Now that we've debugged the memory leak, I don't think we need this
anymore.
This reverts commit 1bdc7ee2f72bdebb1cdc94601247834a434614d6.
Conflicts:
puppet/zulip/files/cron.d/rabbitmq-numconsumers
puppet/zulip/files/supervisor/conf.d/zulip.conf
(imported from commit ff87f2aebcbc71013fa7a05aedb24e2dcad82ae6)
Unbundle the push notifications from the missed message queue processors
and handlers. This makes notifications more immediate, and sets things up
for better badge count handling, and possibly per-stream filtering.
(imported from commit 11840301751b0bbcb3a99848ff9868d9023b665b)
This fixes a small memory leak in our queue workers, where we don't
reset the accumulated times contained in our query logging data.
Longer-term, we may want to make something mergable for mainline where
we only store on the connection object the totals; that would be a
fixed amount of emmory per connection and thus not have this problem.
(imported from commit 914fa13acfb576f73c5f35e0f64c2f4d8a56b111)
This command should be run continuously via supervisor. It periodically
checks for new email messages to send, and then sends them. This is for
sending email that you've queued via the Email table, instead of mandrill
(as is the case for our localserver/development deploys).
(imported from commit a2295e97b70a54ba99d145d79333ec76b050b291)
Errors are sent to a queue processor that posts them to staging,
just like the feedback bot.
(imported from commit 4a8d099672a1b3e48a8bc94148d8b53db73d2c64)
We also now separate out the times for the socket overhead, the
request service time, and the queuing delays.
(imported from commit e1683f7f28b968b86ebb701b0ac29b00ac6d67c3)
One quirk here is that the Request object is built in the
message_sender worker, not Tornado. This means that the request time
only counts time taken for the actual sending and does not account
for socket overhead. For this reason, I've left the fake logging in
for now so we can compare the two times.
(imported from commit b0c60a3017527a328cadf11ba68166e59cf23ddf)
The register_json_consumer() function now expects its callback
function to accept a single argument, which is the payload, as
none of the callbacks cared about channel, method, and properties.
This change breaks down as follows:
* A couple test stubs and subclasses were simplified.
* All the consume() and consume_wrapper() functions in
queue_processors.py were simplified.
* Two callbacks via runtornado.py were simplified. One
of the callbacks was socket.respond_send_message, which
had an additional caller, i.e. not register_json_consumer()
calling back to it, and the caller was simplified not
to pass None for the three removed arguments.
(imported from commit 792316e20be619458dd5036745233f37e6ffcf43)
Subclasses of QueueProcessingWorker that don't override start() will
have their consume() functions wrapped by consume_wrapper(), which
will catch exceptions and log data from troublesome events to a log
file.
We need to do a puppet apply to create /var/log/zulip/queue_error.
(imported from commit 3bd7751da5fdef449eeec3f7dd29977df11e2b9c)
The DigestWorker isn't do any batching of events, so it can
be implemented with a simple consume() handler.
(imported from commit 0560eaf1a6e510adf97448fbe8933d9903d30016)
This prevents the ones with external side-effects (like sending real
email) from being accidentally run in dev instances.
(imported from commit 6d9861d721abb29136bfff974de01a9264051436)
New dependency: sockjs-tornado
One known limitation is that we don't clean up sessions for
non-websockets transports. This is a bug in Tornado so I'm going to
look at upgrading us to the latest version:
https://github.com/mrjoes/sockjs-tornado/issues/47
(imported from commit 31cdb7596dd5ee094ab006c31757db17dca8899b)
The queued email gets deleted if the user signs up before it gets sent.
Otherwise, they are reminded in 2 days that they still haven't signed up.
This addressses Trac #1812
(imported from commit c1bdc09c03ac576b08986e56994de72d52fd293b)
Deployment instructions: I think all the queue workers get
restarted automatically, so there is probably nothing special
to do here in the deploy itself, but we will want to monitor
it closely, and the change should make our number of locks go
down.
QueueProcessingWorker.start() now calls consume_and_commit(),
which ensures that we don't hold locks after work actions
by using Django's commit_on_success() decorator.
Obviously, workers that override start() will not call consume_and_commit()
through this code path. SlowQueryWorker calls commit_on_success()
in its start() method now, and I hope to address MissedMessageWorker soon.
(imported from commit f3f38a7f45730eee8f3b5794371ba5b994017676)
Use the commit_on_success() context manager around the call
to internal_send_message() inside of SlowQueryWorker's polling
loop, so that the pending SELECT statement from
get_status_dict_by_realm() gets committed. If we don't do
this, postgres will hold locks on zerver_userprofile, and other
tables, for a long time, which can interfere with migrations.
This is an interim solution until we switch postgres's default
commit behavior. Right now the default transaction isolation
is "read committed," so SELECT statements lead to AccessShareLocks
that do no get closed until the transaction finishes.
(imported from commit f72aeffbbe71a731e327459f15bd7dbebaf9e0b8)
Handled by the queue processor for signups. Added a management command
that accomplishes the same task, in case it's needed for manually added users,
or in case we goof and need to remove queued emails for a given user.
This addresses Trac #1807
(imported from commit 6727b82a07fa6a3ea3d827860c9e60fd0602297a)
Luke Faraone
Zev Benjamin
Tim Abbott
acrefoot
Jessica McKellar
acrefoot
Kevin Mehall
Steve Howell
Leo Franchi
Jessica McKellar