Re: Avoid race condition between fetch and repack/gc?

On Mon, Mar 16, 2020 at 10:27 AM Jeff King [off-list ref] wrote:

  - you don't use config like core.packedGitLimit that encourages Git to
    drop mmap'd windows. On 64-bit systems, this doesn't help at all (we
    have plenty of address space, and since the maps are read-only, the
    OS is free to drop the clean pages if there's memory pressure). On
    32-bit systems, it's a necessity (and I'd expect a 32-bit server to
    run into this issue a lot for large repositories).

I could be wrong, but I'm going to _guess_ from the :7999 in the
example output that the repository is hosted with Bitbucket Server.
Assuming my guess is correct, Bitbucket Server _does_ set
"core.packedgitlimit=256m" (and "core.packedgitwindowsize=32m", for
what it's worth). Those settings are applied specifically because
we've found they _do_ impact Git's overall memory usage when serving
clones in particular, which is important for cases where the system is
serving dozens (or hundreds) of concurrent clones.

Of course, we don't always do a great job of re-testing
once-beneficial settings later, which means sometimes they end up
being based on outdated observations. Perhaps we should prioritize
some additional testing here, especially on 64-bit systems. (We've
been setting "core.packedgitlimit" since back when Bitbucket Server
was called Stash and supported Git 1.7.6 on the server.)

That said, though, you note "core.packedgitlimit" is necessary on
32-bit servers and, unfortunately, we do still support Bitbucket
Server on 32-bit OSes. Maybe we should investigate applying (or not)
the flags depending on the platform. Sadly, that's not necessarily
simple to do since just because the _OS_ is 64-bit doesn't mean _Git_
is; it's pretty trivial to run 32-bit Git on 64-bit Windows, for
example.

Bryan