We may revisit this in the future, but similar to is:private, the
current Zulip user experience makes users expect that in the
is:mentioned view, they should really be able to mark messages as
read.
Further, the practice use case for not marking them as read is very
low, since it's rare for someone to have so many mentions that
revisiting the mentions view isn't sufficient to see everything that
needs their attention.
Previously, is_exactly() had already been repalced with can_bucket_by().
This commit removes is_exactly() and replaces its usage in our tests
with can_bucket_by().
For Manage Streams, when we render the subscriptions
template, a significant amount of time is taken
by the "t" helper.
Obviously for the first call, we expect "t" to be
somewhat expensive, but subsuquent calls should be
fast, but i18next seems to have some overhead.
Also, we can save a tiny bit of overhead (marking it
as a safe string) that comes from our helper.
As an aside, are we sure it's ok to mark translations
as safe strings?
To test before and after, use blueslip.timings before
and after this commit. When I tested with about 300
streams, the difference is pretty striking:
without cache: 100ms
with cache: 20ms
This is particularly interesting, since the subscriptions
templates have long strings for things like the SVG-based
checkmarks, but they're not really the bottleneck.
Unfortunately, this doesn't seem to be a huge win
elsewhere. In some places we don't call "t", but of
course those might change in the future and benefit from
the cache. And in other places we have smart widgets
that avoid rendering all N objects at one (e.g. buddy
list and list_render).
So this might be too big a hammer to speed up one
screen (albeit a really slow one). It's possible
that we should simply move the i18n.t step **outside**
of certain templates to avoid doing them in a loop.
We now incorporate people.get_message_people() in our
logic for compose/PM typeaheads. This not only gives
users better results in some cases, but it will also
improve performance for large realms in some cases.
We'll use this in two places coming up, so it's
worth extracting, plus I wanted to add the
fairly lengthy comment here. (Tim, feel free
to edit down the comment as you see fit).
This is relatively unobtrusive, and we don't send
anything to the server.
But any user can now enter blueslip.timings in the
console to see a map of how long things take in
milliseconds. We only record one timing per
event label (i.e. the most recent).
It's pretty easy to test this by just clicking
around. For 300 users/streams most things are
fast except for:
- initialize_everything
- manage streams (render_subscriptions)
Both do lots of nontrivial work, although
"manage streams" is a bit surprising, since
we're only measuring how long to build the
HTML from the templates (whereas the real
time is probably browser rendering costs).
This change sets us up to optimize how we
filter users in the admin user settings.
See #13554 for more context on the user
facing issues.
This fix is basically three related things:
- Add filterer options to list_render.
- Add helper method to people.js.
- Use filterer in settings_users.js.
The filter "callback" was only a "callback" in the
most general sense of the word.
It's just a filter predicate that returns a bool.
This is to prepare for another filtering option,
where the caller can filter the whole list
themselves. I haven't figured out what I will name
the new option yet, but I know I want to make the
two options have specific names.
We are already providing callbacks everywhere, so
it would be nice to eliminate some dead code.
This also speeds things up ever so slightly (no
longer type-checking the option every time through
the loop).
We also split out exports.filter to make unit testing
easier. The function seems kinda silly now, being so
small, but I hope to add another filtering option soon.
It's a bit confusing when you read this code to know
where the original list was created. I'm not a huge
fan of the cache scheme here, but it does seem to
work for live updates.
Zulip has had a small use of WebSockets (specifically, for the code
path of sending messages, via the webapp only) since ~2013. We
originally added this use of WebSockets in the hope that the latency
benefits of doing so would allow us to avoid implementing a markdown
local echo; they were not. Further, HTTP/2 may have eliminated the
latency difference we hoped to exploit by using WebSockets in any
case.
While we’d originally imagined using WebSockets for other endpoints,
there was never a good justification for moving more components to the
WebSockets system.
This WebSockets code path had a lot of downsides/complexity,
including:
* The messy hack involving constructing an emulated request object to
hook into doing Django requests.
* The `message_senders` queue processor system, which increases RAM
needs and must be provisioned independently from the rest of the
server).
* A duplicate check_send_receive_time Nagios test specific to
WebSockets.
* The requirement for users to have their firewalls/NATs allow
WebSocket connections, and a setting to disable them for networks
where WebSockets don’t work.
* Dependencies on the SockJS family of libraries, which has at times
been poorly maintained, and periodically throws random JavaScript
exceptions in our production environments without a deep enough
traceback to effectively investigate.
* A total of about 1600 lines of our code related to the feature.
* Increased load on the Tornado system, especially around a Zulip
server restart, and especially for large installations like
zulipchat.com, resulting in extra delay before messages can be sent
again.
As detailed in
https://github.com/zulip/zulip/pull/12862#issuecomment-536152397, it
appears that removing WebSockets moderately increases the time it
takes for the `send_message` API query to return from the server, but
does not significantly change the time between when a message is sent
and when it is received by clients. We don’t understand the reason
for that change (suggesting the possibility of a measurement error),
and even if it is a real change, we consider that potential small
latency regression to be acceptable.
If we later want WebSockets, we’ll likely want to just use Django
Channels.
Signed-off-by: Anders Kaseorg <anders@zulipchat.com>
Currently, if we change stream we see the immediate saving of stream, but
it is more convenient to have "Save" and "Discard" buttons as we use
everywhere else in the organization setting subsystem.
This is a preliminary commit for further commits where we will be using the
newly created function `save_discard_widget_status_handler` in click
handler for changing the notification stream.
This refactors `discard_property_element_changes` and
`check_property_changed` function to move conditional statements of
properties that need to be handled separately. It's a preliminary commit in
the series of using save/discard widget for notification stream setting.
As the part of making notification stream settings to change using
"save/discard" widget instead of immediate saving, we need to access the
stream id which is being selected at the moment.
(This is another preliminary commit in the direction of having
"save/discard" widget show up rather than saving immediately.)
The code for selecting and processing the stream for both types of
notifications is almost the same, so de-duplicated.
For "New stream notifications" and "New user notifications" it is more
intuitive to just use the new system for showing success/saving status
feedback.
This fixes the error where we pass `user_id` of 'string' type as the
argument instead of 'integer' to `exports.get_person_from_user_id` which
further passes `user_id` to InDict.has() function which accepts integer
argument only.
Extracting the function makes it a bit easier to
test and use in a generic way.
Also, I wanted this to live in stream_data, so that
it's easier to find if we change how we model
subscriber data.
Finally, I use _.every to do the subset check
instead of `_.difference`, since _.difference
is actually N-squared:
_.difference = restArguments(function(array, rest) {
rest = flatten(rest, true, true);
return _.filter(array, function(value){
return !_.contains(rest, value);
});
});
And we don't actually want to build a list only
to check that it's zero vs. nonzero length.
We now do this, which short circuits as soon
as it finds any key that is only in sub1:
return _.every(sub1.subscribers.keys(), (key) => {
return sub2_set.has(key);
});
First, there are no more convoluted signals.
We also simplify the parameter to just the "mentioned"
object corresponding to either a user or a broadcast
mention.
For the user group scenario, this has always been dead
code, which you only realized when you got to the comment
at the bottom. Now we actually do nothing.
And I moved the relevant commment to the
the typeahead code (with new wording).
I also moved the is_silent check to the caller. I don't
feel too strongly about that either way. It's kind of silly
to call a function only to give that function an additional
responsibility to worry about. On the other hand, I see
the logic of that function enforcing everything. I went
with the former for now.
Arguably we should have a warning for silent mentions,
since doing a silent mention of somebody not on a stream
is a good indication of a typo. I do understand the use
case, but the user can always ignore the warning. Anyway,
we have decent test coverage on this.
This isn't really an extraction; it's more giving
a name to an anonymous function and moving it to
higher module scope.
We convert this to an ordinary function call, which
allows us to move it out of intialize().
Since there's just one simple parameter now (linked_stream),
we can avoid some error checking.
We also avoid the comment that describes the function,
since it now has a name.
And then one minor tweak is to do the inexpensive
`invite_only` higher in the function. This will be
a nice speedup when you link to really large public
streams.
The unit tests are also a bit easier to read now--less
setup and more explicit names.
This is easier to read and faster, because it
avoids some unnecessary encoding on the pm-with
part, plus just a lot of extra logic that amounts
to just appending the slug.
Performance for this function is relevant because it is used
for every user every time we rerender the right sidebar.
Previously, if you tried to invite a user whose account had been
deactivated, we didn't provide a clear path forward for reactivating
the users, which was confusing.
We fix this by plumbing through to the frontend the information that
there is an existing user account with that email address in this
organization, but that it's deactivated. For administrators, we
provide a link for how to reactivate the user.
Fixes#8144.
This experimental setting disables sending private messages in Zulip
in a crude way (i.e. users get an error when they try to send one).
It makes no effort to adjust the UI to avoid advertising the idea of
sending private messages.
Fixes#6617.
The sort_recipients helper is used for many different
typeaheads, such as compose PMs, compose mentions,
and some settings-related code.
We now avoid unnecessary sorting steps in cases
where we have plenty of results in the top buckets
(such as users who match on prefix).
This change should not have any user-facing
implications.
This method is a bit complex, but I think it's
worthwhile to force PM autocompletes and mention
autocompletes through the same code path.
We also kill off this method:
typeahead_helper.sort_people_and_user_groups
This fixes some regressions from a recent
commit that might not have been deployed.
9f7be51ce8
Even if that change had been deployed, it
should not have been user-facing, but it
would have spammed us with blueslip errors
every time somebody used the stream/topic
popovers in the left sidebar.
There's no reason any more to have a single function
filter both persons and groups. Instead, we just
filter each cohort with a more direct function.
This is a minor performance speedup (avoiding the
conditional in the loop), but I mostly wanted to
simplify the code.
The sort_people_and_user_groups function's only
value-add over sort_recipients is to split out
groups and users, but the caller already had
them split out, so it was kinda silly to concatenate
them back together.
I doubt this was a dumb decision at the time; I think
it was probably a consequence of how bootstrap's
normal approach is kinda inflexible when you're
using typeahead to pull data from multiple sources.
I wanted to kill off sort_people_and_user_groups
completely, but the mention/silent_mention autocompletes
are still a bit awkward to refactor.
For historical reasons pm_list was handling just
one possible edge case of where is:private was
combined with other search terms, namely the
pm-with operator.
The code was correct in realizing the is:private
was redundant there, but now we handle that
upstream in Filter.fix_operators (see previous
commit).
Now we just look for any is:private term.
This change makes these two functions more alike:
- get_search_result
- get_search_result_legacy
To test the UI modify zerver/views/home.py by
replacing `settings.SEARCH_PILLS_ENABLED` with
`True`. I only did a quick sanity check, since
any bugs with the new system are more likely due
to bitrot than any changes I have made here.
The history is this:
Tim cloned the code (before the smaller
helpers were extracted):
db4f6e278f
In 8b153f6452
Shubham removed get_operator_subset_suggestions but
accidentally left a `concat` statement in that got
misapplied to the previous suggestions:
- suggestions = get_operator_subset_suggestions(operators);
result = result.concat(suggestions);
The error there was carried over in some recent changes,
but this commit fixes that strangeness.
In 73e4f3b3fa
Shubham made this change, which makes sense only for
pills, and this code remains intact.
- if (operators.length > 0) {
- last = operators.slice(-1)[0];
+ if (query_operators.length > 0) {
+ last = query_operators.slice(-1)[0];
+ } else {
+ // If query_operators = [] then last will remain
+ // {operator: '', operand: '', negated: false}; from above.
+ // `last` has not yet been added to operators/query_operators.
+ // The code below adds last to operators/query_operators
+ operators.push(last);
+ query_operators.push(last);
}
Mohit made a couple changes to both old and new.
Anders made a couple non-substantive changes related to
the ES6 migration.
Steve (me) made several structural changes to the code. For
some of them I only changed the legacy code, not the pills
code. I didn't fix Shubham's mistake until this change.
Now the two functions should look similar except in the places
where they are intentionally different. I also added a comment
explaining the get_operator_subset_suggestions difference.
Fixes#13609
There may be a deeper issue that various JavaScript logic expects
every message to have a `.message_content` element, but we definitely
should have the `.rendered_markdown` class on any markdown content.
Fixes#13634.
The reason we use functions here will be clear
in the next change.
This is a "prefactor" commit that doesn't change
any user-facing behavior nor significantly change
performance.
A few reasons to extract it:
- we can shorten lines (and not repeat query
every time)
- we can scope the big block comment explaining
why util.prefix_sort is a strange name
- the name is better (it's an O(N) operation that
mostly partitions)
- we may want to swap it out with a true partition
function that's truly a partition (since the
case checks done by prefix_sort are possibly
either a non-feature or mostly overridden by
the other sorts)
The recip.id || recip.user_id idiom has only been
needed for some old unit tests.
It was previously required as a bad workaround for the
local echo issue fixed in dd1a6a97bd
where we would get `display_recipient` values added in an invalid format.
I want to be able to easily test this without
having to simulate all the jQuery side effects.
This simply preserves the old logic, which seems
to handle one edge case without handling every
possible edge case. The edge cases aren't super
important here, though, since the only thing it affects
is bolding "Private Messages", and when to do that
is somewhat up to personal tastes.
Having said that, we could definitely improve
this code and possibly should move some of this
logic to either narrow_state.js or filter.js.
Instead of doing various ad-hoc calculations of
which PM is "active" and plumbing it through various
functions and then updating it via jQuery instead of
just the template, we now just calculate `is_active`
in `_build_private_messages_list` with a little
helper function.
In 3cfc3ca24b I removed
the feature that limited PM conversations to five or
less (including the active conversation), but I
didn't clean up this parameter. I think lint was
confused by the fact that we did mutate it.
I am wondering if this started out as an experiment
and was never fully polished before the push? Or
maybe I was just careless. Anyway, I don't
think were any symptoms here--it was just dead code
that we didn't need.
This fixes a rebase issue between the int_dict introduction and use
for people.js with the introduce of filter_values on dict.js and use
inside people.js.
Note that we haven't fully swept this for Dict,
since some dicts are keyed by strings. For
example PM counts can have a huddle like
"101,102,103" as a key.
This should be slightly more performant, and we
often call this function N times, such as when
rendering the buddy list.
There's a minor change to pm_list to avoid
an unnecessary computation on huddles that would
otherwise trigger a blueslip warning for the
huddles case.
Once we get past the special check for fake
person objects already having `pm_recipient_count`,
we can rely on the object being a `person`
object with `user_id` set.
When we are pulling data from message.display_recipient
for private messages, the user_id field is always
called 'id', not 'user_id', so we can simplify
some defensive code.
This required lots of manual testing:
- search/navigate user presence
- send PM and mention user
- pay attention to compose fade
- send stream msg and mention user
- open Private Messages in top-left and click
- test unread counts
- invite user who already has account
- search for users in search bar
- check user settings
- User Groups
- Users
- Deactivated Users
- Bots
- create a bot
- mention user groups
- send group PM then click on lower right
- view/edit/create streams
If there are still pieces of code that don't convert
ids to ints, the code should still work but report
blueslip errors.
I try to mostly convert user_ids to ints in the callers,
since often the callers are dealing with small amounts
of data, like user ids from huddles.
We only ever show 3 or 4 people in search suggestions
(possibly w/a couple variations, like pm-with/sender/etc.),
so we can try to search a smaller subset of people
before going through the entire realm.
We use message_store.user_ids() for this, since you
typically want to search messages for people that
have sent messages recently, and we already sort
based on PM conversations.
This should avoid some memory allocations.
We also use build_person_matcher to avoid
repeating the same logic over and over
again to process the query into termlets.
We also remove people.get_all_persons() and
people.person_matches_query().
This may actually be a slowdown for the worst case
scenario, but it sets us up to be able to easily
short circuit the removal of diacritic characters
for users that have pure ascii names.
For example, czo has lots of names like this:
- Tim Abbott
- Steve Howell
Since they're pure ascii, we can do a one-time
check. A subsequent commit will show how we use
this.
This looks like simple code cleanup, but it's more
than that.
The code cleanup here is that we don't have three
callbacks to get a list of typeaheads for bootstrap.
Instead, we just have one function that does all the
main work.
And then the speedup comes from the fact we no longer
need to remove diacritics from the query for every
time through our loop of seeing if a person matches
the query.
It's a bit subtle to see in the diff, but these are
the relevant lines:
const matcher = exports.get_person_or_user_group_matcher(query);
const filtered_results = _.filter(people_and_groups, matcher);
Before this, bootstrap was doing $.grep, and we'd have
to reinitialize the matcher for every person.
If you profile this before and after, you'll see that
remove_diacritics gets called fewer times.
To profile this, you want to loads lots of users into
your DB and try to autocomplete "Extra", as in "Extra1 User".
If you try to autocomplete something else, then my patch
won't really help, and `remove_diacritics` will still
show up as expensive. Because it is that expensive a function.
These had to be done in tandem, since they were
both kinda coupled to the function that is now
called query_matches_name_description.
(This commit slightly negatively impacts PM
lookups, but this is addressed in the subsequent
commit, which makes PMs much faster. The impact
is super minimal--it's just an extra function
dispatch.)
This may seem silly now, since we are returning a function
that still dispatches over all flavors of search for
every item, but subsequent commits will make it obvious
why I'm doing this.
We want to do our own matching of items, rather than
just giving a callback to bootstrap, which does $.grep
on all the items.
Doing our own matching gives us flexibility for future
improvements like custom data structures for searching
through big amounts of data. Even in the short term
we can speed up searches by pulling expensive operations
outside the grep/filter call.
This architecture has been in place for our search
bar since ~2014.
We have ~5 years of proof that we'll probably never
extend Dict with more options.
Breaking the classes into makes both a little faster
(no options to check), and we remove some options
in FoldDict that are never used (from/from_array).
A possible next step is to fine-tune the Dict to use
Map internally.
Note that the TypeScript types for FoldDict are now
more specific (requiring string keys). Of course,
this isn't really enforced until we convert other
modules to TS.
We had a potentially nasty bug where we
weren't guaranteeing that all/stream/everyone
collated in consistent ways inside of
`compare_people_for_relevance`, which can
send certain types of sort algorithms into
an infinite loop. I doubt this ever happened
in practice, but it's obviously worth fixing.
Now we also have a clear tiebreaker between
any two all/everyone/stream mentions, which
is the idx field.
Finally, this should be a bit more efficient.
This name was misleading, since this code is used
in sort_recipients, which happens when you, for
example, autocomplete persons in the "To:" box
when composing (and has nothing to do with
mentioning).
This makes it a bit easier to find common patterns,
plus it sets us up to pull the calls even further
up the stack.
The first rule of dealing with user data is sanitize
at the edges, not deep down in some function that
has many callers. Putting this code so deep down
in the stack means it's more likely to be called in
a loop.
Anders Kaseorg
Tim Abbott
Rohitt Vashishtha
Steve Howell
Pragati Agrawal
Tlazypanda