The unique index on `(user_id, message_id)` that is the
`zerver_usermessage` table is rather specific, and even the PostgreSQL
extended statistics are not enough for it to realize there is a
correlation between the `realm_id` in the message table and the
`user_id` in the usermessage table. This means that adding the
`realm_id` limit when there is a join to `zerver_usermessage` flips
the query plan from a nested loop of unique usermessage index-only
scan, with an index scan of the messages pkey -- to a parallel hash
join of the messages limit with a index scan of just the user_id limit
on usermessages. It thinks this is necessary because it thinks that
the `realm_id` limit may remove a large number of messages from the
usermessage set -- which is totally untrue.
Remove the `realm_id` limit if we have a usermessage join.
Removes the JsonErrorBase and JsonError schemas as all error
responses in the API docs use the CodedErrorBase or CodedError
schemas.
Removes the AddSubscriptionsResponse schema since it's no longer
incorrectly used as a shared schema for error responses, and
instead documents the specific success response properties in the
endpoint.
Adds an InvalidStreamError schema for errors that return a 'msg'
field with the string: "Invalid stream ID". Updates endpoints that
have this error 'str' documented to use the shared schema.
Updates documentation of ResourceNotFoundErrors for unknown draft
and scheduled message IDs to include the 'code' field, have an
HTTP status code of 404 in the documentation, and to follow the
general description format of errors in the API documentation.
This endpoint verifies that the services that Zulip needs to function
are running, and Django can talk to them. It is designed to be used
as a readiness probe[^1] for Zulip, either by Kubernetes, or some other
reverse-proxy load-balancer in front of Zulip. Because of this, it
limits access to only localhost and the IP addresses of configured
reverse proxies.
Tests are limited because we cannot stop running services (which would
impact other concurrent tests) and there would be extremely limited
utility to mocking the very specific methods we're calling to raising
the exceptions that we're looking for.
[^1]: https://kubernetes.io/docs/tasks/configure-pod-container/configure-liveness-readiness-startup-probes/
The `expected` flag was incredibly confusing, as you
couldn't tell from the calling code what you were
actually expecting to happen.
I avoid the context manager idiom in order to force
the callers to create simple helper functions, and
I de-duplicate some code in some places.
I also force the caller to explicitly soft-deactivate
the user with one simple line of code, so that the
person reading the test doesn't have to research
the side effects of the helper. (And I make it
very easy for new authors to follow the practice
going forward.)
This is also somewhat of a prep commit to avoid
the obfuscated use of refresh_from_db.
The get_user function is poorly named, but I don't want to
sweep the entire codebase yet.
It's also nice to have a test wrapper for little experiments
like profiling tests or hunting down calls to refresh_from_db.
It's possible that we would also just change the new wrapper
to more directly call Django. The `get_user` function isn't
used in a ton of real-world places, so we might want the test
code to just bypass the cache.
I add a bunch of cute helper methods to make
the test a bit more readable.
And then I make sure to get clean objects,
which precludes the need for our callback
functions to refresh the user objects.
And finally I make sure that our validation
functions don't cause any round trips (assuming
we have fetched objects using a standard
Zulip helper, which example_user ensures.)
In feature levels 153 and 154, a new value of "partially_completed"
for `result` in a success (HTTP status code 200) was added for two
endpoints that process messages in batches: /api/delete-topic and
/api/mark-all-as-read.
Prior to these changes, `result` was either "success" or "error" for
all responses, which was a useful API invariant to have for clients.
So, here we remove "partially_completed" as a potential value for
"result" in a response. And instead, for the two endpoints noted
above, we return a boolean field "complete" to indicate if the
response successfully deleted/marked as read all the targeted
messages (complete: true) or if only some of the targeted messages
were processed (complete: false).
The "code" field for an error string that was also returned as part
of a partially completed response is removed in these changes as
well.
The web app does not currently use the /api/mark-all-as-read
endpoint, but it does use the /api/delete-topic endpoint, so these
changes update that to check the `complete` boolean instead of the
string value for `result`.
For arrays of objects in return values of API endpoints, any
general description of the objects in the arrays should be
documented in the description of the array. A description at the
level of the items in the array will not be rendered in the API
documentation. Descriptions of each property of the object will
be rendered, but these are specific to the property and not the
object as a whole.
Updates the pms, streams and huddles arrays of objects included
in the unread_msgs object of the register response so that the
descriptions are at the array level in the OpenAPI documentation.
When unread_msgs data was added to the register queue response, see
commit 4f0110e, the `user_ids_string` field in the `huddles` array
of objects with information about unread group direct messages, had
the user IDs in the string sorted numerically.
Documents that these strings include the current users's ID and are
sorted numerically and separated by commas so that the documentation
is clear for client implementations.
This adds support for syncing user role via the newly added "role"
attribute, which can be set to either of
['owner', 'administrator', 'moderator', 'member', 'guest'].
Removes durable=True from the atomic decorator of do_change_user_role,
as django-scim2 runs PATCH operations in an atomic block.
This is a prep commit to separate the single test
'test_stream_send_message_events' into two separate tests named
'test_stream_send_message_events' & test_stream_update_message_events'
to verify the events related to send and update message, respectively.
As a part of introducing two new user settings
* 'automatically_follow_topics_policy'
* 'automatically_unmute_topics_policy'
in the next commit, we will extend 'test_stream_send_message_events'.
This logical separation helps in avoiding a single, super-long test.
This commit removes the stray values, i.e., [1, 2, 3], used
in the tests for desktop_icon_count_display.
We use 'UserProfile.DESKTOP_ICON_COUNT_DISPLAY_CHOICES' instead.
'test_change_user_settings' in 'UserDisplayActionTest' excludes
the notification settings and tests only the display settings.
The code block excluding the notification settings doesn't exclude
'modern_notification_settings'. It only excludes the
'notification_settings_legacy'.
This commit replaces 'notification_settings_legacy' with
'notification_setting_types', which consists of all the
notification settings.
Expands API changelog feature level 134 entry and adds the related
Changes notes to the events documentation for the updates made in
commit f4fcedd: "stream op: create" and "subscription op: peer_add"
events being sent when a private stream is made public.
Those changes were made after the feature level 133 updates, but
before the feature level 134 updates, which is why 134 is the
feature level for the change that is documented for clients.
In commit ada2991f1c, when a user gains access to a stream due to
a role change, in addition to sending "stream op: create" events,
"subscription op: peer_add" events are sent for streams that the
user gains access to due to their role change. Updates the API
changelog entry for feature level 205.
Updates the "subcription op: peer_add" event documentation to be
more accurate in for the general use cases of this event, which
are to provide updated subscriber information for streams that
a user has access to.
Since the cache is flushed when the cutoff or realm changes, the
maximum size of the cache should cap out at the number of streams in
the realm. Raise the max cache size, now that this will not simply
lead to useless cache space for smaller servers.
There is now no longer any reason to have the scheduled_email
enqueuing wait until all of the users' contexts have been generated.
Switch to returning the contexts as an iterator, and send them as we
compute them.
The query plan for fetching recent messages from the arbitrary set of
streams formed by the intersection of 30 random users can be quite
bad, and can descend into a sequential scan on `zerver_recipient`.
Worse, this work of pulling recent messages out is redone if the
stream appears in the next batch of 30 users.
Instead, pull the recent messages for a stream on a one-by-one basis,
but cache them in an in-memory cache. Since digests are enqueued in
30-user batches but still one-realm-at-a-time, work will be saved both
in terms of faster query plans whose results can also be reused across
batches.
This requires that we pull the stream-id to stream-name mapping for
_all_ streams in the realm at once, but that is well-indexed and
unlikely to cause performance issues -- in fact, it may be faster
than pulling a random subset of the streams in the realm.
This is common in cases where the reverse proxy itself is making
health-check requests to the Zulip server; these requests have no
X-Forwarded-* headers, so would normally hit the error case of
"request through the proxy, but no X-Forwarded-Proto header".
Add an additional special-case for when the request's originating IP
address is resolved to be the reverse proxy itself; in these cases,
HTTP requests with no X-Forwarded-Proto are acceptable.
The type annotation for functools.partial uses unchecked Any for all
the function parameters (both early and late). returns.curry.partial
uses a mypy plugin to check the parameters safely.
https://returns.readthedocs.io/en/latest/pages/curry.html
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This is designed to help PostgreSQL have better specificity and
locality in its indexes. Subsequent commits will adjust the code to
make sure that we use these indexes rather than the `realm_id`-less
versions.
We do not add a `realm_id` variation to the full-text index, since
it is a GIN index; multi-column GIN indexes are not terribly
performant, require the `btree_gin` extension for `int` types (which
requires superuser privileges on PostgreSQL 12 and earlier), and
cannot be consistently added concurrently on running instances.
After all indexes have been made, we also run `CREATE STATISTICS` in
order to give PostgreSQL the opportunity to realize that recipient and
sender are highly correlated with message realm, allowing it to
estimate that `(realm_id, recipient_id)` is likely as specific as
matching a given `recipient_id`, instead of as likely as matching
`realm_id` times matching a `recipient_id`. Finally, those statistics
must be filled by `ANALYZE zerver_message`, which is run last.
Matching the topic exactly, as opposed to case-insensitively, is not a
common operation, and one that we want to make difficult to do
accidentally. Inline the single use case of it.
We now have a `realm_id` on Message; use it, rather than having to
check the sender's realm. This is theoretically different for
cross-realm bots, but these changes are all in tests where that does
not apply.
This algorithm existed in multiple places, with different queries.
Since we only access properties in the UserMessage table, we
standardize on the much simpler and faster Index Only Scan, rather
than a merge join.
When searching for links inside a topic name, the question mark (?)
was used to split the topic. If a URL had a query after the URL
(e.g., "?foo=bar"), then the query was trimmed from the URL.
Removing the question mark from `basic_link_splitter` is sufficient
to fix this issue. The `get_web_link_regex` function then removes
the trailing punctuation if any, including literal question marks.
Fixes#26368.
When there was no space right after `/todo` but there was content on a
new line, the message would be rendered plainly, not as a todo widget.
This was because we split on only the space character to then check if
the first token was a valid widget.
Now we split on both spaces and newlines to extract the widget name,
irrespective of whether it is followed by a space or a newline. This
results in the message being rendered as a todo widget as expected.
Rename existing shortened references to demo organizations, like
`is_demo_org` or `demo-org-warning`, that have been used in the
codebase so far and replace them to be like the `models.py`
variable: `Realm.demo_organization_scheduled_deletion_date`.
This REDOS was not exploitable, as its content is only read from
checked-in files; regardless, simplify it to not backtrack. We also
do not actually have any location which use leading or trailing
whitespace, so remove those optional bits.
Our logic for extracting strings from templates did not properly
handle the syntax for code containing whitespace control characters,
resulting in a couple strings from subscribe_to_more_streams.hbs not
being processed.
The Librato webhook requires a mapping (which should be considered
immutable) with a default value. Ruff reports a false-positive due to
the Json wrapper.
Instead of a WildValue, the JSON/Sentry webhook expect the request body to be a
dict.
For the JSON webhook, json.dumps accepts other types of input as well and the
constraint is not necessary, but this serve as a good example of an alternative
use of WebhookPayload to describe a payload that is intended to be parsed from
the entire request body from JSON, into a type other than WildValue.
Transifex has parameters that need to be parsed from JSON and converted
to int. Note that we use Optional[Json[int]] instead of
Json[Optional[int]] to replicate the behavior of json_validator. This
caveat is explained in a new test called test_json_optional.
These webhooks have some URL query params that do not need additional
validation or parsing from JSON. So WebhookPaylaod is not applicable to
these webhooks.
This converts most webhook integration views to use @typed_endpoint instead
of @has_request_variables, rewriting REQ parameters. For these
webhooks, it simply requires switching the decorator, rewriting the
type annotation of payload/message to WebhookPayload[WildValue], and
removing the REQ default that defines the to_wild_value converter.
This function is used by almost all webhooks.
To support it, we use the "api_ignore_parameter" flag so that positional
arguments like topic and body that are not intended to be parsed from
the request can be ignored.
This demonstrates the use of BaseModel to replace a check_dict_only
validator.
We also add support to referring to $defs in the OpenAPI tests. In the
future, we can descend down each object instead of mapping them to dict
for more accurate checks.
This demonstrates some basic use cases of the Json[...] wrapper with
@typed_endpoint.
Along with this change we extend test_openapi so that schema checking
based on function signatures will still work with this new decorator.
Pydantic's TypeAdapter supports dumping the JSON schema of any given type,
which is leveraged here to validate against our own OpenAPI definitions.
Parts of the implementation will be covered in later commits as we
migrate more functions to use @typed_endpoint.
See also:
https://docs.pydantic.dev/latest/api/type_adapter/#pydantic.type_adapter.TypeAdapter.json_schema
For the OpenAPI schema, we preprocess it mostly the same way. For the
parameter types though, we no longer need to use
get_standardized_argument_type to normalize type annotation, because
Pydantic dumps a JSON schema that is compliant with OpenAPI schema
already, which makes it a lot convenient for us to compare the types
with our OpenAPI definitions.
Do note that there are some exceptions where our definitions do not match
the generated one. For example, we use JSON to parse int and bool parameters,
but we don't mark them to use "application/json" in our definitions.
We want to reject ambiguous type annotations that set ApiParamConfig
inside a Union. If a parameter is Optional and has a default of None, we
prefer Annotated[Optional[T], ...] over Optional[Annotated[T, ...]].
This implements a check that detects Optional[Annotated[T, ...]] and
raise an assertion error if ApiParamConfig is in the annotation. It also
checks if the type annotation contains any ApiParamConfig objects that
are ignored, which can happen if the Annotated type is nested inside
another type like List, Union, etc.
Note that because
param: Annotated[Optional[T], ...] = None
and
param: Optional[Annotated[Optional[T], ...]] = None
are equivalent in runtime prior to Python 3.11, there is no way for us
to distinguish the two. So we cannot detect that in runtime.
See also: https://github.com/python/cpython/issues/90353
The goal of typed_endpoint is to replicate most features supported by
has_request_variables, and to improve on top of it. There are some
unresolved issues that we don't plan to work on currently. For example,
typed_endpoint does not support ignored_parameters_supported for 400
responses, and it does not run validators on path-only arguments.
Unlike has_request_variables, typed_endpoint supports error handling by
processing validation errors from Pydantic.
Most features supported by has_request_variables are supported by
typed_endpoint in various ways.
To define a function, use a syntax like this with Annotated if there is
any metadata you want to associate with a parameter, do note that
parameters that are not keyword-only are ignored from the request:
```
@typed_endpoint
def view(
request: HttpRequest,
user_profile: UserProfile,
*,
foo: Annotated[int, ApiParamConfig(path_only=True)],
bar: Json[int],
other: Annotated[
Json[int],
ApiParamConfig(
whence="lorem",
documentation_status=NTENTIONALLY_UNDOCUMENTED
)
] = 10,
) -> HttpResponse:
....
```
There are also some shorthands for the commonly used annotated types,
which are encouraged when applicable for better readability and less
typing:
```
WebhookPayload = Annotated[Json[T], ApiParamConfig(argument_type_is_body=True)]
PathOnly = Annotated[T, ApiParamConfig(path_only=True)]
```
Then the view function above can be rewritten as:
```
@typed_endpoint
def view(
request: HttpRequest,
user_profile: UserProfile,
*,
foo: PathOnly[int],
bar: Json[int],
other: Annotated[
Json[int],
ApiParamConfig(
whence="lorem",
documentation_status=INTENTIONALLY_UNDOCUMENTED
)
] = 10,
) -> HttpResponse:
....
```
There are some intentional restrictions:
- A single parameter cannot have more than one ApiParamConfig
- Path-only parameters cannot have default values
- argument_type_is_body is incompatible with whence
- Arguments of name "request", "user_profile", "args", and "kwargs" and
etc. are ignored by typed_endpoint.
- positional-only arguments are not supported by typed_endpoint. Only
keyword-only parameters are expected to be parsed from the request.
- Pydantic's strict mode is always enabled, because we don't want to
coerce input parsed from JSON into other types unnecessarily.
- Using strict mode all the time also means that we should always use
Json[int] instead of int, because it is only possible for the request
to have data of type str, and a type annotation of int will always
reject such data.
typed_endpoint's handling of ignored_parameters_unsupported is mostly
identical to that of has_request_variables.
_default_manager is the same as objects on most of our models. But
when a model class is stored in a variable, the type system doesn’t
know which model the variable is referring to, so it can’t know that
objects even exists (Django doesn’t add it if the user added a custom
manager of a different name). django-stubs used to incorrectly assume
it exists unconditionally, but it no longer does.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
Commit cf0eb46afc added this to let
Django understand the CREATE INDEX CONCURRENTLY statement that had
been hidden in a RunSQL query in migration 0244. However, migration
0245 explained that same index to Django in a different way by setting
db_index=True. Move that to 0244 where the index is actually created,
using SeparateDatabaseAndState.
Also remove the part of the SQL in 0245 that was mirrored by dummy
state_operations, and replace it with real operations.
Signed-off-by: Anders Kaseorg <anders@zulip.com>
This is important because the "guests" value isn't one that we'd
expect anyone to pick intentionally, and in particular isn't an
available option for the similar/adjacent "email invitations" setting.
Earlier whenever a new invitation is created a event was sent
to only admin users. So, if invites by a non-admins user are changed
the invite panel does not live update.
This commit makes changes to also send event to non-admin
user if invites by them are changed.
This commit rename the existing setting `Who can invite users to this
organization` to `Who can send email invitations to new users` and
also renames all the variables related to this setting that do not
require a change to the API.
This was done for better code readability as a new setting
`Who can create invite links` will be added in future commits.
This commit does the backend changes required for adding a realm
setting based on groups permission model and does the API changes
required for the new setting `Who can create multiuse invite link`.
Alex Vandiver
Mateusz Mandera
David Rosa
Sahil Batra
Lauryn Menard
Steve Howell
Aman Agrawal
Prakhar Pratyush
Prakhar Pratyush
Lalit
Anders Kaseorg
N-Shar-ma
Adrián Oliva
Tim Abbott
Zixuan James Li
Ujjawal Modi