11:00, Stepan Neretin <slpmcf(at)gmail(dot)com> wrote:

> Hi hackers,
>
> We’ve been thinking about how to make WAL archiving faster.
>
> This topic was previously discussed in [1], and we’ve taken a first step
> by implementing the attached patch, which adds support for archiving
> multiple WAL files in one go.
>
> The idea is straightforward: instead of invoking the archive command or
> callback once per WAL file, we allow passing a batch of files. The patch
> introduces support for new placeholders:
>
> -
>
> %F – list of WAL file names
> -
>
> %P – list of their full paths
> -
>
> %N – number of files
>
> Since PostgreSQL already reads multiple files into memory and caches them
> before archiving, this change avoids repeated fork() calls and reduces
> overhead in high-throughput setups.
>
> Of course, there are trade-offs. After discussing with Andrey Borodin, we
> noted that if even one file in the batch fails to archive, we currently
> have to retry the whole batch. While it’s technically possible to return a
> list of successfully archived files, that would complicate the API and
> introduce messy edge cases.
>
> So we’re also exploring a more flexible idea: an asynchronous archiver
> mode.
>
> The idea is to have PostgreSQL write WAL file names (marked .ready) into
> a FIFO or pipe, and let an archive process or library asynchronously
> consume and archive them. It would send back confirmations (or failures)
> through another pipe, allowing PostgreSQL to retry failed files as needed.
> This could decouple archiving from the archiver loop and open the door to
> more efficient and parallel implementations.
>
> We’d appreciate feedback on both directions:
>
> -
>
> Do you think the idea in the attached patch — batching WAL files for
> archiving — is viable? Is it something worth pursuing?
> -
>
> What do you think about the async archiver concept? Would it fit
> PostgreSQL’s architecture and operational expectations?
>
> Thanks,
> Stepan Neretin
>
> [1]
> https://www.postgresql.org/message-id/flat/BC335D75-105B-403F-9473-976C8BBC32E3%40yandex-team.ru#d45caa9d1075734567164f73371baf00
>
>