Re: [RFC PATCH] Modify fetch-pack to no longer die on error?

From: Jonathan Tan <jonathantanmy@google.com>
Date: Jul 29, 2020, 18:53 UTC
Message-ID: <20200729185347.653706-1-jonathantanmy@google.com>
In-Reply-To: <20200728200847.GA1019822@coredump.intra.peff.net>

Show 19 quoted lines

> On Tue, Jul 28, 2020 at 12:23:50PM -0700, Jonathan Tan wrote:
> 
> > > For this particular case, with the performance and all, I agree that
> > > the stupid and robust approach would be the best.
> > 
> > I'm concerned that we will be painting ourselves into a corner here - I
> > have been appreciating the richer interface that a C call provides us,
> > compared to sub-processes where we have to communicate through a single
> > input stream and a single output stream. For example, "wanted-refs" and
> > [...]
> 
> Yeah, that's a compelling reason. I'd have thought for this use case you
> could just say "hey, make sure these objects exist", which doesn't
> require a lot of communication. But often when I think things like that
> and start coding, it turns out to be much more complicated. So I am
> perfectly willing to believe you that it doesn't apply here. And even
> if it did, you're right that we may run into other spots that do need to
> pass back more information, but benefit from more lib-ified code that
> doesn't die().

Thanks. Just to be clear, I can't think of any hard stops to implementing lazy fetch in a sub-process right now, but (as you said) while programming I could discover something, or we could need something in the future.

Show 12 quoted lines

> Just to play devil's advocate for a moment...
> 
> > (Also, I think that debugging within a process is easier than debugging
> > across processes, but that might not be a concern that other people
> > share.)
> 
> This is definitely true sometimes, but I think is sometimes the
> opposite. When we push things out to a sub-process, then the interface
> between the two processes has to be well-defined (e.g., writing results
> to a file with a particular format). And that can make debugging easier,
> because you can pick up from that intermediate state (munging it in the
> middle, or even generating it from scratch for testing).

Well, unless there is some sort of interactivity like in remote helpers :-)

Show 5 quoted lines

> Likewise, that can result in a more flexible and robust system from the
> perspective of users. If we had invented "git log" first, we probably
> wouldn't have "git rev-list | git diff-tree --stdin" at all. But having
> that as two separate tools is sometimes useful for people doing things
> _besides_ log, since it gives different entry points to the code.

That's true. And the lazy fetch might be one of them - after looking at the code, I think that I can get to where I want just by implementing a null negotiator, which will be useful for end users. (In particular, simulating a lazy fetch might be useful sometimes, and re-fetching a packfile could be a crude way of repairing object corruption.)

Show 5 quoted lines

> That said, I think I could buy the argument that "fetch" works pretty
> well as a basic building block for users. It's pretty rare to actually
> use fetch-pack as a distinct operation. This is all a monolith vs module
> tradeoff question, and the tradeoff around modularity for fetch isn't
> that compelling.

If we are going the sub-process route, I was planning to use "fetch" as the sub-process, actually, not "fetch-pack" - among other things, "fetch" allows us to specify a fetch negotiator. So it seems like you are saying that if we had to use "fetch-pack", you have no problem with libifying it and calling it in the same process, but if we can use "fetch", we should use it as a sub-process?

In any case, I'll look into using "fetch" as a sub-process and see if it works.