Previously, an active production Zulip server would experience a class
of deadlocks caused by two or more concurrent bulk update operations
on the UserMessage table.
This is because UPDATE ... SET ... WHERE statements that execute in
parallel take row-level UPDATE locks as they get results; since the
query plans may result in getting rows in different orders between two
queries, this can result in deadlocks.
Some databases allow ORDER BY on their UPDATE ... WHERE statements;
PostgreSQL does not. In PostgreSQL, the answer is to do a sub-select
with an ORDER BY ... FOR UPDATE to ensure consistent ordering on row
locks.
We do this all code paths using bitand or bitor as part of bulk
editing message flags, which should ensure that these concurrent
operations obtain row level locks on the table in the same order.
Fixes#19054.
Previous behavior was logging only the uuid if it was provided by the
remote server, but that's insufficient, because the user may actually
have no devices registered with uuis and we (at the bouncer) end up
sending notifications to id-based registrations. Not having that id
logged makes it impossible to figure out what's going on.
Fixes#18017.
In previous commits, the change to the bouncer API was introduced to
support this and then a series of migrations added .uuid to
UserProfiles.
Now the code for self-hosted servers that makes requests
to the bouncer is changed to make use of it.
This is the first step to making the full switch to self-hosted servers
use user uuids, per issue #18017. The old id format is still supported
of course, for backward compatibility.
This commit is separate in order to allow deploying *just* the bouncer
API change to production first.
When removing notifications, we skip the access control on if the user
still can read them -- they should not have a notification of them,
both because they currently cannot read the message, as well as
because they have already done so.
In steady-state, requests to FCM take about a second; however, in
cases where the remote FCM server is unstable, the request has been
observed to block for more than a minute.
As noted in the previous commit, pushes must complete within 30s;
fail fast, and let the retries and exponential backoff handle errors.
The worst-case total time taken with timeouts and errors for an FCM
notification is 19.5s. Unfortunately, `aioapns` does not appear to
have any timeouts, and thus this commit cannot guarantee a total of
fewer than 30s.
This reverts bc15085098 (which provided
not justification for its change) and moves further, down to 2 retries
from the default of 5.
10 retries, with exponential backoff, is equivalent to sleeping 2^11
seconds, or just about 34 minutes (though the code uses a jitter which
may make this up to 51 minutes). This is an unreasonable amount of
time to spend in this codepath -- as only one worker is used, and it
is single-threaded, this could effectively block all missed message
notifications for half an hour or longer.
This is also necessary because messages sent through the push bouncer
are sent synchronously; the sending server uses a 30-second timeout,
set in PushBouncerSession. Having retries which linger longer than
this can cause duplicate messages; the sending server will time out
and re-queue the message in RabbitMQ, while the push bouncer's request
will continue, and may succeed.
Limit to 2 retries (APNS currently uses 3), and results expected max
of 4 seconds of sleep, potentially up to 6. If this fails, there
exists another retry loop above it, at the RabbitMQ layer (either
locally, or via the remote server's queue), which will result in up to
3 additional retries -- all told, the request will me made to FCM up
to 12 times.
Calling `get_apns_context` opens (and caches) an open connection to
the APNs servers. Since `apns_enabled` is called from Django
codepaths, this means that the Django processes hold unnecessary
connections open to the APNs servers.
Switch `apns_enabled` to checking what `get_apns_context` checks when
we're just returning True/False.
aioapns 2.1 removed the loop parameter from the aioapns.APNs
constructor, because Python 3.10 removed the loop parameter from the
asyncio.Lock constructor.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Stop using `access_user_group_by_id` in notifications codepaths, as it
is meant to be used to check for _write_ access, not read
access (which is not limited). In the notification codepaths, there
are no ACLs to apply, and the ID is known-good; just load it
directly. The `for_mention` flag is removed, as it was not used in the
mention codepaths at all, only the notification ones.
This replaces the temporary (and testless) fix in
24b1439e93 with a more permanent
fix.
Instead of checking if the user is a bot just before
sending the notifications, we now just don't enqueue
notifications for bots. This is done by sending a list
of bot IDs to the event_queue code, just like other
lists which are used for creating NotificationData objects.
Credit @andersk for the test code in `test_notification_data.py`.
This makes logging more consistent between FCM and APNs codepaths, and
makes clear which user-ids are for local users, and which are opaque
integers namespaced from some remote zulip server.
Being able to determine how many distinct users are getting push
notifications per remote host is useful, as is the distribution of
device counts. This parallels the log line in
handle_push_notification for push notifications from local realms,
handled via the event queue.
This parallels fe25517295, but for mobile notifications. It also
adds a test, which verifies that such content does not crash either
mobile or email notifications.
aioapns already has a retry loop. By default it retries forever on
ConnectionError and ConnectionClosed, so our own retry loop would
never be reached. Remove our retry loop, and configure aioapns to
retry APNS_MAX_RETRIES times on ConnectionError like the previous
version did. It still retries forever on ConnectionClosed; that’s not
configurable but probably fine.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
de04f0ad67 changed now notifications recipients were calculated, in
a manner that caused them to be sent when they should not have been.
ac70a2d2e1 was supposed to resolve this, but appears to have been
insufficient, as all three of these cases have been observed to still
happen.
Add safety checks immediately before notification, until the
underlying logic error can be sussed out.
We do not allow any user to edit the system user groups (including
renaming, deleting, adding or removing members, etc.) from the
API. These user groups will change only by the code when a new
user is added or role of a user is changed.
This is implemented by rejecting access_user_group_by_id always
except the case when it is use to get the user group for sending
email and push notifications, as we would need to send notifications
to the mentioned user group.
Previously, we checked for the `enable_offline_email_notifications` and
`enable_offline_push_notifications` settings (which determine whether the
user will receive notifications for PMs and mentions) just before sending
notifications. This has a few problem:
1. We do not have access to all the user settings in the notification
handlers (`handle_missedmessage_emails` and `handle_push_notifications`),
and therefore, we cannot correctly determine whether the notification should
be sent. Checks like the following which existed previously, will, for
example, incorrectly not send notifications even when stream email
notifications are enabled-
```
if not receives_offline_email_notifications(user_profile):
return
```
With this commit, we simply do not enqueue notifications if the "offline"
settings are disabled, which fixes that bug.
Additionally, this also fixes a bug with the "online push notifications"
feature, which was, if someone were to:
* turn off notifications for PMs and mentions (`enable_offline_push_notifications`)
* turn on stream push notifications (`enable_stream_push_notifications`)
* turn on "online push" (`enable_online_push_notifications`)
then, they would still receive notifications for PMs when online.
This isn't how the "online push enabled" feature is supposed to work;
it should only act as a wrapper around the other notification settings.
The buggy code was this in `handle_push_notifications`:
```
if not (
receives_offline_push_notifications(user_profile)
or receives_online_push_notifications(user_profile)
):
return
// send notifications
```
This commit removes that code, and extends our `notification_data.py` logic
to cover this case, along with tests.
2. The name for these settings is slightly misleading. They essentially
talk about "what to send notifications for" (PMs and mentions), and not
"when to send notifications" (offline). This commit improves this condition
by restricting the use of this term only to the database field, and using
clearer names everywhere else. This distinction will be important to have
non-confusing code when we implement multiple options for notifications
in the future as dropdown (never/when offline/when offline or online, etc).
3. We should ideally re-check all notification settings just before the
notifications are sent. This is especially important for email notifications,
which may be sent after a long time after the message was sent. We will
in the future add code to thoroughly re-check settings before sending
notifications in a clean manner, but temporarily not re-checking isn't
a terrible scenario either.
This reduces loose strings in the codebase, and allows us to not worry
about the exact naming (`stream_email_enabled` or `stream_emails_enabled`?)
and tense (`mentioned` or `mention`?).
Ideally this new class should have been in `lib/notification_data.py`,
which is our file for things like this. But, the next commit requires
using this data in `models.py`, and importing from `notification_data.py`
to `models.py` causes recursive imports.
Work around Fatal1ty/aioapns#15, by silencing error-level logging from
the aioapns logger. We deal with the results of failed
send_notification calls by examining the `result.description` and
handling them; the extra logging message merely clutters the Sentry
logs.
The code to also notify for wildcard mentions was added in
0ed0bb6828.
But that showed the same text for both the cases. This commit fixes
that.
This is more of change for correctness. The mobile app currently does
not rely on this text for notifications, but constructs the text by
itself from the data in the payload.
This also fixes the "stream_push_notify" case to consistently show
a `#` before the stream name.
The previous string was bold, potentially confusing, and doesn't
explain clearly what's happening. We replace this with a string that's
more or less copied from what we do in email notifications with the
similar setting enabled.
When a user has disabled message content in mobile push notifications,
we send a fixed string (currently "REDACTED") as the content of the
notification. Previously, this string was not tagged for translation;
we fix that here.
Additionally, because mobile push notifications are generated in a
queue worker, they do not have the user's language set by the Django
middleware. Our email notifications solve that problem using
`override_language`; we do the same here.
We choose to do override_language in get_message_payload_apns and
get_message_payload_gcm, rather than the caller, in order to be
consistent with tests.
Tested end-to-end by tabbott by setting a translation for "REDACTED"
manually in German.
Fixes#18713.
django.utils.translation.ugettext is a deprecated alias of
django.utils.translation.gettext as of Django 3.0, and will be removed
in Django 4.0.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This issue isn't something a system administrator needs to take action
on -- it's a likely minor logic bug around organization
administrators moving topics between streams.
As a result, it shouldn't send error emails to administrators.
We include tests for the new implementation to avoid churning the
codebase too much so this can be easily reverted when we are able to
re-enable the feature.
We now do something sensible for spoilers in notifications. A message
like:
```spoiler Luke's father is
Vader. Don't tell anyone else.
```
would be rendered as:
Luke's father is (...)
If the push_notification for the UserMessage is already active,
we don't send any push notification to the user. This may
happen due to race conditions.
Added and fixed test cases affected by this.
do_clear_mobile_push_notifications_for_ids can now be used to
clear push_notification for multiple users at once. This method
loops over users, so no performance optimization is gained.
This makes the system store and track PushDeviceToken objects on
the local Zulip server when using the push notifications bouncer
and includes tests for this.
This is something we need to implement end-to-end encryption for
push notifications. We'll add the encryption key as an additional
property on the local PushDeviceToken object.
It also likely adds some value in the case that a server were to
switch between using the bouncer service and sending notifications
directly, though in practice that's unlikely to happen.
This adds a new function `get_apns_badge_count()` to
fetch count value for a user push notification and
then sends that value with the APNs payload.
Once a message is read from the web app, the count is
decremented accordingly and a push notification with
`event: remove` is sent to the iOS clients.
Fixes#10271.
Fixes#2665.
Regenerated by tabbott with `lint --fix` after a rebase and change in
parameters.
Note from tabbott: In a few cases, this converts technical debt in the
form of unsorted imports into different technical debt in the form of
our largest files having very long, ugly import sequences at the
start. I expect this change will increase pressure for us to split
those files, which isn't a bad thing.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
The test_management_commands use in particular was causing pickling
errors when the test failed, because Python 3 filter returns an
iterator, not a list.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
mock is just a backport of the standard library’s unittest.mock now.
The SAMLAuthBackendTest change is needed because
MagicMock.call_args.args wasn’t introduced until Python
3.8 (https://bugs.python.org/issue21269).
The PROVISION_VERSION bump is skipped because mock is still an
indirect dev requirement via moto.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Generated by `pyupgrade --py3-plus --keep-percent-format` on all our
Python code except `zthumbor` and `zulip-ec2-configure-interfaces`,
followed by manual indentation fixes.
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
We use the plumbing introduced in a previous commit, to now raise
PushNotificationBouncerRetryLaterError in send_to_push_bouncer in case
of issues with talking to the bouncer server. That's a better way of
dealing with the errors than the previous approach of returning a
"failed" boolean, which generally wasn't checked in the code anyway and
did nothing.
The PushNotificationBouncerRetryLaterError exception will be nicely
handled by queue processors to retry sending again, and due to being a
JsonableError, it will also communicate the error to API users.
We add PushNotificationBouncerRetryLaterError as an exception to signal
an error occurred when trying to communicate with the bouncer and it
should be retried. We use JsonableError as the base class, because this
signal will need to work in two roles:
1. When the push notification was being issued by the queue worker
PushNotificationsWorker, it will signal to the worker to requeue the
event and try again later.
2. The exception will also possibly be raised (this will be added in the
next commit) on codepaths coming from a request to an API endpoint (for
example to add a token, to users/me/apns_device_token). In that case,
it'll be needed to provide a good error to the API user - and basing
this exception on JsonableError will allow that.
This includes adding a new endpoint to the push notification bouncer
interface, and code to call it appropriately after resetting a user's
personal API key.
When we add support for a user having multiple API keys, we may need
to add an additional key here to support removing keys associated with
just one client.
Fixes#1727.
With the server down, apply migrations 0245 and 0246. 0246 will remove
the pub_date column, so it's essential that the previous migrations
ran correctly to copy data before running this.
This brings us in line, and also allows us to style these more like
unordered lists, which is visually more appealing.
On the backend, we now use the default list blockprocessor + sane list
extension of python-markdown to get proper list markup; on the
frontend, we mostly return to upstream's code as they have followed
CommonMark on this issue.
Using <ol> here necessarily removes the behaviour of not renumbering
on lists written like 3, 4, 7; hopefully users will be OK with the
change.
Fixes#12822.
Historically, Zulip's implementation of wildcard mentions never
triggered either email or push notifications, instead being limited to
desktop notifications and the "mentions" counter.
We fix this just by plumbing the "wildcard_mentioned" flag through our
system.
Implements much of
https://github.com/zulip/zulip/issues/6040#issuecomment-510157264.
We're also now ready to seriously work on #3750.
This contains email of the user to whom notification is being
send. This has not been used in any past mobile releases, so it is
safe to remove it.
As user_id will be stable for the user, but not email. So it's better to
start consuming `user_id` instead of email on mobile.
When a bunch of messages with active notifications are all read at
once -- e.g. by the user choosing to mark all messages, or all in a
stream, as read, or just scrolling quickly through a PM conversation
-- there can be a large batch of this information to convey. Doing it
in a single GCM/FCM message is better for server congestion, and for
the device's battery.
The corresponding client-side logic is in zulip/zulip-mobile#3343 .
Existing clients today only understand one message ID at a time; so
accommodate them by sending individual GCM/FCM messages up to an
arbitrary threshold, with the rest only as a batch.
Also add an explicit test for this logic. The existing tests
that happen to cause this function to run don't exercise the
last condition, so without a new test `--coverage` complains.
Anders Kaseorg
Christopher Chong
Zixuan James Li
Alex Vandiver
Mateusz Mandera
Abhijeet Prasad Bodas
Tim Abbott
Sahil Batra
Hashir Sarwar
Shelly
Rohitt Vashishtha
Aman Agrawal
Anders Kaseorg
Chris Bobbe
Mateusz Mandera
Tim Abbott
Shikhar Varshney
Thomas Ip
Vishwesh Jainkuniya
Greg Price