Reporting this as requested by Antoine Pitrou: Under certain circumstances in Python 3.2 (r32:88445) it's possible for buffered I/O to lose data before it is written and/or return the wrong results when reading. I tripped over this issue whilst writing an assembler which first writes out a bunch of binary data and then goes back over it in a somewhat arbitrary order patching addresses. The following code demonstrates the issue: [code] START = 0 MID = 1 LENGTH = 4 def test(buffering): f = open("test.bin", "w+b", buffering = buffering) for i in range(LENGTH): f.write(b'\x00') f.seek(MID) f.read(1) f.write(b'\x00') f.seek(MID) f.write(b'\x01') f.seek(START) f.seek(MID) print(f.read(1)) f.close() print("Buffered result: ") test(-1) print("Unbuffered result:") test(0) [end code] Output on both Gentoo and Vista is: Buffered result: b'\x00' Unbuffered result: b'\x01' Expected output is b'\x01' in both cases. The issue is reproducible with larger files provided that the constants are increased ~proportionally (START 0, MID 500, LENGTH 1000 for example). Transposing the buffered/unbuffered tests and/or using different buffer sizes for the buffered test seem have no effect. The lengthy code at this URL also demonstrates the issue: http://pastebin.com/xqrzKr5D The above was produced from this code, which was autogenerated by intercepting my assembler's output.
open("test.bin", "w+b", buffering=-1) creates a BufferedRandom object. In the _pyio module, BufferedRandom overrides the write() method to "undo readahead", whereas the _io module reuses bufferedwriter_write() for bufferedrandom_methods and bufferedwriter_methods. I suppose that the problem is "just" that _io.BufferedRandom.write() doesn't undo readahead.
