Skip to content
  • Manuel Jacob's avatar
    py3: make stdout line-buffered if connected to a TTY · f9734b2d59cc
    Manuel Jacob authored
    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)
    f9734b2d59cc