Bug #10586
closed
Python keep client (CollectionWriter) appears to deadlock
Description
Some of the jobs across our cluster (those that are neither stuck due to #10585 nor one of the handful that are still running) appear to be stuck in our python crunch script in one of the calls to arvados.CollectionWriter()
Our crunch script is stuck after printing "...writing output to keep" but before "...validating it", which means it is in one of these three calls: https://github.com/wtsi-hgi/arvados-pipelines/blob/master/crunch_scripts/gatk-haplotypecaller-cram.py#L73-L80
It seems likely that issue 10585 and this one could be due to the same underlying issue, which would be some sort of deadlock in the Python keep client, assuming that arv-mount has some supervisor process that eventually notices things are hung and kills them off (whereas our crunch script doesn't have that).
Updated by Tom Morris about 8 years ago
- Subject changed from python keep client appears to deadlock to Python keep client (CollectionWriter) appears to deadlock
- Target version set to 2016-12-14 sprint
Updated by Tom Clegg about 8 years ago
10586-writer-pool-deadlock
Updated by Peter Amstutz about 8 years ago
I would feel better if this exit path also called self.pending_tries_notification.notify_all() (or maybe wrap the whole method in a try/finally that calls notify_all()) since spurious wakeups are preferrable to spurious deadlocks.
keep.py:529
if self.pending_copies() < 1:
# Drain the queue and then raise Queue.Empty
while True:
self.get_nowait()
self.task_done()
This should be if e is not self.TaskFailed ("is" is the object identity operator in Python).
keep.py:598
except Exception as e:
if e != self.TaskFailed:
_logger.exception("Exception in KeepWriterThread")
This method is supposed to return a 2-tuple, but if service.finished() it will return None, which will raise an iteration error when it tries to unpack into locator, copies (maybe you want this to raise TaskFailed?)
keep.py:606
def do_task(self, service, service_root):
if service.finished():
return
Updated by Tom Clegg about 8 years ago
Peter Amstutz wrote:
I would feel better if this exit path also called
self.pending_tries_notification.notify_all()(or maybe wrap the whole method in a try/finally that callsnotify_all()) since spurious wakeups are preferrable to spurious deadlocks.keep.py:529
[...]
It's impossible for any other thread to be stuck in wait() at this point, because write_success() has already called notify_all() after pending_copies() dropped to zero (or it was always zero, in which case no thread could have even called wait()). But I agree that an extra notify_all() is cheap insurance so I've added that.
This should be
if e is not self.TaskFailed("is" is the object identity operator in Python).keep.py:598
[...]
Fixed.
This method is supposed to return a 2-tuple, but
if service.finished()it will returnNone, which will raise an iteration error when it tries to unpack intolocator, copies(maybe you want this to raiseTaskFailed?)keep.py:606
[...]
Indeed. Fixed this by addressing it back in get_next_task(), so we don't pile on more "last result was an error" outcomes after a service starts reporting finished() (which includes "success"!).
Updated by Peter Amstutz about 8 years ago
10586-writer-pool-deadlock LGTM @ e4664336c420836bf26f423faf2af9316302da93
Updated by Tom Clegg about 8 years ago
- Status changed from Feedback to Resolved
Can't be certain that the reported deadlock is the one that got fixed, but it seems likely enough, so closing.