git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 1/2] maintenance: account for promisor pack geometry

From
Patrick Steinhardt <ps@pks.im>
Date
Aug 11, 2026, 10:01 UTC
Message-ID
<anry8wAbkxNfVgfh@pks.im>
In-Reply-To
<ann0nnSGfSJ7y7YK@com-79390>
On Mon, Aug 10, 2026 at 10:56:14AM -0500, Taylor Blau wrote:
Show 41 quoted lines
> On Mon, Aug 10, 2026 at 05:11:32PM +0200, Patrick Steinhardt wrote:
> > > Check both progressions before choosing the repack mode. If either
> > > leaves a pack above its split, geometric repack still avoids rewriting
> > > that pack, whereas the all-into-one fallback would rewrite it. Use the
> > > fallback only when neither progression leaves a pack behind. That
> > > preserves the reason for the fallback: let the all-into-one repack
> > > handle unreachable objects when it is not rewriting more packs than the
> > > geometric repack.
> >
> > Okay. The consequence of the status quo could be that we perform an
> > all-into-one repack more frequent than really desired because the set of
> > non-promised packs is small, and thus even writing a small set of new
> > objects could cause a full repack.
> 
> Right. I stumbled on this after a few colleagues had reported that their
> geometric maintenance task didn't seem to be doing anything. When
> looking into it, I found that they had many promisor packs, but the
> non-promisor packs were already in a geometric progression, and thus we
> did an all-into-one repack.
> 
> > This might create the reverse situation though, where we don't perform
> > the all-into-one repack at all anymore. We could come up with a clever
> > solution here, like for example considering both sequences together and
> > repacking when we cross a certain combined threshold. But I'm not sure
> > it's worth it for now, and we can still evolve the strategy as needed.
> 
> The change in this patch means that we will perform a geometric repack
> when doing so would result in a new geometrically-repacked series of
> promisor packs, in addition to non-promisor ones.
> 
> Is your concern that the non-promisor packs might be in a state where we
> should compact them into a single pack, but that the sequence of
> promisor packs would prevent us from doing so? In that case, we will
> perform a geometric repack on both sets of packs independently. If the
> non-promisor packs should be rolled up into a single pack (i.e.,
> "geometry.split == geometry.pack_nr"), then the geometric repack *will*
> produce a single pack, as if we had performed an all-into-one repack on
> the set of non-promisor packs.
> 
> So I am not sure that I understand your concern here, but please let me
> know if I am missing some aspect of it.

The concern is that it's quite unlikely that both the geometric and non-geometric sequence will merge all packs together at the same point in time. Consequently, we'll never hit the case where we perform an all-into-one pack to prune unreachable objects, and that may cause us to never prune objects at all.

So what I'm wondering is whether we should be a bit more clever about that and perform an all-into-one repack under a new condition, like for example when the objects we're about to repack exceed a certain percentage of the repository size.

Hope that clarifies it a bit :)
Thanks!
Patrick
Previous: Taylor BlauNext: Taylor Blau
Message 5 of 10 in “maintenance: handle geometric repack tasks with promisor pack(s)”
  1. 0/2 maintenance: handle geometric repack tasks with promisor pack(s)Taylor Blau, Aug 5, 2026
  2. 1/2 maintenance: account for promisor pack geometryTaylor Blau, Aug 5, 2026
  3. Patrick SteinhardtAug 10, 2026
  4. Taylor BlauAug 10, 2026
  5. Patrick SteinhardtAug 11, 2026
  6. Taylor BlauAug 11, 2026
  7. Patrick SteinhardtAug 11, 2026
  8. 2/2 maintenance: trigger --auto for promisor rollupsTaylor Blau, Aug 5, 2026
  9. Patrick SteinhardtAug 10, 2026
  10. Taylor BlauAug 10, 2026

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.