Creating the QueueProcessingWorker objects when the ThreadedWorker is
created can lead to a race which caused confusing error messages:
1. A thread tries to call `self.worker = get_worker()`
2. This call raises an exception, which is caught by
`log_and_exit_if_exception`
3. `log_and_exit_if_exception` sends our process a SIGUSR1, _but
otherwise swallows the error_.
4. The thread's `.run()` is called, which tries to access
`self.worker`, which was never set, and throws another exception.
5. The process handles the SIGUSR1, restarting.
Move the creation of the worker to when it is started, so the worker
object does not need to be stored, and possibly have a decoupled
failure.
Unhandled exceptions propagating to process_queue were not caught there,
causing improper logging - errors didn't land in errors.log as expected.
Exceptions should be caught and explicitly logged by the process_queue
logger. Exceptions occurring during consuming events are caught and
handled inside the worker's logic - however those that happen while
setting up the worker were not addressed at all, and that's the core bug
we mean to address here.
Furthermore, in multi-threaded mode we want the autoreload mechanism to
be working - which it doesn't without catching the exceptions. The
correct approach is to - again - catch the exception, log it and then
send SIGUSR1 signal to trigger exit and autoreload.
SIGALRM is the simplest way to set a specific maximum duration that
queue workers can take to handle a specific message. This only works
in non-threaded environments, however, as signal handlers are
per-process, not per-thread.
The MAX_CONSUME_SECONDS is set quite high, at 10s -- the longest
average worker consume time is embed_links, which hovers near 1s.
Since just knowing the recent mean does not give much information[1],
it is difficult to know how much variance is expected. As such, we
set the threshold to be such that only events which are significant
outliers will be timed out. This can be tuned downwards as more
statistics are gathered on the runtime of the workers.
The exception to this is DeferredWorker, which deals with quite-long
requests, and thus has no enforceable SLO.
[1] https://www.autodesk.com/research/publications/same-stats-different-graphs
Django 2.2.x is the next LTS release after Django 1.11.x; I expect
we'll be on it for a while, as Django 3.x won't have an LTS release
series out for a while.
Because of upstream API changes in Django, this commit includes
several changes beyond requirements and:
* urls: django.urls.resolvers.RegexURLPattern has been replaced by
django.urls.resolvers.URLPattern; affects OpenAPI code and related
features which re-parse Django's internals.
https://code.djangoproject.com/ticket/28593
* test_runner: Change number to suffix. Django changed the name in this
ticket: https://code.djangoproject.com/ticket/28578
* Delete now-unnecessary SameSite cookie code (it's now the default).
* forms: urlsafe_base64_encode returns string in Django 2.2.
https://docs.djangoproject.com/en/2.2/ref/utils/#django.utils.http.urlsafe_base64_encode
* upload: Django's File.size property replaces _get_size().
https://docs.djangoproject.com/en/2.2/_modules/django/core/files/base/
* process_queue: Migrate to new autoreload API.
* test_messages: Add an extra query caused by .refresh_from_db() losing
the .select_related() on the Realm object.
* session: Sync SessionHostDomainMiddleware with Django 2.2.
There's a lot more we can do to take advantage of the new release;
this is tracked in #11341.
Many changes by Tim Abbott, Umair Waheed, and Mateusz Mandera squashed
are squashed into this commit.
Fixes#10835.
This is a preparatory commit for using isort for sorting all of our
imports, merging changes to files where we can easily review the
changes as something we're happy with.
These are also files with relatively little active development, which
means we don't expect much merge conflict risk from these changes.
Previous cleanups (mostly the removals of Python __future__ imports)
were done in a way that introduced leading newlines. Delete leading
newlines from all files, except static/assets/zulip-emoji/NOTICE,
which is a verbatim copy of the Apache 2.0 license.
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
Using sys.exit in a management command makes it impossible
to unit test the code in question. The correct approach to do the same
thing in Django management commands is to raise CommandError.
Followup of b570c0dafa
- For threaded workers:
Django's autoreloader catches SIGQUIT(3) to reload the program. If
a process being watched by autoreloader exits with status code 3,
reloader will restart the process. To reload, we send SIGUSR1(10)
signal from consumers to a handler in process_queue which then
exits with status code 3.
- For single worker per process:
Catch the SIGUSR1 and quit; supervisorctl will restart the worker
automatically.
Fixes#5512
While running queue processors multithreaded will limit the
performance available to very small systems, it's easy to fix that by
adding more RAM, and previously, Zulip didn't work on such systems at
all, so this is unambiguously an improvement there.
Fixes#32.
Fixes#34.
(Commit message expanded significantly by tabbott.)
This change drops the memory used for Python processes run by Zulip in
development from about 1GB to 300MB on my laptop.
On the front of safety, http://pika.readthedocs.org/en/latest/faq.html
explains "Pika does not have any notion of threading in the code. If
you want to use Pika with threading, make sure you have a Pika
connection per thread, created in that thread. It is not safe to share
one Pika connection across threads.". Since this code only connects
to rabbitmq inside the individual threads, I believe this should be
safe.
Progress towards #32.
Alex Vandiver
Anders Kaseorg
Mateusz Mandera
rht
Tim Abbott
Anders Kaseorg
Vishnu Ks
derAnfaenger
Vishnu Ks
Umair Khan
Rafid Aslam
Steve Howell
Tim Abbott
rahuldeve
Reid Barton
Zev Benjamin