py3: make stdout line-buffered if connected to a TTY
Status messages that are to be shown on the terminal should be written to the
file descriptor before anything further is done, to keep the user updated.
One common way to achieve this is to make stdout line-buffered if it is
connected to a TTY. This is done on Python 2 (except on Windows, where libc,
which the CPython 2 streams depend on, does not properly support this).
Python 3 rolls it own I/O streams. On Python 3, buffered binary streams can't be
set line-buffered. The previous code (added in
227ba1afcb65) incorrectly
assumed that on Python 3, pycompat.stdout (sys.stdout.buffer) is already
line-buffered. However the interpreter initializes it with a block-buffered
stream or an unbuffered stream (when the -u option or the PYTHONUNBUFFERED
environment variable is set), never with a line-buffered stream.
One example where the current behavior is unacceptable is when running
`hg pull https://www.mercurial-scm.org/repo/hg` on Python 3, where the line
"pulling from https://www.mercurial-scm.org/repo/hg" does not appear on the
terminal before the hg process blocks while waiting for the server.
Various approaches to fix this problem are possible, including:
1. Weaken the contract of procutil.stdout to not give any guarantees about
buffering behavior. In this case, users of procutil.stdout need to be
changed to do enough flushes. In particular,
1. either ui must insert enough flushes for ui.write() and friends, or
2. ui.write() and friends get split into flushing and fully buffered
methods, or
3. users of ui.write() and friends must flush explicitly.
2. Make stdout unbuffered.
3. Make stdout line-buffered. Since Python 3 does not natively support that for
binary streams, we must implement it ourselves.
(2.) is problematic because using unbuffered I/O changes the performance
characteristics significantly compared to line-buffered (which is used on
Python 2) and this would be a regression.
(1.2.) and (1.3) are a substantial amount of work. It’s unclear whether the
added complexity would be justified, given that raw performance doesn’t matter
that much when writing to a terminal much faster than the user could read it.
(1.1.) pushes complexity into the ui class instead of separating the concern of
how stdout is buffered. Other users of procutil.stdout would still need to take
care of the flushes.
This patch implements (3.). The general performance considerations are very
similar to (1.1.). The extra method invocation and method forwarding add a
little more overhead if the class is used. In exchange, it doesn’t add overhead
if not used.
For the benchmarks, I compared the previous implementation (incorrect on Python
3), (1.1.), (3.) and (2.). The command was chosen so that the streams were
configured as if they were writing to a TTY, but actually write to a pager,
which is also the default:
HGRCPATH=/dev/null python3 ./hg --cwd ~/vcs/mozilla-central --time --pager yes --config pager.pager='cat > /dev/null' status --all
previous:
time: real 7.880 secs (user 7.290+0.050 sys 0.580+0.170)
time: real 7.830 secs (user 7.220+0.070 sys 0.590+0.140)
time: real 7.800 secs (user 7.210+0.050 sys 0.570+0.170)
(1.1.) using Yuya Nishihara’s patch:
time: real 9.860 secs (user 8.670+0.350 sys 1.160+0.830)
time: real 9.540 secs (user 8.430+0.370 sys 1.100+0.770)
time: real 9.830 secs (user 8.630+0.370 sys 1.180+0.840)
(3.) using this patch:
time: real 9.580 secs (user 8.480+0.350 sys 1.090+0.770)
time: real 9.670 secs (user 8.480+0.330 sys 1.170+0.860)
time: real 9.640 secs (user 8.500+0.350 sys 1.130+0.810)
(2.) using a previous patch by me:
time: real 10.480 secs (user 8.850+0.720 sys 1.590+1.500)
time: real 10.490 secs (user 8.750+0.750 sys 1.710+1.470)
time: real 10.240 secs (user 8.600+0.700 sys 1.590+1.510)
As expected, there’s no difference on Python 2, as exactly the same code paths
are used:
previous:
time: real 6.950 secs (user 5.870+0.330 sys 1.070+0.770)
time: real 7.040 secs (user 6.040+0.360 sys 0.980+0.750)
time: real 7.070 secs (user 5.950+0.360 sys 1.100+0.760)
this patch:
time: real 7.010 secs (user 5.900+0.390 sys 1.070+0.730)
time: real 7.000 secs (user 5.850+0.350 sys 1.120+0.760)
time: real 7.000 secs (user 5.790+0.380 sys 1.170+0.710)
#require no-windows
ATTENTION: logtoprocess runs commands asynchronously. Be sure to append "| cat"
to hg commands, to wait for the output, if you want to test its output.
Otherwise the test will be flaky.
Test if logtoprocess correctly captures command-related log calls.
$ hg init
$ cat > $TESTTMP/foocommand.py << EOF
> from __future__ import absolute_import
> from mercurial import registrar
> cmdtable = {}
> command = registrar.command(cmdtable)
> configtable = {}
> configitem = registrar.configitem(configtable)
> configitem(b'logtoprocess', b'foo',
> default=None,
> )
> @command(b'foobar', [])
> def foo(ui, repo):
> ui.log(b'foo', b'a message: %s\n', b'spam')
> EOF
$ cp $HGRCPATH $HGRCPATH.bak
$ cat >> $HGRCPATH << EOF
> [extensions]
> logtoprocess=
> foocommand=$TESTTMP/foocommand.py
> [logtoprocess]
> command=(echo 'logtoprocess command output:';
> echo "\$EVENT";
> echo "\$MSG1") > $TESTTMP/command.log
> commandfinish=(echo 'logtoprocess commandfinish output:';
> echo "\$EVENT";
> echo "\$MSG1";
> echo "canonical: \$OPT_CANONICAL_COMMAND") > $TESTTMP/commandfinish.log
> foo=(echo 'logtoprocess foo output:';
> echo "\$EVENT";
> echo "\$MSG1") > $TESTTMP/foo.log
> EOF
Running a command triggers both a ui.log('command') and a
ui.log('commandfinish') call. The foo command also uses ui.log.
Use sort to avoid ordering issues between the various processes we spawn:
$ hg fooba
$ sleep 1
$ cat $TESTTMP/command.log | sort
command
fooba
logtoprocess command output:
#if no-chg
$ cat $TESTTMP/commandfinish.log | sort
canonical: foobar
commandfinish
fooba exited 0 after * seconds (glob)
logtoprocess commandfinish output:
$ cat $TESTTMP/foo.log | sort
a message: spam
foo
logtoprocess foo output:
#endif
Confirm that logging blocked time catches stdio properly:
$ cp $HGRCPATH.bak $HGRCPATH
$ cat >> $HGRCPATH << EOF
> [extensions]
> logtoprocess=
> pager=
> [logtoprocess]
> uiblocked=echo "\$EVENT stdio \$OPT_STDIO_BLOCKED ms command \$OPT_COMMAND_DURATION ms" > $TESTTMP/uiblocked.log
> [ui]
> logblockedtimes=True
> EOF
$ hg log
$ sleep 1
$ cat $TESTTMP/uiblocked.log
uiblocked stdio [0-9]+.[0-9]* ms command [0-9]+.[0-9]* ms (re)
Try to confirm that pager wait on logtoprocess:
Add a script that wait on a file to appears for 5 seconds, if it sees it touch
another file or die after 5 seconds. If the scripts is awaited by hg, the
script will die after the timeout before we could touch the file and the
resulting file will not exists. If not, we will touch the file and see it.
$ cat >> fakepager.py <<EOF
> import sys
> printed = False
> for line in sys.stdin:
> sys.stdout.write(line)
> printed = True
> if not printed:
> sys.stdout.write('paged empty output!\n')
> EOF
$ cat > $TESTTMP/wait-output.sh << EOF
> #!/bin/sh
> for i in \`$TESTDIR/seq.py 50\`; do
> if [ -f "$TESTTMP/wait-for-touched" ];
> then
> touch "$TESTTMP/touched";
> break;
> else
> sleep 0.1;
> fi
> done
> EOF
$ chmod +x $TESTTMP/wait-output.sh
$ cat >> $HGRCPATH << EOF
> [extensions]
> logtoprocess=
> pager=
> [pager]
> pager = "$PYTHON" $TESTTMP/fakepager.py
> [logtoprocess]
> commandfinish=$TESTTMP/wait-output.sh
> EOF
$ hg version -q --pager=always
Mercurial Distributed SCM (version *) (glob)
$ touch $TESTTMP/wait-for-touched
$ sleep 0.2
$ test -f $TESTTMP/touched && echo "SUCCESS Pager is not waiting on ltp" || echo "FAIL Pager is waiting on ltp"
SUCCESS Pager is not waiting on ltp