This change makes the cache entries smaller for message
dictionaries. It also ensures we get valid data put into
message dictionaries if, for example, the sender's avatar
changes.
After this change, all of the attributes for a message
sender are only fetched during post-processing with two
exceptions:
* We get sender_id for "free" from the message,
and it's the primary key that we need to figure
out which data to fetch in post-processing.
* We need sender_realm_id to be able to cache topic
links, and a sender's realm id will never change,
so it's not a concern for invalidating cache rows.
All the other attributes are either likely to change (e.g.
sender avatar_version) and/or impact the size of cache
entries more severely than the two small id fields above.
This change should improve our overall system performance
by reducing the amount of memory used by every N message
rows we cache, and typically N will be in the thousands or
so on a large realm.
The other major implication of this change is that when
a user changes their avatar, and then later messages that
the user sent are fetched, all of the fields that go into
computing the avatar url will be pulled from the database,
not from cache.
Message.get_raw_db_rows is moved to MessageDict, since its
implementation details are highly coupled to other methods
in MessageDict.
And then sew_messages_and_reactions comes along for the
ride.
We eventually want to move Reaction.get_raw_db_rows to there
as well.
We now populate the avatar url as part of the post
processing step of building message dictionaries,
so that the avatar url is no longer in cache.
This change makes the cache slimmer, because instead
of caching the avatar url (which often includes a long
hash), we just cache the smaller fields that are used
to compute the url.
Note that this commit still has the problem that we're
essentially computing the avatar url from cached fields
that can be invalid. We will address that a few commits
later.
An immediate benefit of this change is that how we compute
avatar urls (or whether we compute them all) is now decoupled
from caching concerns. We will address this later as
well. (Some clients will be capable of computing their
own gravatar urls, for example.)
We're about to have multiple post-processing stages for building
message dictionaries. Rather than having individual "hydration"
methods remove intermediate values, we just wait until the end.
This decouples the hyrdration steps. The potentional problem
here is that we may have a field like sender_is_mirror_dummy
that isn't part of the final payload, but we need it for
calculating display recipients and avatars. We don't want to
delete it too early from the objects.
The `is_mentioned` flag in message events was buggy. We now
look directly at flags.
We will kill off `is_mentioned` in a subsequent commit.
We also remove some debugging code in the test that was failing
before this fix. The test would only fail when `is_mentioned`
was wrong, which never happened when you ran a single test, and
which would happen randomly when you ran multiple tests.
This removes sender names from the message cache, since
they aren't guaranteed to be valid, and they're inexpensive
to add.
This commit will make the message cache entries smaller
by removing sender___full_name and sender__short_name
fields.
Then we add in the sender fields to the message payloads
by doing a query against the unique sender ids of the
messages we are processing.
This change leads to 2 extra database hops for most of
our message-related codepaths. The reason there are 2 hops
instead of 1 is that we basically re-calculate way too
much data to get a no-markdown dictionary.
Introduce MessageDict.post_process_dicts() will allow us
the ability to do the following:
* use less memory in the cache for repeated data
* prevent cache invalidation
* format data according to different client needs
The first use of this function is pretty inconsequential, but
it sets us up for more consequential changes.
In this commit we defer the MessageDict.hydrate_recipient_info
step until after we pull data out of the cache. This impacts
cache size as follows:
* streams - negligibly bigger
* PMs/huddles - slimmer due to not needing to repeat
sender data like email/full_name
Again, the main point of this change is to start setting up
the infrastructure to do post-processing.
This is a first step to eventually slimming the message cache,
but there are still some moving parts there to be worked through.
The more immediate benefit of extracting this function is that
we can put tests on it. Also, it isolates some functionality
that may go away as our clients gets smarter.
The logic to apply events to page_params['unread_msgs'] was
complicated due to the aggregated data structures that we pass
down to the client.
Now we defer the aggregation logic until after we apply the
events. This leads to some simplifications in that codepath,
as well as some performance enhancements.
The intermediate data structure has sets and dictionaries that
generally are keyed by message_id, so most message-related
updates are O(1) in nature.
Also, by waiting to compute the counts until the end, it's a
bit less messy to try to keep track of increments/decrements.
Instead, we just update the dictionaries and sets during the
event-apply phase.
This change also fixes some corner cases:
* We now respect mutes when updating counts.
* For message updates, instead of bluntly updating
the whole topic bucket, we update individual
message ids.
Unfortunately, this change doesn't seem to address the pesky
test that fails sporadically on Travis, related to mention
updates. It will change the symptom, slightly, though.
We now have two helper functions:
* get_raw_unread_data
* aggregate_unread_data
Separating the concerns is nice. The first function does
all the data collection. The second function should be fast,
and it only re-organizes the data into an aggregated form
that makes the page_params payload smaller and easier for
clients to work with.
For the first function, we try to return data structures
that are easier to manipulate than the end result. This
will allow us to apply events more easily, in a subsequent
commit.
There is no reason for either render_incoming_message() or
render_markdown() to require full UserProfile objects just to
triage alert words.
By only asking for user_ids, we save extra queries in two
callpaths and we make it easier to start using user_ids in
do_send_messages().
This never made sense to be a flag on the UserMessage table, since
it's not per-user state. And in fact it doesn't need to be in a
database at all, since it's easily computed from content anyway.
Fixes#1099.
This is the first part of a larger migration to convert Zulip's
reactions storage to something based on the codepoint, not the emoji
name that the user typed in, so that we don't need to worry about
changes in the names we're using breaking the emoji storage.
This field is convenient for bankruptcy checks. Clients could
calculate it from page_params.unread_msgs before this change, but
it would kind of a painful calculation.
To add count, we had to simplify the mypy annotations, which weren't
really accurate before.
We were exiting this function in certain cases before updating
mentions. This bug was always there, but it was flaky in terms
of database setup whether the tests would fail, so now the
relevant test sends three consecutive messages.
We also avoid putting duplicate message ids in mentions.
We are adding a new list of unread message ids grouped by
conversation to the queue registration result. This will allow
clients to show accurate unread badges without needing to load an
unbound number of historic messages.
Jason started this commit, and then Steve Howell finished it.
We only identify conversations using stream_id/user_id info;
we may need a subsequent version that includes things like
stream names and user emails/names for API clients that don't
have data structures to map ids -> attributes.
Instead of using dict's `get()` method use the subscript syntax
so that we can assert correctly that the reaction row contains
all the fields and if not raise the `KeyError` instead of silently
returning None.
This fixes a major performance issue, where we would fetch
user_profile objects inside a code path that had already bulk-fetched
the necessary user objects.
Like the similar related changes we just made, the fix is to marshall
and pass the data into the avatar library directly.
Also puts them into a processing queue, though the queue processor
does nothing.
Rewritten by tabbott to avoid unnecessary database queries in
do_send_messages.
In some cases here we simplify things by calling avatar_url()
instead of get_avatar_url(), when we have a user_profile record
handy. For other cases we pass in an extra avatar_version
parameter to get_avatar_url(), including from avatar_url().
This changes bugdown to use the realm passed in by the caller (if any)
for rendering, fixing a problem where bots such as the notification
bot would have their messages rendering using the admin realm's
settings, not the settings of the realm their messages are being sent
into.
Also adds a test for the notification bot case.
Fixes#3215.
This moves the realm_filter_key variable, primarily used for clarity,
up from Bugdown into the render_markdown function.
We'll need this for the upcoming commits.
This should substantially improve the clarity of the code, since
inside bugdown, this is only being used as a hash key that happens to
usually be a realm ID, not used as a Realm ID.
This change adds support for displaying inline open graph previews for
links posted into Zulip.
It is designed to interact correctly with message editing.
This adds the new settings.INLINE_URL_EMBED_PREVIEW setting to control
whether this feature is enabled.
By default, this setting is currently disabled, so that we can burn it
in for a bit before it impacts users more broadly.
Eventually, we may want to make this manageable via a (set of?)
per-realm settings. E.g. I can imagine a realm wanting to be able to
enable/disable it for certain URLs.
With reactions and other upcoming features, we'll be adding several
places where we need to check whether a particular user can access a
particular message. It's best to just have a single helper function
for this purpose that we can use everywhere.
The command to render old messages now looks for all messages
not matching the bugdown version, and it no longer directly calls
into model code. We should still be extremely cautious about
using this code.
This pulls message-related code from models.py into a new
module called message.py, and it starts to break some bugdown
dependencies. All the methods here are basically related to
serializing Message objects as dictionaries for caches and
events.
extract_message_dict
stringify_message_dict
message_to_dict
message_to_dict_json
MessageDict.to_dict_uncached
MessageDict.to_dict_uncached_helper
MessageDict.build_dict_from_raw_db_row
MessageDict.build_message_dict
This fix also removes a circular dependency related
to get_avatar_url.
Also, there was kind of a latent bug in Message.need_to_render_content
where it was depending on other calls to Message to import bugdown
and set it globally in the namespace. We really need to just
eliminate the function, since it's so small and used by code that
may be doing very sketchy things, but for now I just fix it. (The
bug would possibly be exposed by moving build_message_dict out to the
library.)
Steve Howell
rht
neiljp (Neil Pilgrim)
Tim Abbott
Aditya Bansal
Steve Howell
Jason Michalski
Vaida Plankyte
Harshit Bansal
Elliott Jin
Christian Hudon
vaibhav
Rishi Gupta
Susan Salituro
Raghav Jajodia
Sampriti Panda
Bojidar Marinov
Robert Hönig
Kracekumar R
Igor Tokarev
Sidhant Bhavnani