Re: [PATCH 5/8] commit-graph: not compatible with replace objects

[Derrick Stolee <stolee@gmail.com>]

On 7/18/2018 3:52 PM, Derrick Stolee wrote:
> On 7/18/2018 3:46 PM, Jeff King wrote:
>> On Wed, Jul 18, 2018 at 08:15:41AM -0700, Derrick Stolee via 
>> GitGitGadget wrote:
>>
>>> From: Derrick Stolee <dstolee@microsoft.com>
>>>
>>> Create new method commit_graph_compatible(r) to check if a given
>>> repository r is compatible with the commit-graph feature. Fill the
>>> method with a check to see if replace-objects exist. Test this
>>> interaction succeeds, including ignoring an existing commit-graph and
>>> failing to write a new commit-graph.
>> I think this approach is sensible. These are optimizations, and it's not
>> a big deal to just punt no cases we can't handle.
>>
>> I do have one complaint about the implementation, though:
>>
>>> +static int commit_graph_compatible(struct repository *r)
>>> +{
>>> +    prepare_replace_object(r);
>>> +    if (hashmap_get_size(&r->objects->replace_map->map))
>>> +        return 0;
>>> +
>>> +    return 1;
>>> +}
>> If I'm reading the code correctly, this will predicate the decision
>> entirely on the presence of refs in refs/replace/. But you may have a
>> situation where those refs exist, but you are not respecting them in
>> this run.
>>
>> For instance, imagine[1] a server that hosts arbitrary repositories, but
>> wants to use commit graphs to speed up server-side operations (e.g.,
>> serving fetches, but also perhaps a web interface doing --contains,
>> etc). If it runs all of the server-side commands with
>> GIT_NO_REPLACE_OBJECTS=1, then there should be no problem. But if a user
>> pushes anything to refs/replace (because they _do_ use replace refs
>> locally, and want to share them with other clients), that would disable
>> graphs on the server.
>>
>> So I think this should at least be loosened to:
>>
>>    if (check_replace_refs) {
>>     prepare_replace_object(r);
>>     if (...)
>>    }
>>
>> which would make this case work. I'd even go so far as to say that for
>> writing, we should just always ignore replace refs and generate the full
>> graph (just like pack-objects does so when writing a pack). Then the
>> resulting graph can be used selectively by disabling replace refs for
>> particular commands. But for the scenario I described above, the
>> distinction is mostly academic, as replacements would be disabled anyway
>> during the write command anyway.
>>
>> -Peff
>>
>> [1] You can probably guess that this is how GitHub handles replace refs.
>>      We ran into this long ago because replacements and grafts mess up
>>      any other caches external to Git that rely on the immutability of
>>      the hash.
>>
>>      We do it with a config option, though, which requires a trivial
>>      patch. I'll post that momentarily.
>
> Thanks for the details! I never considered that someone would have 
> these refs around, but would ignore them most of the time.
>
> The biggest reason I wanted to punt here was that it is easy to toggle 
> between using replace refs and not using them. Writing and reading as 
> long as we are ignoring those refs is a good idea, and I'll use that 
> approach in v2.


Thanks, Peff, for improving the check_replace_refs interaction [1].

Since this series now has two dependencies (including Stefan's ref-store 
fix [2] that I had included in my v1), I'll let those topics settle 
before I send a new v2.

If there are more comments about how I'm handling these cases, then 
please jump in and tell me. I'll revisit this topic in a few weeks.

Thanks,

-Stolee

[1] 
https://public-inbox.org/git/20180718204449.GA31816@sigill.intra.peff.net/T/#u

      [PATCH 1/3] check_replace_refs: fix outdated comment

[2] 
https://public-inbox.org/git/20180717224935.96397-1-sbeller@google.com/T/#u

     [PATCH 0/2] RFC ref store to repository migration