On Fri, Apr 26 2019, Junio C Hamano wrote:
> Ævar Arnfjörð Bjarmason <avarab@xxxxxxxxx> writes:
>
>> On Wed, Apr 24 2019, Derrick Stolee via GitGitGadget wrote:
>>
>>> NOTE: this series was rebased onto ab/commit-graph-fixes, as the conflicts
>>> were significant and subtle.
>>
>> Sorry, hopefully it helped more than it harmed :)
>>
>> A few unrelated things:
>
> Thanks always for your careful review and thoughtful comments, by
> the way.
>
>> Now as noted in my series we now on 'master' downgrade that to a warning
>> (along with the rest of the errors):
>>
>> $ ~/g/git/git --exec-path=$PWD status
>> error: commit-graph version 2 does not match version 1
>> On branch master
>> [...]
>>
>> ...and this series sets the default version for all new graphs to v2.
>
> The phrasing seems odd. It is unclear, even to me who is vaguely
> familiar with the word "commit-graph" and is aware of the fact that
> the file format is being updated, what
>
> "commit-graph version 2 does not match version 1"
Yeah it should really say:
"commit-graph is of version 2, our maximum supported version is 1"
Hindsight is 20/20, but more generally I wonder if we should have these
format versions match that of the git version (unlikely to change it
twice in the same release...) which would allow us to say things like:
"commit-graph needs v2.22.0 or later, we have the version written by v2.18.0..v2.21.0"
But of course dealing with those larger integers in the code/gaps is
also messy :)
> wants to say. Do I have version #2 on disk and the running binary
> only understands version #1? Or do I have version #1 on disk and
> the binary expected version #2? How would I get out of this
> situation? Is it sufficient to do "rm -f .git/info/commit-graph*"
> and is it safe?
Yeah. An rm of .git/info/commit-graph is safe, so is "-c
core.commitGraph=false" as a workaround.
I'd say "let's improve the error", but that ship has sailed, and we can
do better than an error here, no matter how it's phrased...
>> I think this is *way* too aggressive of an upgrade path. If these
>> patches go into v2.22.0 then git clients on all older versions that grok
>> the commit graph (IIRC v2.18 and above) will have their git completely
>> broken if they're in a mixed-git-version environment.
>>
I should note that "all older versions..." here is those that have
core.commitGraph=true set. More details in 43d3561805 ("commit-graph
write: don't die if the existing graph is corrupt", 2019-03-25).
>> Is it really so important to move to v2 right away that we need to risk
>> those breakages? I think even with my ab/commit-graph-fixes it's still
>> too annoying (I was mostly trying to fix other stuff...). If only we
>> could detect "we should make a new graph now" ....
>
> True.
Having slept on my earlier
https://public-inbox.org/git/87a7gdspo4.fsf@xxxxxxxxxxxxxxxxxxx/ I think
I see a better way to deal with this than my earlier suggestion that we
perform some version flip-flop dance on the single "commit-graph" file:
How about just writing .git/objects/info/commit-graph-v2, and for the
upcoming plan when where they'll be split have some dir/prefix there
where we include the version?
That means that:
1. If there's an existing v1 "commit-graph" file we don't write a v2 at
that path in v2.22, although we might have some "write v1 (as well
as v2?) for old client compat" config where we opt-in to do that.
2. By default in v2.22 we read/write a "commit-graph-v2" file,
preferring it over the v1 "commit-graph", falling back on earlier
versions if it's not there (until gc --auto kicks in on v2.22 and
makes a v2 graph).
3. If you have concurrent v2.21 and v2.22 clients accessing the repo
you might end up generating one commit-graph or the other depending
on who happens to trigger "gc --auto".
Hopefully that's a non-issue since an out-of-date graph isn't
usually a big deal, and client versions mostly march forward. But
v2.22 could also learn some "incremental gc" where it says "my v2 is
older, v1 client must have refreshed it, I'll refresh mine/both".
4. v2.22 and newer versions will have some code in git-gc where we'll
eventually readdir() .git/objects/info and remove graphs that are
too old per some new config (say
"gc.pruneOlderCommitGraphVersions=180 days").
This means that:
A. GOOD: Now and going forward we can fearlessly create new versions of
the graph without worrying/testing how older clients deal with it.
B. BAD: We are going to eat ~2x the disk space for commit-graphs while
such transitions are underway. I think that's fine. They're
relatively small compared to .git/objects, and we'll eventually "gc"
the old ones.
C. BAD: Different versions of git might perform wildly differently (new
version slower) since their respective preferred graph versions
might have a very different/up-to-date number of commits v.s. what's
in the packs.
I think "A" outweighs "B" && "C" in this case. It's "just" a caching
data structure, and git works without it. So we can be a lot looser than
say updating the index/pack format.
Worst case things slow down but still work, and as noted in #3 above if
we care it can be mitigated (but I don't, I think we can safely assume
"client versions march forward").
