On Fri, 24 Feb 2006, Johannes Schindelin wrote:
Sorry, but no. Really no. Pipes have several advantages over temporary
files:
- The second program can already work on the data before the first
finishes.
This really is a _huge_ issue in general, although probably not a very
big one in this case.
This is what I talked about when I said "streaming" data. Look at the
difference between
git whatchanged -s drivers/usb
and
git log drivers/usb
in the kernel repo. They give almost the same output, but...
Notice how one starts _immediately_, while the other starts after a few
seconds (or, if you have a slow machine, and an unpacked archive, after
tens of seconds or longer).
And the reason is that "git log" uses "git-rev-list" with a path limiter,
and currently that ends up having to walk basically the whole history in
order to generate a minimal graph.
In contrast, "git-whatchanged" uses "git-diff-tree" to limit the output,
and git-diff-tree doesn't care about "minimal graph" or crud like that: it
just cares about discarding any local commits that aren't interesting. It
doesn't need to worry about updating parent chains etc, so it can do it
all incrementally - and can thus start output as soon as it gets anything
at all.
Now, maybe you think that "a few seconds" isn't a big deal. Sure, it's
actually fast as hell, considering what it is doing, and anybody should be
really really impressed that we can do that at all.
But (a) it _is_ a huge deal. Responsiveness is really important. And
worse: (b) it scales badly with repository size. Creating the whole
data-set before starting to output it really doesn't scale.
Now, I have ways to make "git-rev-list" better. It doesn't really need to
walk the _whole_ history for its path limiting before it can start
outputting stuff: it really _could_ do things more incrementally. However,
it's a real bitch sometimes to work with incremental data when you don't
know everything, so it gets a lot more complicated.
So my point isn't that "git log drivers/usb" will get less and less
responsive over time. I can fix that - eventually. My point is that in
order to make it more responsive, I need to make it less synchronous. More
"streaming".
And that is where a pipe is so much better than a file. It's very
fundamentally a streaming interface.
However, I suspect some of these issues are non-issues for the perl
programs that work with a few entries at a time.
Linus