We already trust ids that are put on our queue
for deferred work. For example, see the code for
"mark_stream_messages_as_read_for_everyone"
We now pass stream_recipient_id when we queue
up work for do_mark_stream_messages_as_read.
This generally saves about 3 queries per
user when we unsubscribe them from a stream.
We get two speedups:
* The query to get existing subscribers only
gets the two fields we need. We no longer
need all the overhead of user_profile
and recipient data being returned in the
query.
* We avoid Django making extra hops to the
database to get user info.
Previously, the transaction.atomic() was not properly scoped to ensure
that RealmAuditLog entries were created in the same transaction,
making it possible for state changes to not be properly recorded in
RealmAuditLog.
When apps like mobile register for "streams", we
will now just use active streams as our baseline,
rather than "occupied" streams.
This means we will send a stream that is active,
even if it happens to have zero occupants. It's
actually pretty rare that a stream has zero occupants,
and it's not exactly clear that we want to exclude
a non-occupied but otherwise active stream from
our list of streams.
It also happens to be fairly expensive to compute
whether a stream is occupied.
This change only affects API clients (including
possibly our mobile app). The main webapp never
used the data from this codepath.
We replace get_peer_user_ids_for_stream_change
with two bulk functions to get peers and/or
subscribers.
Note that we have three codepaths that care about
peers:
subscribing existing users:
we need to tell peers about new subscribers
we need to tell subscribed user about old subscribers
unsubscribing existing users:
we only need to tell peers who unsubscribed
subscribing new user:
we only need to tell peers about the new user
(right now we generate send_event
calls to tell the new user about existing
subscribers, but this is a waste
of effort that we will fix soon)
The two bulk functions are this:
bulk_get_subscriber_peer_info
bulk_get_peers
They have some overlap in the implementation,
but there are some nuanced differences that are
described in the comments.
Looking up peers/subscribers in bulk leads to some
nice optimizations.
We will save some memchached traffic if you are
subscribing to multiple public streams.
We will save a query in the remove-subscriber
case if you are only dealing with private streams.
This will ensure that we always fully execute the database part of
modifying subscription objects. In particular, this should prevent
invariant failures like #16347 where Subscription objects were created
without corresponding RealmAuditLog entries.
Fixes#16347.
We don't need the select_related('user_profile')
optimization any more, because we just keep
track of user info in our own data structures.
In this codepath we are never actually modifying
users; we just occasionally need their ids or
emails.
This can be a pretty substantive improvement if
you are adding a bunch of users to a stream
who each have a bunch of their own subscriptions.
We could also limit the number of full rows in this
query by adding an extra hop to the DB just to
get colors (using values_list), and then only get
full sub info for the streams that we're adding, rather
than getting every single subscription, in full, for each user.
Apart from finding what colors the user has already
used, the only other reason we need all the columns
in Subscription here is to handle streams that
need to be reactivated. Otherwise we could do
only("id", "active", "recipient_id", "user_profile_id")
or similar. Fortunately, Subscription isn't
an overly wide table; it's mostly bool fields.
But by far the biggest thing to avoid is bringing
in all the extra user_profile data.
We have pretty good coverage on query counts here,
so I think this fix is pretty low risk.
This class removes a lot of the annoying tuples
we were passing around.
Also, by including the user everywhere, which
is easily available to us when we make instances
of SubInfo, it sets the stage to remove
select_related('user_profile').
We used to send occupy/vacate events when
either the first person entered a stream
or the last person exited.
It appears that our two main apps have never
looked at these events. Instead, it's
generally the case that clients handle
events related to stream creation/deactivation
and subscribe/unsubscribe.
Note that we removed the apply_events code
related to these events. This doesn't affect
the webapp, because the webapp doesn't care
about the "streams" field in do_events_register.
There is a theoretical situation where a
third party client could be the victim of
a race where the "streams" data includes
a stream where the last subscriber has left.
I suspect in most of those situations it
will be harmless, or possibly even helpful
to the extent that they'll learn about
streams that are in a "quasi" state where
they're activated but not occupied.
We could try to patch apply_event to
detect when subscriptions get added
or removed. Or we could just make the
"streams" piece of do_events_register
not care about occupy/vacate semantics.
I favor the latter, since it might
actually be what users what, and it will
also simplify the code and improve
performance.
The query to get "occupied" streams has been expensive
in the past. I'm not sure how much any recent attempts
to optimize that query have mitigated the issue, but
since we clearly aren't sending this data, there is no
reason to compute it.
Using web_public_guest for anonymous users is confusing since
'guest' is actually a logged-in user compared to
web_public_guest which is not logged-in and has only
read access to messages. So, we rename it to
web_public_visitor.
We no longer do O(N) queries to get existing streams.
This is a somewhat contrived use case--generally, we
are not trying to re-subscribe a user to several
streams. Still, we want to avoid this.
This commit also makes `test_bulk_subscribe_many`
do more work, and the change to the test helped
me discover this bug.
If a user asks to be subscribed to a stream
that they are already subscribed to, then
that stream won't be in new_stream_user_ids,
and we won't need to send an event for it.
This change makes that happen more automatically.
Let
U = number of users to subscribe
S = number of streams to subscribe
We were technically doing N^3 amount of work
when we sent certain events, or to be more
precise, U * S * S amount of work. For each
stream, we were looping through a list of tuples
of size U * S to find the users for the stream.
In practice either U or S is usually 1, so the
performance gains here are probably negligible,
especially since the constant factors here
were just slinging around Python data.
But the code is actually more readable now, so
it's a double win.
We rename needs_new_sub (which sounds like
a boolean!) to new_recipient_ids, and we
calculate it explicitly within the loop, so
that we don't need to worry as much about
subsequent passes through the loop mutating it.
This allows us to also remove recipient_ids,
which in turn lets us remove recipients_map,
albeit with a small tweak for stream_map.
I also introduce the my_subs local, which
I use to more directly populate used_colors,
as well as using it as the loop var.
I think it's important that the callers understand
that bulk_add_subscriptions assumes all streams
are being created within a single realm, so I make
it an explicit parameter.
This may be overkill--I would also be happy if we
just included the assertions from this commit.
This function now does all the work that we used
to do with notify_subscriptions_added happening
inside a loop.
There's a small fine-tuning here, where we only
get recent traffic on streams that we're actually
sending events for.
We now just pass in all_subscribers_by_stream, rather
than a callback.
We also move sub_tuples_by_user closer to the
loop where we call notify_subscriptions_added.
This preserves the alpha layer on GIF images that need to be resized
before being uploaded. Two important changes occur here:
1. The new frame is a *copy* of the original image, which preserves the
GIF info.
2. The disposal method of the original GIF is preserved. This
essentially determines what state each frame of the GIF starts from
when it is drawn; see PIL's docs:
https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html#saving
for more info.
This resolves some but not all of the test cases in #16370.
This low-level interface allows consuming from a queue with timeouts.
This can be used to either consume in batches (with an upper timeout),
or one-at-a-time. This is notably more performant than calling
`.get()` repeatedly (what json_drain_queue does under the hood), which
is "*highly discouraged* as it is *very inefficient*"[1].
Before this change:
```
$ ./manage.py queue_rate --count 10000 --batch
Purging queue...
Enqueue rate: 11158 / sec
Dequeue rate: 3075 / sec
```
After:
```
$ ./manage.py queue_rate --count 10000 --batch
Purging queue...
Enqueue rate: 11511 / sec
Dequeue rate: 19938 / sec
```
[1] https://www.rabbitmq.com/consumers.html#fetching
Despite its name, the `queue_size` method does not return the number
of items in the queue; it returns the number of items that the local
consumer has delivered but unprocessed. These are often, but not
always, the same.
RabbitMQ's queues maintain the queue of unacknowledged messages; when
a consumer connects, it sends to the consumer some number of messages
to handle, known as the "prefetch." This is a performance
optimization, to ensure the consumer code does not need to wait for a
network round-trip before having new data to consume.
The default prefetch is 0, which means that RabbitMQ immediately dumps
all outstanding messages to the consumer, which slowly processes and
acknowledges them. If a second consumer were to connect to the same
queue, they would receive no messages to process, as the first
consumer has already been allocated them. If the first consumer
disconnects or crashes, all prior events sent to it are then made
available for other consumers on the queue.
The consumer does not know the total size of the queue -- merely how
many messages it has been handed.
No change is made to the prefetch here; however, future changes may
wish to limit the prefetch, either for memory-saving, or to allow
multiple consumers to work the same queue.
Rename the method to make clear that it only contains information
about the local queue in the consumer, not the full RabbitMQ queue.
Also include the waiting message count, which is used by the
`consume()` iterator for similar purpose to the pending events list.
We modify access_stream_for_delete_or_update function to return
Subscription object also along with stream. This change will be
helpful in avoiding an extra query to get subscription object in
code for updating subscription role.
For streams in which only full members are allowed to post,
we block guest users from posting there.
Guests users were blocked from posting to admin only streams
already. So now, guest users can only post to
STREAM_POST_POLICY_EVERYONE streams.
This is not a new feature but a bugfix which should have
happened when implementing full member stream policy / guest users.
Replaced ImageOps.fit by ImageOps.pad, in zerver/lib/upload.py, which
returns a sized and padded version of the image, expanded to fill the
requested aspect ratio and size.
Fixes part of #16370.
Currently, drain_queue and json_drain_queue ack every message as it is
pulled off of the queue, until the queue is empty. This means that if
the consumer crashes between pulling a batch of messages off the
queue, and actually processing them, those messages will be
permanently lost. Sending an ACK on every message also results in a
significant amount lot of traffic to rabbitmq, with notable
performance implications.
Send a singular ACK after the processing has completed, by making
`drain_queue` into a contextmanager. Additionally, use the `multiple`
flag to ACK all of the messages at once -- or explicitly NACK the
messages if processing failed. Sending a NACK will re-queue them at
the front of the queue.
Performance of a no-op dequeue before this change:
```
$ ./manage.py queue_rate --count 50000 --batch
Purging queue...
Enqueue rate: 10847 / sec
Dequeue rate: 2479 / sec
```
Performance of a no-op dequeue after this change (a 25% increase):
```
$ ./manage.py queue_rate --count 50000 --batch
Purging queue...
Enqueue rate: 10752 / sec
Dequeue rate: 3079 / sec
```
Part of #16094.
Moved the language selection preference logic from home.py to a new
function in i18n.py to avoid repetition in analytics views and home
views.
This prevents the memcached connection from being shared across
multiple processes, and hopefully addresses unexpected behavior from
cached functions like get_user_profile_by_id invoked inside the worker
processes.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
We call build_message_send_dict from check_message instead of
do_send_messages.
This is a prep commit for adding a new setting for handling
wildcard mentions in large streams.
We extract the loop for building message dict in
do_send_messages in a separate function named
build_message_send_dict.
This is a prep commit for moving the code for building
of message dict in check_message.
There is a bug where we send event for even
those messages which do not have embedded links
as we are using single set 'links_for_embed' to
check whether we have to send event for
embedded links or not.
This commit fixes the bug by adding 'links_for_embed'
in message dict itself and send the event only
if that message has embedded links.
This commit removes the unnecessary comment which was added in
9454683108, when we were using message.get() for keys which
were also passed as args in do_send_messages, but there are no
such keys in the current code.
This commit removes the unnecessary line of code to get
rendered_content from message dict sent by check_message
when it actually does not inlcude 'rendered_content' key.
This line was added in 9454683108, but now we do not send
rendered_content in the message dict as we render the message
in do_send_messages itself.
A later commit alters `authenticate` of EmailAuthBackend to
add a store `needs_to_change_password` variable to session
which is useful to insist users on changing their weak password.
The tests start failing with that change because client.login()
runs `authenticate` without a `request` object. So, this commit
sends a request object with `request.session=self.client.session`
to self.client.login() in tests wherever needed.
We now no longer define any schemas in test_events--all
of them are in event_schema, which helps our tooling
cross-check schemas for openapi and node tests.
It happens that whether you add a reaction or remove
a reaction, we send the exact same fields, just using
a different op code.
This sort of symmetry is actually kind of rare, as
usually "add" events have more fields, and "remove" events
might just send an id of something to remove.
Our openapi schema treats these as two seperate events,
so we are more consistent with it, and it helps our
schema-checking tooling for node fixtures, too.
Note that we now have to exempt the two events from
our openapi checks, due to the is_mirror_dummy field
in the deprecated user block. We can decide how to
handle this later--one possibility is to just add it
as an optional field on the event_schema side.
Note that we use value_type for value instead of
bool, since properties can be non-bool things
like color, which we just don't test now. We
should test them.
We more than compensate for this by checking
the actual value of the value in
check_subscription_update.
There is a legacy format where we send
singular "message_id" instead of plural
"message_ids".
Then there are different fields for "private"
and "stream" message types.
Note that we make the schema for profile_data
slightly more realistic, but it doesn't actually get
exercised by our current tests (apart from
making sure it's a dict), since we don't have
profile data for our test realm.
We also don't have the optional fields for bots,
since our tests don't exercise that, nor
delivery_email.
So we exempt realm_user_add_event from openapi
checks for now.
When we try to match the openapi specs better, we
will probably want to add a few tests to test_events.
Obviously getting good coverage for adding users
would be nice for all these scenarios:
* delivery_email matters
* bots
* realm has profile fields
This is a prep commit for supporting "presence"
events, where the key of the dictionary is some
arbitrary string like "website" but the value
of the dictionary is another dictionary itself
with keys that are more like variable names.
This also forces us to create TupleType.
We exempt this from the openapi check,
since we haven't figured out how to model
tuples in openapi with the same precision
as event_schema (and it may be impossible).
Long term we just want to stop dealing in
tuples, of course.
StringDict is a data type for representing dictionaries where
all keys and values are strings. Add this data type to data_types.py
and edit other files so that this data type is put to use and tested.
(slightly tweaked by @showell to remove a comment and shorten
a var name now that we have a proper data type)
We also make our schema in event_schema reflect this,
which in turn makes us match the already accurate
openapi spec, so we no longer need to exempt four
types of events from our sanity checks.
We might want to rename the tool to something more
general now, since we are really reconciling three
things:
- node fixtures
- event_schema checkers for test_events
- openapi specs
The way we compare python and openapi schemas is
as follows:
- first convert openapi schemas to be build
from DictType, ListType, etc. with from_opeapi
- do a diff on the schemas
Most of the new code is just having the FooType
family of classes serialize themselves with schema().
Defining types with an object hierarchy
of type classes will allow us to build
functionality that was impossible (or
really janky) with the validators.py
approach of composing functions.
Most of the changes to event_schema.py
were automated search/replaces.
This patch doesn't really yet take
advantage of the new FooType classes,
but we will use it soon to audit our
openapi specs.
user_profile will be None for web_public_guests here. Hence, for
settings (of which most be inaccessible by web public guest),
which require a user_profile, we either set an empty value for
them or set them to a default value. This will help render
the frontend or extend support to our clients without breaking
a lot of code.
Tweaked by tabbott to add many comments.
This argument does not define if an endpoint "is a webhook"; it is set
for "/api/v1/messages", which is not really a webhook, but allows
access from webhooks.
If multiple filters match the same string, we run into an infinite
loop of converting string into urls. To fix it, we mark the matched
string as atomic after first conversion.
In ae58ed5a7 we decided to echo back the text, when no Pygments lexer
matching that language was found. When we do so, we must take care to
HTML escape the lang before wrapping it in a data-code-language attribute.
Tweaked by tabbott to make clear the escaping is defensive.
Calling `render()` in a middleware before LocaleMiddleware has run
will pick up the most-recently-set locale. This may be from the
_previous_ request, since the current language is thread-local. This
results in the "Organization does not exist" page occasionally being
in not-English, depending on the preferences of the request which that
thread just finished serving.
Move HostDomainMiddleware below LocaleMiddleware; none of the earlier
middlewares call `render()`, so are safe. This will also allow the
"Organization does not exist" page to be localized based on the user's
browser preferences.
Unfortunately, it also means that the default LocaleMiddleware catches
the 404 from the HostDomainMiddlware and helpfully tries to check if
the failure is because the URL lacks a language component (e.g.
`/en/`) by turning it into a 304 to that new URL. We must subclass
the default LocaleMiddleware to remove this unwanted functionality.
Doing so exposes a two places in tests that relied (directly or
indirectly) upon the redirection: '/confirmation_key'
was redirected to '/en/confirmation_key', since the non-i18n version
did not exist; and requests to `/stats/realm/not_existing_realm/`
incorrectly were expecting a 302, not a 404.
This regression likely came in during f00ff1ef62, since prior to
that, the HostDomainMiddleware ran _after_ the rest of the request had
completed.
When converting fenced code markdown, we add the language (if specified)
in a data-attribute by tweaking the HTML generated. Doing so, allows the
frontend to make use of this attr to display view-in-playground option
for codeblocks.
We use pygments to get the lexer subclass name and use that instead of
directly using the language in the data-attribute. Doing so, helps us
map different language aliases (like `js` and `javascript`) into a common
variable (like `JavaScript`) - and avoids the client from dealing with
multiple tags corresponding to the same language.
The html structure for a message like this:
``` js
..content..
```
would now be:
<div class="codehilite" data-codehilite-language="JavaScript">
<pre>..content..</pre>
</div>
Tests and fixtures amended.
This was a broken abstraction that returned to its caller within
multiple forked processes on exceptions, and encouraged ignoring the
error code (as all of its callers did).
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
Fixes#16284.
Most of the work for this was done when we implemented correct
behavior for guest users, since they treat public streams like private
streams anyway.
The general method involves moving the messages to the new stream with
special care of UserMessage.
We delete UserMessages for subs who are losing access to the message.
For private streams with protected history, we also create UserMessage
elements for users who are not present in the old stream, since that's
important for those users to access the moved messages.
Previously, S3UploadBackend.delete_export_tarball failed to strip the
leading ‘/’ from the export path. This mistake is now caught by Moto
1.3.15. I expect it caused deletion failures in the real S3, although
I haven’t verified this.
We store export_path in the audit log with a leading ‘/’, but the
actual S3 keys do not have a leading ‘/’. Changing either system
would require a migration. So the new convention is that the
variables named ‘export_path’ have a leading ‘/’, while variables
named ‘path_id’ or ‘key’ do not.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
The previous code only worked by accident and hyperlink 20.0.0 breaks
it.
>>> hyperlink.parse("example.com").replace(scheme="https")
DecodedURL(url=URL.from_text('https:example.com'))
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Django treats path("<name>") like re_path(r"(?P<name>[^/]+)") and
path("<path:name>") like re_path(r"(?P<name>.+)").
This is more readable and consistent than the mix of slightly
different regexes we had before, and fixes various bugs:
• The r'apps/(.*)$' regex was missing a start anchor ^, so it
incorrectly matched all URLs that included apps/ as a substring
anywhere.
• The r'accounts/login/(google)/$' regex was missing a start anchor ^,
so it incorrectly matched all URLs that ended with
accounts/login/google/.
• The type annotation of zerver.views.realm_export.delete_realm_export
takes export_id as an int, but it was previously passed as a string.
• The type annotation of zerver.views.users.avatar takes medium as a
bool, but it was previously passed as a string.
• The [0-9A-Za-z]+ pattern for uidb64 was missing the - and _
characters that can validly be part of a base64url encoded
string (although I think the id is actually a decimal integer here,
in which case only 012345ADEIMNOQTUYcgjkwxyz are present in its
base64url encoding).
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Steve Howell
Tim Abbott
Aman Agrawal
Anders Kaseorg
Cody Piersall
Alex Vandiver
sahil839
akshatdalton
Abhijeet Prasad Bodas
Vishnu KS
palash
Dinesh
Clara Dantas
orientor
Wes Galbraith
Amitsinghyadav
Sumanth V Rao
Anders Kaseorg