[PATCH 1 of 2 V2] log: flush stdout/err for the first log entries

[Pierre-Yves David]

On 09/28/2015 04:17 PM, Matt Mackall wrote:
> On Mon, 2015-09-28 at 00:15 +0900, Yuya Nishihara wrote:
>> On Sat, 26 Sep 2015 15:45:45 -0700, Pierre-Yves David wrote:
>>> On 09/26/2015 01:03 AM, Yuya Nishihara wrote:
>>>> On Fri, 25 Sep 2015 13:08:09 -0500, Matt Mackall wrote:
>>>>>>> diff -r b80b2ee71a08 mercurial/ui.py
>>>>>>> --- a/mercurial/ui.py	Thu Sep 24 00:34:15 2015 -0700
>>>>>>> +++ b/mercurial/ui.py	Thu Sep 24 17:45:42 2015 -0500
>>>>>>> @@ -104,6 +104,7 @@
>>>>>>>             self._ucfg = config.config() # untrusted
>>>>>>>             self._trustusers = set()
>>>>>>>             self._trustgroups = set()
>>>>>>> +        self._outcount = 0
>>>>>>>             self.callhooks = True
>>>>>>>
>>>>>>>             if src:
>>>>>>> @@ -617,6 +618,9 @@
>>>>>>>             else:
>>>>>>>                 for a in args:
>>>>>>>                     self.fout.write(str(a))
>>>>>>> +            if self._outcount < 50:
>>>>>>> +                self.fout.flush()
>>>>>>> +                self._outcount += 1
>>>>>>>
>>>>>>>         def write_err(self, *args, **opts):
>>>>>>>             self._progclear()
>>>>>>>
>>>>>>> ("hg log -k sullivan -T." is a pretty good behavior test here.)
>>>>
>>>> So we want line-buffering for first 50 lines?
>>>>
>>>> I've timed it with line-buffered stdout. It's slightly slower than fully-
>>>> buffered io, but perhaps we don't care it if we use the pager.
>>>
>>> Look slike you got your hand on the way to control the buffer setting. \o/
>>>
>>> We should definitely goes for line (or probably even none) buffering for
>>> the pager pipe (the same as what we do for no-pager).
>>
>> Okay, I'll look for the best way to control the buffering. IIRC, the buffer
>> resides in FILE API of libc as of Python 2. There are several ways to change it.
>>
>>   a) sys.stdout = fdopen(sys.stdout.fileno(), 'wb', 1)  # 1 for line buffering
>>
>>      can't replace all FILE* objects because sys.stdout may be aliased to
>>      different names, such as ui.fout.
>
> I couldn't get this to work for some reason when I tried a few months
> back. But if you can get it working, this is probably the cleanest
> approach.

do we have a bug tracking all the patches who died of disentry in there 
journay to the valley of fixes?

>

-- 
Pierre-Yves David