On 06 Sep 2016, at 12:06 AM, Stefan Fritsch <[email protected]> wrote:
> in trunk, when having a lot of slow long running transfers, most of them seem > to hog a thread and only a few of them go into async write completion mode. > Compare this to 2.4, where many transfers are doing async write completion > and > only a small number of threads are busy. > > Is this a known issue? I’ve seen this before, but couldn’t reproduce it reliably. If you can catch a long running transfer can you get a stacktrace from it? One possible option is that a bucket is insisting on being consumed in one go and thus refuses to leave the handler hook and enter write completion, the second option is that the bucket ends up in write completion but blocks, and we would need to know why. There are still a number of filters out there that insist on swallowing whole bucket brigades in one go. Any of these filters will prevent write completion from working. This will show up in the stacktrace. Regards, Graham —
