Re: suppress fatal pathspec errors from "git add"?

From: Jeff King <peff@peff.net>
Date: Jan 3, 2010, 08:28 UTC
Message-ID: <20100103082804.GA23748@sigill.intra.peff.net>
In-Reply-To: <7vk4vz38v0.fsf@alter.siamese.dyndns.org>

On Sun, Jan 03, 2010 at 12:12:19AM -0800, Junio C Hamano wrote:

Show 19 quoted lines

> If your primary activity that happens in the work tree were:
> 
> 	while :
>         do
>         	file=$(date +"random-%H%M%S")
>                 >"$file"
>                 rm -f "$file"
> 	done
> 
> and your add were done in
> 
> 	while sleep 3600
>         do
> 		git add $(for i in *; do test -e $i && echo $i; done)
>                 git commit -m "hourly snapshot"
> 	done
> 
> totally asynchronously without coordinating with what the primary activity
> is doing, your "test -e && echo" can race with a concurrent "rm".

Of course. But in such a situation, you are almost certainly better off doing "git add -A". The external loop is really only useful if you want to add a subset of the files that may or may not exist for some reason. Which is getting pretty specific and crazy.

I haven't checked to see whether "add -A" has a race in discovering files versus actually adding them (I suppose there has to be one...even if we open() immediately and use fstat() and friends for everything else, there still must be a race between getting the name from readdir() and calling open()).

And I can see that as a potentially useful workflow: you are doing regular snapshots of a set of files which are being changed by an outside force. You care about errors, but not files disappearing between readdir() and stat(). The "find" command has an "--ignore-readdir-race" option, which is what you would want.

But that is not "--ignore-errors should also ignore missing files". It is "I want to ignore missing files but not other errors".

Show 8 quoted lines

> Even though I think it is an insane use pattern that we shouldn't bother
> to bend too much backwards to support it, --ignore-errors were added
> primarily for a similar use case (i.e. by the time we try to read it, it
> is either gone or it cannot be read by the user who runs "git add"), so in
> that sense it _might_ make sense to ignore all errors with the option.  If
> we choose to go in that direction, it would also make tons of sense to
> update the documentation of the option to caution more strongly that its
> use is a sign of insanity and is discouraged at the same time.

I tracked down the original thread and it looks like the motivation was to handle repositories with a mixed set of readable and inaccessible (due to permissions) files:

 http://article.gmane.org/gmane.comp.version-control.git/75676

So I think it's not _totally_ insane there. You would do "git add --ignore-errors ." instead of having to manually write the shell loop of accessible items (though personally, I think I would just write the shell loop in that situation).

That's just my two cents on the matter. I'm not personally planning on writing patches for either case.

-Peff