* PERF: Improve database query perf when loading topics for a category.
Instead of left joining the `topics` table against `categories` by filtering with `categories.id`,
we can improve the query plan by filtering against `topics.category_id`
first before joining which helps to reduce the number of rows in the
topics table that has to be joined against the other tables and also
make better use of our existing index.
The following is a before and after of the query plan for a category
with many subcategories.
Before:
```
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------
Limit (cost=1.28..747.09 rows=30 width=12) (actual time=85.502..2453.727 rows=30 loops=1)
-> Nested Loop Left Join (cost=1.28..566518.36 rows=22788 width=12) (actual time=85.501..2453.722 rows=30 loops=1)
Join Filter: (category_users.category_id = topics.category_id)
Filter: ((topics.category_id = 11) OR (COALESCE(category_users.notification_level, 1) <> 0) OR (tu.notification_level > 1))
-> Nested Loop Left Join (cost=1.00..566001.58 rows=22866 width=20) (actual time=85.494..2453.702 rows=30 loops=1)
Filter: ((COALESCE(tu.notification_level, 1) > 0) AND ((topics.category_id <> 11) OR (topics.pinned_at IS NULL) OR ((t
opics.pinned_at <= tu.cleared_pinned_at) AND (tu.cleared_pinned_at IS NOT NULL))))
Rows Removed by Filter: 1
-> Nested Loop (cost=0.57..528561.75 rows=68606 width=24) (actual time=85.472..2453.562 rows=31 loops=1)
Join Filter: ((topics.category_id = categories.id) AND ((categories.topic_id <> topics.id) OR (categories.id = 1
1)))
Rows Removed by Join Filter: 13938306
-> Index Scan using index_topics_on_bumped_at on topics (cost=0.42..100480.05 rows=715549 width=24) (actual ti
me=0.010..633.015 rows=464623 loops=1)
Filter: ((deleted_at IS NULL) AND ((archetype)::text <> 'private_message'::text))
Rows Removed by Filter: 105321
-> Materialize (cost=0.14..36.04 rows=30 width=8) (actual time=0.000..0.002 rows=30 loops=464623)
-> Index Scan using categories_pkey on categories (cost=0.14..35.89 rows=30 width=8) (actual time=0.006.
.0.040 rows=30 loops=1)
Index Cond: (id = ANY ('{11,53,57,55,54,56,112,94,107,115,116,117,97,95,102,103,101,105,99,114,106,1
13,104,98,100,96,108,109,110,111}'::integer[]))
-> Index Scan using index_topic_users_on_topic_id_and_user_id on topic_users tu (cost=0.43..0.53 rows=1 width=16) (a
ctual time=0.004..0.004 rows=0 loops=31)
Index Cond: ((topic_id = topics.id) AND (user_id = 1103877))
-> Materialize (cost=0.28..2.30 rows=1 width=8) (actual time=0.000..0.000 rows=0 loops=30)
-> Index Scan using index_category_users_on_user_id_and_last_seen_at on category_users (cost=0.28..2.29 rows=1 width
=8) (actual time=0.004..0.004 rows=0 loops=1)
Index Cond: (user_id = 1103877)
Planning Time: 1.359 ms
Execution Time: 2453.765 ms
(23 rows)
```
After:
```
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=1.28..438.55 rows=30 width=12) (actual time=38.297..657.215 rows=30 loops=1)
-> Nested Loop Left Join (cost=1.28..195944.68 rows=13443 width=12) (actual time=38.296..657.211 rows=30 loops=1)
Filter: ((categories.topic_id <> topics.id) OR (topics.category_id = 11))
Rows Removed by Filter: 29
-> Nested Loop Left Join (cost=1.13..193462.59 rows=13443 width=16) (actual time=38.289..657.092 rows=59 loops=1)
Join Filter: (category_users.category_id = topics.category_id)
Filter: ((topics.category_id = 11) OR (COALESCE(category_users.notification_level, 1) <> 0) OR (tu.notification_level > 1))
-> Nested Loop Left Join (cost=0.85..193156.79 rows=13489 width=20) (actual time=38.282..657.059 rows=59 loops=1)
Filter: ((COALESCE(tu.notification_level, 1) > 0) AND ((topics.category_id <> 11) OR (topics.pinned_at IS NULL) OR ((topics.pinned_at <= tu.cleared_pinned_at) AND (tu.cleared_pinned_at IS NOT NULL))))
Rows Removed by Filter: 1
-> Index Scan using index_topics_on_bumped_at on topics (cost=0.42..134521.06 rows=40470 width=24) (actual time=38.267..656.850 rows=60 loops=1)
Filter: ((deleted_at IS NULL) AND ((archetype)::text <> 'private_message'::text) AND (category_id = ANY ('{11,53,57,55,54,56,112,94,107,115,116,117,97,95,102,103,101,105,99,114,106,113,104,98,100,96,108,109,110,111}'::integer[])))
Rows Removed by Filter: 569895
-> Index Scan using index_topic_users_on_topic_id_and_user_id on topic_users tu (cost=0.43..1.43 rows=1 width=16) (actual time=0.003..0.003 rows=0 loops=60)
Index Cond: ((topic_id = topics.id) AND (user_id = 1103877))
-> Materialize (cost=0.28..2.30 rows=1 width=8) (actual time=0.000..0.000 rows=0 loops=59)
-> Index Scan using index_category_users_on_user_id_and_last_seen_at on category_users (cost=0.28..2.29 rows=1 width=8) (actual time=0.004..0.004 rows=0 loops=1)
Index Cond: (user_id = 1103877)
-> Index Scan using categories_pkey on categories (cost=0.14..0.17 rows=1 width=8) (actual time=0.001..0.001 rows=1 loops=59)
Index Cond: (id = topics.category_id)
Planning Time: 1.633 ms
Execution Time: 657.255 ms
(22 rows)
```
* PERF: Optimize index on topics bumped_at.
Replace `index_topics_on_bumped_at` index with a partial index on `Topic#bumped_at` filtered by archetype since there is already another index that covers private topics.
Currently, pinned topics are ordered by the `bumped_at` column. This behavior is not desired because it gives admins no control over the order of pinned topics. This PR makes pinned topics ordered by the `pinned_at` column. A topic that is pinned last appears first in topic lists. If an admin wants an already pinned topic to appear first in the list of pinned topics, they'll have to unpin that topic and pin it again.
Meta topic: https://meta.discourse.org/t/how-do-i-set-the-order-of-pinned-topics/16935/23?u=osama.
When dismissing new topics for the Tracked filter, the dismiss was
limited to 30 topics which is the default per page count for TopicQuery.
This happened even if you specified which topic IDs you were
selectively dismissing. This PR fixes that bug, and also moves
the per_page_count into a DEFAULT_PER_PAGE_COUNT for the TopicQuery
so it can be stubbed in tests.
Also moves the unused stub_const method into the spec helpers
for cases like this; it is much better to handle this in one place
with an ensure. In a follow up PR I will clean up other specs that
do the same thing and make them use stub_const.
I merged this PR in yesterday, finally thinking this was done https://github.com/discourse/discourse/pull/12958 but then a wild performance regression occurred. These are the problem methods:
1aa20bd681/app/serializers/topic_tracking_state_serializer.rb (L13-L21)
Turns out date comparison is super expensive on the backend _as well as_ the frontend.
The fix was to just move the `treat_as_new_topic_start_date` into the SQL query rather than using the slower `UserOption#treat_as_new_topic_start_date` method in ruby. After this change, 1% of the total time is spent with the `created_in_new_period` comparison instead of ~20%.
----
History:
Original PR which had to be reverted **https://github.com/discourse/discourse/pull/12555**. See the description there for what this PR is achieving, plus below.
The issue with the original PR is addressed in 92ef54f402
If you went to the `x unread` link for a tag Chrome would freeze up and possibly crash, or eventually unfreeze after nearly 10 mins. Other routes for unread/new were similarly slow. From profiling the issue was the `sync` function of `topic-tracking-state.js`, which calls down to `isNew` which in turn calls `moment`, a change I had made in the PR above. The time it takes locally with ~1400 topics in the tracking state is 2.3 seconds.
To solve this issue, I have moved these calculations for "created in new period" and "unread not too old" into the tracking state serializer.
When I was looking at the profiler I also noticed this issue which was just compounding the problem. Every time we modify topic tracking state we recalculate the sidebar tracking/everything/tag counts. However this calls `forEachTracked` and `countTags` which can be quite expensive as they go through the whole tracking state (and were also calling the removed moment functions).
I added some logs and this was being called 30 times when navigating to a new /unread route because `sync` is being called from `build-topic-route` (one for each topic loaded due to pagination). So I just added a debounce here and it makes things even faster.
Finally, I changed topic tracking state to use a Map so our counts of the state keys is faster (Maps have .size whereas objects you have to do Object.keys(obj) which is O(n).)
<!-- NOTE: All pull requests should have tests (rspec in Ruby, qunit in JavaScript). If your code does not include test coverage, please include an explanation of why it was omitted. -->
Original PR which had to be reverted **https://github.com/discourse/discourse/pull/12555**. See the description there for what this PR is achieving, plus below.
The issue with the original PR is addressed in 92ef54f402
If you went to the `x unread` link for a tag Chrome would freeze up and possibly crash, or eventually unfreeze after nearly 10 mins. Other routes for unread/new were similarly slow. From profiling the issue was the `sync` function of `topic-tracking-state.js`, which calls down to `isNew` which in turn calls `moment`, a change I had made in the PR above. The time it takes locally with ~1400 topics in the tracking state is 2.3 seconds.
To solve this issue, I have moved these calculations for "created in new period" and "unread not too old" into the tracking state serializer.
When I was looking at the profiler I also noticed this issue which was just compounding the problem. Every time we modify topic tracking state we recalculate the sidebar tracking/everything/tag counts. However this calls `forEachTracked` and `countTags` which can be quite expensive as they go through the whole tracking state (and were also calling the removed moment functions).
I added some logs and this was being called 30 times when navigating to a new /unread route because `sync` is being called from `build-topic-route` (one for each topic loaded due to pagination). So I just added a debounce here and it makes things even faster.
Finally, I changed topic tracking state to use a Map so our counts of the state keys is faster (Maps have .size whereas objects you have to do Object.keys(obj) which is O(n).)
The aim of this PR is to improve the topic tracking state JavaScript code and test coverage so further modifications can be made in plugins and in core. This is focused on making topic tracking state changes easier to respond to with callbacks, and changing it so all state modifications go through a single method instead of modifying `this.state` all over the place. I have also tried to improve documentation, make the code clearer and easier to follow, and make it clear what are public and private methods.
The changes I have made here should not break backwards compatibility, though there is no way to tell for sure if other plugin/theme authors are using tracking state methods that are essentially private methods. Any name changes made in the tracking-state.js code have been reflected in core.
----
We now have a `_trackedTopicLimit` in the tracking state. Previously, if a topic was neither new nor unread it was removed from the tracking state; now it is only removed if we are tracking more than `_trackedTopicLimit` topics (which is set to 4000). This is so plugins/themes adding topics with `TopicTrackingState.register_refine_method` can add topics to track that aren't necessarily new or unread, e.g. for totals counts.
Anywhere where we were doing `tracker.states["t" + data.topic_id] = newObject` has now been changed to flow through central `modifyState` and `modifyStateProp` methods. This is so state objects are not modified until they need to be (e.g. sometimes properties are set based on certain conditions) and also so we can run callback functions when the state is modified.
I added `onStateChange` and `onMessageIncrement` methods to register callbacks that are called when the state is changed and when the message count is incremented, respectively. This was done so we no longer need to do things like `@observes("trackingState.states")` in other Ember classes.
I split up giant functions like `sync` and `establishChannels` into smaller functions for readability and testability, and renamed many small functions to _functionName to designate them as private functions which not be called by consumers of `topicTrackingState`. Public functions are now all documented (well...at least ones that are not immediately obvious).
----
On the backend side, I have changed the MessageBus publish events for TopicTrackingState to send back tags and tag IDs for more channels, and done some extra code cleanup and refactoring. Plugins may override `TopicTrackingState.report` so I have made its footprint as small as possible and externalised the main parts of it into other methods.
This is a try to simplify logic around dismiss new topics to have one solution to work in all places - dismiss all-new, dismiss new in a specific category or even in a specific tag.
Disabling shared drafts used to leave topics in an inconsistent state
where they were not displayed as shared drafts and thus there was no
way of publishing them. Moreover, they were accessible just to users
who have permissions to create shared drafts.
This commit adds another permission check that is used for most
operations and the old can_create_shared_draft? remains used just when
creating a new shared draft.
This is an edge-case of 9fb3629. An admin could set the shared draft category to one where both TL2 and TL3 users have access but only give shared draft access to TL3 users. If something like this happens, we need to make sure that TL2 users won't be able to see them, and they won't be listed on latest.
Before this change, `SharedDrafts` were lazily created when a destination category was selected. We now create it alongside the topic and set the destination to the same shared draft category.
Previously thumbnails were only preloaded for queries using `TopicQuery#default_results`, which meant that requests for PM topic lists would lead to N+1 queries.
This commit moves the preloading into TopicList#load_topics, along with other similar preloads (e.g. plugin custom fields)
The direct call to `ActiveRecord::Associations::Preloader#preload` is necessary because `@topics` can be an array, not an `ActiveRecord::Relation`
When that site setting is enabled, the category counts (new/unread)
include the subcategory definition topics, but the topics aren't included
in the list. This fixes that discrepancy.
This is intended for use by plugins which are building their own topic lists, and want to include PMs alongside regular topics (e.g. discourse-assign). It does not get used directly in core.
Like "default watching" and "default tracking" categories option now the "regular" categories support is added. It will be useful for sites that are muted by default. The user option will be displayed only if `mute_all_categories_by_default` site setting is enabled.
This stops sync of tracking state when list is filtered, in the past this
would cause the tracking state to go off wack.
Additionally this introduces an alias for "filter=tracking", called "f=tracking"
This was done cause the term "filter" is used internally in 2 different ways
the main way is for /unread /new filtering.
Trying to also call a query param "filter" causes enormous amounts of
internal pain, this circumvents the issue.
This adds a special filter to topic lists that will filter to tracked and
watched categories.
To use it you can visit:
`https://sitename/?filter=tracked`
`https://sitename/unread?filter=tracked`
and so on
Note, we do not include explicitly tracked and watched topics **outside** of
the tracked categories and tags.
We can consider a `filter=all_tracked` to cover this edge case.
Before this change:
- first full page load would get category defaults defined un cateory settings
- a navigation to a topic and then back to categories list would reset defaut to the ones defined in discovery/topics
This introduces new APIs for obtaining optimized thumbnails for topics. There are a few building blocks required for this:
- Introduces new `image_upload_id` columns on the `posts` and `topics` table. This replaces the old `image_url` column, which means that thumbnails are now restricted to uploads. Hotlinked thumbnails are no longer possible. In normal use (with pull_hotlinked_images enabled), this has no noticeable impact
- A migration attempts to match existing urls to upload records. If a match cannot be found then the posts will be queued for rebake
- Optimized thumbnails are generated during post_process_cooked. If thumbnails are missing when serializing a topic list, then a sidekiq job is queued
- Topic lists and topics now include a `thumbnails` key, which includes all the available images:
```
"thumbnails": [
{
"max_width": null,
"max_height": null,
"url": "//example.com/original-image.png",
"width": 1380,
"height": 1840
},
{
"max_width": 1024,
"max_height": 1024,
"url": "//example.com/optimized-image.png",
"width": 768,
"height": 1024
}
]
```
- Themes can request additional thumbnail sizes by using a modifier in their `about.json` file:
```
"modifiers": {
"topic_thumbnail_sizes": [
[200, 200],
[800, 800]
],
...
```
Remember that these are generated asynchronously, so your theme should include logic to fallback to other available thumbnails if your requested size has not yet been generated
- Two new raw plugin outlets are introduced, to improve the customisability of the topic list. `topic-list-before-columns` and `topic-list-before-link`
We have the `# frozen_string_literal: true` comment on all our
files. This means all string literals are frozen. There is no need
to call #freeze on any literals.
For files with `# frozen_string_literal: true`
```
puts %w{a b}[0].frozen?
=> true
puts "hi".frozen?
=> true
puts "a #{1} b".frozen?
=> true
puts ("a " + "b").frozen?
=> false
puts (-("a " + "b")).frozen?
=> true
```
For more details see: https://samsaffron.com/archive/2018/02/16/reducing-string-duplication-in-ruby
In some cases CTE caused pathologically bad query plans.
This optimises it so query runs by itself and caches for lifetime
of the topic query object.
This lightweight caching is done cause topic query will often
execute two queries (one for pinned and one for non pinned)
This feature adds the ability to define synonyms for tags, and the ability to merge one tag into another while keeping it as a synonym. For example, tags named "js" and "java-script" can be synonyms of "javascript". When searching and creating topics using synonyms, they will be mapped to the base tag.
Along with this change is a new UI found on each tag's page (for example, `/tags/javascript`) where more information about the tag can be shown. It will list the synonyms, which categories it's restricted to (if any), and which tag groups it belongs to (if tag group names are public on the `/tags` page by enabling the "tags listed by group" setting). Staff users will be able to manage tags in this UI, merge tags, and add/remove synonyms.
Instead of enabling `suppress_from_latest` setting on many categories now we can enable `mute_all_categories_by_default` site setting. Then users should opt-in to categories for them to appear in the latest and categories pages.
Also:
Move includes call higher which makes it possible to run all of the
intermediate queries for easier debugging.
Add tests for TagsController with categories in the path.
If we are searching for categories by their slugs, it doesn't make sense
to include subcategories since a slug, by itself, does not necessarily
uniquely identify a subcategory.
Similarly, the empty string as a slug is not a good category identifier.
Doing .pluck(:column).first is a very common pattern in Discourse and in
most cases, a limit cause isn't being added. Instead of adding a limit
clause to all these callsites, this commit adds two new methods to
ActiveRecord::Relation:
pluck_first, equivalent to limit(1).pluck(*columns).first
and pluck_first! which, like other finder methods, raises an exception
when no record is found
Zeitwerk simplifies working with dependencies in dev and makes it easier reloading class chains.
We no longer need to use Rails "require_dependency" anywhere and instead can just use standard
Ruby patterns to require files.
This is a far reaching change and we expect some followups here.
In some very specific cases (large sites) shared drafts can introduce a
performance hit due to the mechanism used to filter out topics
This avoids the entire process when shared drafts are not enabled
* Revert "Revert "FEATURE: Publish read state on group messages. (#7989) [Undo revert] (#8024)""
This reverts commit 36425eb9f0.
* Fix: Show who read only if the attribute is enabled
* PERF: Precalculate the last post readed by a group member
* Use book-reader icon instear of far-eye
* FIX: update topic groups correctly
* DEV: Tidy up read indicator update on write
* Reenable: "FEATURE: Publish read state on group messages. (#7989)"
This reverts commit 67f5cc1ce8.
* FIX: Read indicator only appears when the group setting is enabled
* Enable or disable read state based on group attribute
* When read state needs to be published, the minimum unread count is calculated in the topic query. This way, we can know if someone reads the last post
* The option can be enabled/disabled from the UI
* The read indicator will live-updated using message bus
* Show read indicator on every post
* The read indicator now shows read count and can be expanded to see user avatars
* Read count gets updated everytime someone reads a message
* Simplify topic-list read indicator logic
* Unsubscribe from message bus on willDestroyElement, removed unnecesarry values from post-menu, and added a comment to explain where does minimum_unread_count comes from
During profiling looking up topic users popped up as a hot path, this
change more than halved the amount of work it does
It reduces object allocations and method calls and avoids repeate translation
of common terms
This reduces chances of errors where consumers of strings mutate inputs
and reduces memory usage of the app.
Test suite passes now, but there may be some stuff left, so we will run
a few sites on a branch prior to merging
We had quite a few cases in core where inputs are being mutated as a side
effect of calling a method.
This handles all the cases where specs caught this.
Mutating inputs makes code harder to reason about. Eg:
```
frog = "frog"
jump(frog)
puts frog
"fly" # ?????
```
This commit is part of a followup commit that adds # frozen_string_literal
to all our specs.
This new site setting determines the maximum age of unread topics in
suggested. By default if you have any unread topics older than 90 days
they will be omitted from suggested.
This change was added for 2 reasons:
1. A performance safeguard, some users tend to collect a huge amount of
read state so it becomes super expensive to find unread
2. People who collect a large amount of unread are much more interested in
recent unread topics vs ancient unread topics, this makes suggested more
relevant
Also, this is a minor speed up for tests cause 3 expensive tests became 1.
Previously we tried using a sub-query which has some terrible performance
implications, by adding 1 extra query we can eliminate the PG issue
A join to user_stats also is prone to the same slowdown
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.
Historically due to https://meta.discourse.org/t/why-is-discourse-so-slow-on-android/8823
we decreased page sizes of both home page and topic page on android by half.
This was done on the server side and as a side effect and caused page sizes on android
to mismatch between Android and non Android.
Unfortunately about a year ago googlebot started pretending it is Android,
this cause Google to start indexing pages as what android would see. So
it saw double the amount of pages in the index as what exists on desktop.
This in turn caused double the amount of indexing work and a large amount
of broken links on long topics.
This fix removes all special behavior which is no longer needed due to
other performance work in Discourse including raw handlebars on home page
and virtual dom on topic pages.
I tested we do not need this on Blu Advance 5.0 it has 1.3 GHZ mediatec mt6580
This phone retails for around $50 USD.
If we decide long term that we want any hacks like this we will shift them
to the client side. It can just hold data in memory without rendering.
Previously the related PMs were last meaning you would have to work through
all unread to see them.
Also amends it so it either asks for related by group OR user not both.
- By default, behaviour is not changed: tags are made lowercase upon creation and edit.
- If force_lowercase_tags is disabled, then mixed case tags are allowed.
- Tags must remain case-insensitively unique. This is enforced by ActiveRecord and Postgres.
- A migration is added to provide a `UNIQUE` index on `lower(name)`. Migration includes a safety to correct any current tags that do not meet the criteria.
- A `where_name` scope is added to `models/tag.rb`, to allow easy case-insensitive lookups. This is used instead of `Tag.where(name: "blah")`.
- URLs remain lowercase. Mixed case URLs are functional, but have the lowercase equivalent as the canonical.
Alan Guo Xiang Tan
Osama Sayegh
Andrei Prigorshnev
Robin Ward
Martin Brennan
Dan Ungureanu
Kane York
Krzysztof Kotlarek
Roman Rizzi
David Taylor
Penar Musaraj
Mark VanLandingham
Daniel Waterworth
Vinoth Kannan
Vinoth Kannan
Sam Saffron
Joffrey JAFFEUX
Neil Lalonde
Gerhard Schlager
Guo Xiang Tan
Maja Komel
Arpit Jalan
Régis Hanol