Note, to avoid race conditions we are setting last_unread to 10 minutes ago
if there is nothing unread.
This is safer in case of in progress transactions
we don't want to lose unread for any window of time.
This optimisation avoids large scans joining the topics table with the
topic_users table.
Previously when a user carried a lot of read state we would have to join
the entire read state with the topics table. This operation would slow down
home page and every topic page. The more read state you accumulated the
larger the impact.
The optimisation helps people who clean up unread, however if you carry
unread from years ago it will only have minimal impact.
UserStat has some special logic to keep adding time read if repeat calls
are made in intervals less than 100 seconds. This is called regularly
when we update read timings on a topic.
We only need to cache this key in redis for 100 seconds, however previously
we would keep it forever, 1 key per user. This has potential of bloating
a very large amount of keys for no longer active users in redis.
Introduce new patterns for direct sql that are safe and fast.
MiniSql is not prone to memory bloat that can happen with direct PG usage.
It also has an extremely fast materializer and very a convenient API
- DB.exec(sql, *params) => runs sql returns row count
- DB.query(sql, *params) => runs sql returns usable objects (not a hash)
- DB.query_hash(sql, *params) => runs sql returns an array of hashes
- DB.query_single(sql, *params) => runs sql and returns a flat one dimensional array
- DB.build(sql) => returns a sql builder
See more at: https://github.com/discourse/mini_sql
Figuring out what unread topics a user has is a very expensive
operation over time.
Users can easily accumulate 10s of thousands of tracking state rows
(1 for every topic they ever visit)
When figuring out what a user has that is unread we need to join
the tracking state records to the topic table. This can very quickly
lead to cases where you need to scan through the entire topic table.
This commit optimises it so we always keep track of the "first" date
a user has unread topics. Then we can easily filter out all earlier
topics from the join.
We use pg functions, instead of nested queries here to assist the
planner.
* cut down on storage of the work Topic, 3 times per row (in 2 indexes)
* only store one view per user per topic
* only store one view per ip per topic
Introduced badge triggers, introduced concept of badge that happens due to a post but has the post hidden
Delta badge grant happens once a minute, backed by redis
This is information that is not usually needed when representing a user
and is in a separate table with a has one relationship to avoid querying
it all the time.
