Our logic for stream_has_topics never accounted for
us creating essentially "empty" stream buckets that
don't have topics, but we recently added some code
related to unread counts that violated the original
assumptions of the code.
Now we check deeper into the stream bucket to find
actual topics.
This bug manifested in the left sidebar where users
were seeing streams as recently active just because
some muted topics had unread counts, when in fact
the stream was inactive for practical purposes.
We now call topic_data.add_message() and
topic_data.remove_message() when we get info about
incoming messages. The old way of passing in a boolean
made the calling code hard to read and added unncessary
conditional logic to the codepath.
We also have vague plans to change how we handle
removing topics, since increment/decrement logic is now
kind of fragile, so making the "remove" path more explicit
prepares us to something smarter in the future, like just
figure out when the last topic has been removed by calling
a filter function or something outside of topic_data.js.
Another thing to note here is that the code changed here
in echo.js is dead code, since we've disabled
message editing for locally edited messages. I considered
removing this code in a preparatory commit, but there's
other PR activity related to local echo that I don't want
to conflict with.
One nice aspect of removing process_message() is that
the new topic_data.js module does not refer to the legacy
field "subject" any more, nor do its node tests.
This commit introduces a per-stream topic_history class
inside of topic_data.js to better encapsulate how we store topic
history.
To the callers, nothing changes here. (Some of our non-black-box
node tests change their way of setting up data, though, since the
internal data structures are different.)
The new class has the following improvements:
* We use message_id instead of timestamp as our sorting key.
(We could have done this in a prep commit, but it wouldn't
have made the diff much cleaner here.)
* We use a dictionary instead of a sorted list to store the
data, so that writes are O(1) instead of O(NlogN). Reads
now do sorts, so they're O(NlogN) instead of O(N), but reads
are fairly infrequent. (The main goal here isn't actually
performance, but instead it just simplifies the
implementation.)
* We isolate `topic_history` from the format of the messages.
This prepares us for upcoming changes where updates to the
data structure may come from topic history queries as well
as messages.
* We split out the message-add path from the message-remove
path. This prepares us to eventually get rid of the "count"
mechanism that is kind of fragile and which has to be
bypassed for historical topics.
The var in question is indexed by stream_id, so stream_dict seems
like a good name for it. We want to distinguish per-stream data
structures from the stream-level entry points.
This new module tracks the recent topic names for any given
stream.
The code was pulled over almost verbatim from stream_data.js,
with minor renames to the function names.
We introduced a minor one-line function called stream_has_topics.
Steve Howell
Steve Howell