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

Re: [PATCH v2 1/4] odb: decouple source path comparisons from `the_repository`

From
Patrick Steinhardt <ps@pks.im>
Date
Aug 17, 2026, 05:39 UTC
Message-ID
<aoKeeQMps50rjhWi@pks.im>
In-Reply-To
<20260814171724.GB2563235@coredump.intra.peff.net>
On Fri, Aug 14, 2026 at 01:17:24PM -0400, Jeff King wrote:
Show 38 quoted lines
> On Wed, Aug 12, 2026 at 11:13:57AM +0200, Patrick Steinhardt wrote:
> 
> > When registering alternates we deduplicate object database sources by
> > their path so that the same source won't be added twice. Ever since
> > cf2dc1c238 (speed up alt_odb_usable() with many alternates, 2021-07-07)
> > this duplicate check is backed by a map keyed by the source's path,
> > using `fspathhash()` and `fspatheq()` as hash and equality functions,
> > respectively.
> > 
> > These functions are problematic in this context for two reasons:
> > 
> >   - They implicitly depend on `the_repository` instead of the
> >     repository that owns the object database.
> 
> I'm not even sure that using core.ignorecase here is strictly correct.
> It is a property of the containing repository, and the filesystem in
> which it's stored. But there is no guarantee that the alternate
> directories are in the same repository, or even the same filesystem!
> 
> So it is really just a best guess proxy for "this system tends to use or
> not use case insensitive filesystems[1]". It can be wrong in both
> directions (failing to suppress duplicates, and suppressing them when
> they are not actually duplicates).
> 
> I wonder how bad it would be if we just always did case-sensitive
> comparisons and made it the caller's responsibility to spell things
> consistently.  I guess some names ultimately come from things like
> "--reference" command-line arguments, so that would depend on user
> spelling. But having duplicates at all is kind of unlikely (you can't
> get it from one --reference clone, but rather a complex tree of
> interwoven repos with shared roots).
> 
> How bad is a duplicate alternate? It's a minor performance issue, I'd
> think. We would add its packs to the list (though hardly ever look
> through them, as the "first" copy would satisfy most requests, and the
> unused second copies end up at the back of the MRU list). You'd only pay
> the extra lookup cost for an object which we fail to find entirely,
> which is rare-ish (mostly speculative lookups for fetches).

A performance regression is definitely the most likely change in behaviour we might see because of this. One other part I am a bit worried about is housekeeping, but I think we should be fine there as we only consider the primary source as special.

I also had the feeling that case insensitivity is quite a bit lacking, too. What we're really after is whether two directories are actually the exact same path. And whether the path is case-insensitive is only one part of that equation, so it's an imperfect metric by itself already.

Ideally, we should probably use realpath(3p) to at least also resolve symlinks. Unfortunately, it's not guaranteed that this function also knows to canonicalize casing.

Show 10 quoted lines
> And it would fix the unlikely-but-possible opposite case of suppressing
> a non-duplicate. If you have a repo on a case-insensitive filesystem
> with two alternates on a case-sensitive system that differ only in case,
> we erroneously suppress one of them, and commands may fail to find
> objects we should have. Of course that's super unlikely, which is why
> nobody has run into it before.
> 
> So I kind of wonder if we could just do away with considering case
> insensitivity here at all. We'd err on the side of correctness in the
> ambiguous cases, and this code complexity can just go away.

You will of course be able to craft edge cases where that would be a significant regression. But if your alternates file looks like this you may be holding it wrong:

    /path/to/alternate
    /PATH/TO/ALTERNATE
    /pAtH/tO/aLtErNaTe
    /PaTh/To/AlTeRnAtE
Show 7 quoted lines
> Alternatively, I think we could probably make the check more thorough in
> a similar way. Always consider a pair of case-insensitive matches as
> possible duplicates, and then for each possible duplicate use stat() to
> check their st_dev and st_ino values. That keeps things cheap for normal
> cases, and we pay only the stat() before de-duping. It's correct and
> doesn't rely on the repo, though it is a bit more somewhat complicated
> code.

Hm. Weren't there filesystems where `st_ino` and `st_dev` aren't set at all? I think that's the case on Windows, which is unfortunately also the one where we see case insensitive filesystems by default. So that makes it way less effective, as it only works on systems where we typically aren't case-insensitive in the first place (except macOS maybe).

So if we want to go down this path I'm inclined to just unconditionally use case sensitive matching and not introduce any secondary machinery.

Patrick
Previous: Jeff KingNext: Patrick Steinhardt
Message 21 of 47 in “odb: eagerly load alternates”
  1. 0/4 odb: eagerly load alternatesPatrick Steinhardt, Aug 10, 2026
  2. 1/4 odb: decouple source path comparisons from `the_repository`Patrick Steinhardt, Aug 10, 2026
  3. Justin ToblerAug 11, 2026
  4. Patrick SteinhardtAug 12, 2026
  5. 2/4 odb: eagerly initialize alternatesPatrick Steinhardt, Aug 10, 2026
  6. Justin ToblerAug 11, 2026
  7. Patrick SteinhardtAug 12, 2026
  8. 3/4 odb: drop `loaded_alternates` fieldPatrick Steinhardt, Aug 10, 2026
  9. Justin ToblerAug 11, 2026
  10. 4/4 odb: drop `alternates_db` fieldPatrick Steinhardt, Aug 10, 2026
  11. Justin ToblerAug 11, 2026
  12. Patrick SteinhardtAug 12, 2026
  13. 0/4 odb: eagerly load alternatesPatrick Steinhardt, Aug 12, 2026
  14. 1/4 odb: decouple source path comparisons from `the_repository`Patrick Steinhardt, Aug 12, 2026
  15. Karthik NayakAug 13, 2026
  16. Patrick SteinhardtAug 13, 2026
  17. Karthik NayakAug 14, 2026
  18. Jeff KingAug 14, 2026
  19. Junio C HamanoAug 14, 2026
  20. Jeff KingAug 14, 2026
  21. Patrick SteinhardtAug 17, 2026
  22. Patrick SteinhardtAug 17, 2026
  23. Jeff KingAug 17, 2026
  24. Patrick SteinhardtAug 17, 2026
  25. Jeff KingAug 17, 2026
  26. Jeff KingAug 17, 2026
  27. Jeff KingAug 14, 2026
  28. Patrick SteinhardtAug 17, 2026
  29. 2/4 odb: eagerly initialize alternatesPatrick Steinhardt, Aug 12, 2026
  30. 3/4 odb: drop `loaded_alternates` fieldPatrick Steinhardt, Aug 12, 2026
  31. Karthik NayakAug 13, 2026
  32. 4/4 odb: drop `alternates_db` fieldPatrick Steinhardt, Aug 12, 2026
  33. Junio C HamanoAug 12, 2026
  34. Patrick SteinhardtAug 13, 2026
  35. Karthik NayakAug 13, 2026
  36. Justin ToblerAug 13, 2026
  37. 0/5 odb: eagerly load alternatesPatrick Steinhardt, Aug 17, 2026
  38. 1/5 setup: create ref and object databases after config is writtenPatrick Steinhardt, Aug 17, 2026
  39. Karthik NayakAug 20, 2026
  40. 2/5 odb: decouple source path comparisons from `the_repository`Patrick Steinhardt, Aug 17, 2026
  41. Junio C HamanoAug 20, 2026
  42. Patrick SteinhardtAug 21, 2026
  43. Junio C HamanoAug 21, 2026
  44. 3/5 odb: eagerly initialize alternatesPatrick Steinhardt, Aug 17, 2026
  45. 4/5 odb: drop `loaded_alternates` fieldPatrick Steinhardt, Aug 17, 2026
  46. 5/5 odb: drop `alternates_db` fieldPatrick Steinhardt, Aug 17, 2026
  47. Karthik NayakAug 20, 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.