Re: [RFC v3] cat-file: add a --stdin-cmd mode

From: Taylor Blau <me@ttaylorr.com>
Date: Feb 1, 2022, 20:14 UTC
Message-ID: <YfmUiXJ//ZC5NZU+@nand.local>
In-Reply-To: <674CD40F-7060-4892-97CD-1940AEF7C3D7@gmail.com>

On Tue, Feb 01, 2022 at 02:27:54PM -0500, John Cai wrote:

Show 10 quoted lines

> On 1 Feb 2022, at 12:52, Taylor Blau wrote:
> > I'm not sure that I've seen a response along the lines of "we need to
> > control when the output stream is flushed in order to do ..." yet, but I
> > would be interested to see one before moving too much further ahead of
> > where we already are.
>
> This would be useful when there is another process A interacting with
> a long running git cat-file process B that is retrieving object
> information from the odb interactively but also wants to use --buffer
> mode.

Let me try and repeat my understanding of what you said to make sure that I fully grok the use-case you have in mind.

You have a repository and want to have a long-running `git cat-file` process that can serve multiple requests. Because the processes which interact with your long-running `cat-file` may ask for many objects, you don't want to flush the output buffer after each object, and so would ideally like to use `--buffer`.

But that doesn't quite work, since the `cat-file` process may not have decided to flush its output buffer even when process A is about to go away.

I wonder about the viability of accomplishing this via a signal handler, i.e., that `cat-file` would call fflush(2) whenever it receives e.g., SIGUSR1. A couple of possible downsides:

  - SIGUSR1 doesn't exist on Windows AFAIK.

  - There are definitely going to be synchrony issues to contend with.
    What happens if we receive our signal while writing to the output
    stream? I think you would just need to mark a variable that
    indicates we should flush after finishing serving the current
    request, but I haven't thought too hard about it.

So maybe a signal isn't the way to go. But I don't think `--stdin-cmd` is the simplest approach either. At the very least, I don't totally understand your plan after implementing a flush command. You mention that it would be nice to implement other commands, but I'm not totally convinced by your examples[1].

I wonder if we could strike a middle ground, which might look like `git cat-file --batch --buffer`, and just feeding it something which we know for certain isn't an object identifier. In other words, what if we did something as simple as:

--- >8 ---

diff --git a/builtin/cat-file.c b/builtin/cat-file.c
index d94050e6c1..bae162fc18 100644
--- a/builtin/cat-file.c
+++ b/builtin/cat-file.c
@@ -595,6 +595,11 @@ static int batch_objects(struct batch_options *opt)
 	warn_on_object_refname_ambiguity = 0;

 	while (strbuf_getline(&input, stdin) != EOF) {
+		if (!strcmp("<flush>", input.buf)) {
+			fflush(stdout);
+			continue;
+		}
+
 		if (data.split_on_whitespace) {
 			/*
 			 * Split at first whitespace, tying off the beginning

--- 8< ---

On the other hand, something even hackier than the above is that we
flush stdout whenever we get a request to print an object which could
not be found. So if you feed a single "\n" to your `cat-file` process,
you'll get " missing" on its output, and the buffer will immediately be
flushed.

I'm not sure that I'd recommend relying on that behavior exactly, but if
you're looking for a short-term solution, it might work ;).

Thanks,
Taylor

[1]: One that comes to mind is changing the output format mid-stream.
     But how often does it really make sense to change the output
     format? I can understand wanting to flush at the end asking
     cat-file for a bunch of objects, but I don't see how you would want
     to change the output format often enough that shaving off Git's
     negligible startup cost is worthwhile (or couldn't be accomplished
     by just spawning another cat-file process and using that).