This enables the following in Discourse AI
```
plugin.register_modifier(:chat_allowed_bot_user_ids) do |user_ids, guardian|
if guardian.user
mentionables = AiPersona.mentionables(user: guardian.user)
allowed_bot_ids = mentionables.map { |mentionable| mentionable[:user_id] }
user_ids.concat(allowed_bot_ids)
end
user_ids
end
```
some bots that are id < 0 need to be discoverable in search otherwise people can not talk to them.
---------
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
Also, remove experimental setting and simply use top_menu for feature detection
This means that when people eventually enable the hot top menu, there will
be topics in it
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
During extensibility when we add post menu buttons we very much want access
to the topic.
The transformer does not include a `topic` attribute due to historical reasons.
Given we are going to move away from transforming long term and need to give
plugins access to topic when they are adding buttons, just add the extra
property
Full post jump "SHIFT+K/J" not working if related topics are displayed
Amended logic so we unconditionally use full post jump on topic pages by
checking for .post-stream
* FEATURE: shift+j and shift+k will scroll entire posts
When scrolling through topics with very long posts we would like to use
`shift+j` and `shift+k` to quickly move between posts.
This allows users to bypass the scroll within post behavior when zooming
through topics with keyboard shortcuts
This overloads the behavior of shift+k and j which can be used to scroll
through sections (new/latest/etc...)
* remove useless tests
These tests are testing nothing, no point carrying them around
When we insert into the hot set we add things with a score of 0
This means that if hot has more than batch size items in it with a score, then the 0s don't get an initial score
This corrects the situation by always ensuring we re-score:
1. batch size high scoring topics
2. (new) batch size recently bumped topics
* Update spec/models/topic_hot_scores_spec.rb
Co-authored-by: Isaac Janzen <50783505+janzenisaac@users.noreply.github.com>
---------
Co-authored-by: Isaac Janzen <50783505+janzenisaac@users.noreply.github.com>
1. Don't show visited line for hot filter, it is in random order
2. Don't count likes on non regular posts (eg: whispers / small actions)
3. Don't count participants in non regular posts
1. Serial likers will just like a bunch of posts on the same topic, this will
heavily inflate hot score. To avoid artificial "heat" generated by one user only count
the first like on the topic within the recent_cutoff range per topic
2. When looking at recent topics prefer "unique likers", defer to total likes on
older topics cause we do not have an easy count for unique likers
3. Stop taking 1 off like_count, it is not needed - platforms like reddit
allow you to like own post so they need to remove it.
Internal links always notify and add internal connections in topics.
This adds a special feature that lets you append `?silent=true` to a link
to have it excluded from:
1. Notifications - users will not be notified for these links
2. Post links below posts in the UI
This is specifically useful for large reports where adding all these connections
just results in noise.
- Decrease gravity, we come in too hot prioritizing too many new topics
- Remove all muted topics / categories and tags from the hot list
- Punish topics with zero likes in algorithm
This introduces a new experimental hot sort ordering.
It attempts to float top conversations by first prioritizing a topics with lots of recent activity (likes and users responding)
The schedule that updates hot topics is disabled unless the hidden site setting: `experimental_hot_topics` is enabled.
You can control "decay" with `hot_topic_gravity` and `recency` with `hot_topics_recent_days`
Data is stored in the new `topic_hot_scores` table and you can check it out on the `/hot` route once
enabled.
---------
Co-authored-by: Penar Musaraj <pmusaraj@gmail.com>
* FIX: never skip push notifications
According to Apple, silent push notifications are automatically punished per:
https://developer.apple.com/videos/play/wwdc2022/10098/?time=814
> As mentioned when I showed you the code on how to request a push
> subscription, you must promise that pushes will be user visible.
> Handling a push event is not an invitation for your JavaScript to
> get silent background runtime. Doing so would violate both a user’s
> trust and a user’s battery life. When handling a push event, you are
> in fact required to post a notification to Notification Center.
> Other browsers all have countermeasures against violating the promise
> to make pushes user visible, and so does Safari.
> In the beta build of macOS Ventura, after three push events where you
> fail to post a notification in a timely manner, your site’s push
> subscription will be revoked. You will need to go through the permission
> workflow again.
The isIdle check was causing certain push notifications to be silent
Additionally, the auto dismissal logic was causing delays which may cause
the device to think the push was a silent one.
By removing this we hope to ensure push notification delivery is more robust
and consistent on iOS.
post action feedback is the mechanism in which we provide visual feedback
to the user when a post action is clicked, in cases where the action is a
background (hidden to user) for example: copying text to the clipboard
Core uses this to share post links, but other plugins (for example: AI) use
this to share post transcripts via the clipboard.
This adds a proper plugin API to consume this functionality
`addPostMenuButton` can provide a builder that specified a function as the action.
This function will be called with an object that has both the current post and a method for showing feedback.
Previously we hand no tests for `include_raw` which some consumers may
depend on.
Specifically, Discourse AI uses it to get raw markdown for a set of posts
on a topic.
Also cleans up tests so they lint with default ruby
In non secure contexts (HTTP vs HTTPS) which many run in development the
`clipboardCopy` method falls back to and an exec hack.
However, callers expect a promise from this method and the fallback just
returns a boolean.
Reverts
- DEV: maxmind license checking failing tests #24534
- UX: Show if MaxMind key is missing on IP lookup #18993
These changes are leading to surprising results, our logs are now filling up with warnings on dev environments
We need the change to be redone
This improves the implementation of #18993
1. Error message displayed to user is clearer
2. open_db will also be called, even if license key is blank, as it was previously
3. This in turn means no need to keep stubbing 'maxmind_license_key'
We were seeing lots of deadlocks deploying this migration. This improves
the situation in 2 ways.
1. ddl transaction is avoided, so we hold locks for far shorter times
2. we operate in chunks of a maximum of 100_000 posts (though it is heavily filtered down)
* improve code so it is clearer
If a user somehow is looking at an old version of the page and attempts
to like a post they already like. Display a more reasonable error message.
Previously we would display:
> You are not permitted to view the requested resource.
New error message is:
> Oops! You already performed this action. Can you try refreshing the page?
Triggering this error condition is very tricky, you need to stop the
message bus. A possible reason for it could be bad network connectivity.
* FIX: min_personal_message_post_length not applying to first post
Due to the way PostCreator is wired, we were not applying min_personal_message_post_length
to the first post.
This meant that admins could not configure it so PMs have different
limits.
The code was already pretending that this works, but had no reliable way
of figuring out if we were dealing with a private message
This extends search so it can have consumers that:
1. Can split off "term" from various advanced filters and orders
2. Can build a relation of either order or filter
It also moves a lot of stuff around in the search class for clarity.
Two new APIs are exposed:
`.apply_filter` to apply all the special filters to a posts/topics relation
`.apply_order` to force a particular order (eg: order:latest)
This can then be used by semantic search in Discourse AI
Previous to this change when both `normalize_emails` and `hide_email_address_taken`
is enabled the expected `account_exists` email was only sent on exact email
matches.
This expands it so it also sends an email to the canonical email owner.
The category feature that automatically closes topics does it silently
This amends it so `rake topics:apply_autoclose` which does retroactive
closing will also do so silently.
Initializing an EmberObject with a null object leads to an exception. This commit stops that from happening, and introduces an acceptance test for adding/removing banner topics via message-bus.
Co-authored-by: David Taylor <david@taylorhq.com>
Under exceptional cases people may need to resize the notification table.
This only happens on forums with a total of more than 2.5 billion notifications.
This rake task can be used to convert all the notification columns to
bigint to make more room.
https://meta.discourse.org/t/markdown-preview-and-result-differ/263878
The result of this markdown had different results in the composer preview and the post. This is solved by updating Loofah to the latest version and using html5 fragments like our user had reported. While the change was only needed in cooked_post_processor.rb for this fix, other areas also had to be updated due to various side effects.
* Revert "Build(deps): Bump message-bus-client from 4.3.2 to 4.3.3 in /app/assets/javascripts (#22194)"
This reverts commit cdcf6cf0dd.
* Revert "Build(deps): Bump message_bus from 4.3.2 to 4.3.3 (#22188)"
This reverts commit c7a9da1f10.
Anonymization is among the most expensive operations we can perform with
extreme potential to impact the database. To mitigate risk we only allow a
single anonymization across the entire cluster concurrently.
This commit introduces support for `cluster_concurrency 1`. When you set that on a Job it will only allow 1 concurrent execution per cluster.
* FEATURE: reduce avatar sizes to 6 from 20
This PR introduces 3 changes:
1. SiteSetting.avatar_sizes, now does what is says on the tin.
previously it would introduce a large number of extra sizes, to allow for
various DPIs. Instead we now trust the admin with the size list.
2. When `avatar_sizes` changes, we ensure consistency and remove resized
avatars that are not longer allowed per site setting. This happens on the
12 hourly job and limited out of the box to 20k cleanups per cycle, given
this may reach out to AWS 20k times to remove things.
3.Our default avatar sizes are now "24|48|72|96|144|288" these sizes were
very specifically picked to limit amount of bluriness introduced by webkit.
Our avatars are already blurry due to 1px border, so this corrects old blur.
This change heavily reduces storage required by forums which simplifies
site moves and more.
Co-authored-by: David Taylor <david@taylorhq.com>
In some cases reverse chronological can be very important.
- Oldest post by sam
- Oldest topic by sam
Prior to these new filters we had no way of searching for them.
Now the 2 new orders `order:oldest` and `order:oldest_topic` can be used
to find oldest topics and posts
* Update spec/lib/search_spec.rb
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
* Update spec/lib/search_spec.rb
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
---------
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
We have been struggling lately finding site settings due to 30 setting limit
This was introduced for performance reasons a while back but is no longer as
needed given that ember is faster.
Additionally searching is hard, so allow people to use fuzzy search against
setting name.
An extensibility point we support server side is setting meta_data
(topic / post custom fields) with the composer payload.
Previous to this change even though we had a lot of setup code we never
actually sent the payload.
This ensures that on create we send meta_data.
We were giving topics with repeated words extra weight in search index.
This meant that it was trivial to stuff words into title to dominate in search
given we search for exact title matches first.
The following tweak means that:
`invite invited invites`
and
`invite some stuff`
Both rank the same for title searching.
Titles are short and punchy, duplicating words should not give special
weight.
Requires a full reindex to take effect.
This amends it so our cached counting reliant specs run in synchronize mode
When running async there are situations where data is left over in the table
after a transactional test. This means that repeat runs of the test suite
fail.
This new modifier can be used by plugins to modify search ordering.
Specifically plugins such as discourse_solved can amend search ordering
so solved topics bump to the top.
Also correct edge case where low and high sort priority categories did not
order correctly when it came to closed/archived
Many blog posts use these to illustrate and images were previously omitted
Additionally strip superfluous HTML and BODY tags from embed HTML.
This was incorrectly returned from server.
Due to the order we were parsing markdown, bbcode [url] elements were not
handled properly.
`[url]https://example.com/path[/url]` was not currectly parsing cause
linkify was detecting the url as: `https://example.com/path[/url]` which is
legit.
To resolve this I swapped url to use a replace rule, and instead re-parsed
the internal payload and injected the tokens in.
This fix is complex cause we support stuff like
`[url][b]test.com[/b][/url]`
So we need to parse the content inside url `[b]test.com[/b]`
Add a modifier that will allow us to tune the results returned by suggested.
At the moment the modifier allows us to toggle including random results.
This was created for the discourse-ai module. It needs to switch off random
results when it returns related topics.
Longer term we can use it to toggle unread/new and other aspects.
This also demonstrates how to test the contract when adding modifiers.
topic.url is a dangerous method to use, it is incompatible with our `sign_in`
helper.
This removes a bunch of places this was erroneously used and helps to avoid
future cargo culting.
Introduces a new API for plugin data modification without class-based extension overhead.
This commit introduces a new API that allows plugins to modify data in cases where they return different data rather than additional data, as is common with filtered_registers in DiscoursePluginRegistry. This API removes the need for defining class-based extension points.
When a plugin registers a modifier, it will automatically be called if the plugin is enabled. The core will then modify the parameter sent to it using the block registered by the plugin:
```ruby
DiscoursePluginRegistry.register_modifier(plugin_instance, :magic_sum_modifier) { |a, b| a + b }
sum = DiscoursePluginRegistry.apply_modifier(:magic_sum_filter, 1, 2)
expect(sum).to eq(3)
```
Key features of these modifiers:
- Operate in a stack (first registered, first called)
- Automatically disabled when the plugin is disabled
- Pass the cumulative result of all block invocations to the caller
This corrects two issues:
1. We were double serializing topic tracking state (as_json calls were not cached)
2. We were inefficiently serializing items by instantiating extra objects
We perform lookups on sidebar section links based on sidebar_section_id
totally ignoring user. This ensures we have an index to work with.
This removes the previous index `links_user_id_section_id_position` which
partially doubled up `idx_unique_sidebar_section_links`
During search indexing we "stuff" the index with additional keywords for
entities that look like domain names.
This allows searches for `cnn` to find URLs for `www.cnn.com`
The search stuffing attempted to keep indexes aligned at the correct positions
by remapping the indexed terms. However under certain edge cases a single
word can stem into 2 different lexemes. If this happened we had an off by
one which caused the entire indexing to fail.
We work around this edge case (and carry incorrect index positions) for cases
like this. It is unlikely to impact search quality at all given index position
makes almost no difference in the search algorithm.
If you set a category to `default_list_filter` none. Information
was not passed to the tag route and routing was incorrect.
This patch fails, cause on reload route does not point to the right place.
```
-- a/app/assets/javascripts/discourse/app/routes/tag-show.js
+++ b/app/assets/javascripts/discourse/app/routes/tag-show.js
@@ -89,6 +89,8 @@ export default DiscourseRoute.extend(FilterModeMixin, {
filter = `tag/${tagId}/l/${topicFilter}`;
}
const list = await findTopicList(
this.store,
this.topicTrackingState,
@@ -123,7 +125,7 @@ export default DiscourseRoute.extend(FilterModeMixin, {
},
setupController(controller, model) {
- const noSubcategories =
+ this.noSubcategories =
this.noSubcategories === undefined
? model.category?.default_list_filter === NONE
: this.noSubcategories;
@@ -133,7 +135,7 @@ export default DiscourseRoute.extend(FilterModeMixin, {
...model,
period: model.list.for_period,
navMode: this.navMode,
- noSubcategories,
+ noSubcategories: this.noSubcategories,
loading: false,
});
```
Long term we don't want to hide this logic from the routing (even in
the category case) it just cause unneeded confusion and fragility.
Every time we created a topic list we would leak a state change callback
This happens on any topic list -> topic -> topic list sequence
This can cause corruption of tracking state and memory bloating given that
all information may be sent to the sync function.
Forcing distributed muted to raise when a notify reviewable job is running
leads to excessive errors in the logs under many conditions.
The new pattern
1. Optimises the counting of reviewables so it is a lot faster
2. Holds the distributed lock for 2 minutes (max)
The downside is the job queue can get blocked up when tons of notify
reviewables are running at the same time. However this should be very
rare in the real world, as we only notify when stuff is flagged which
is fairly infrequent.
This also give a fair bit more time for the notifications which may be
a little slow on large sites with tons of mods.
- Reduce duplication of terms in post index from unlimited to 6. This will
result in reduced index size and reduced weighting for posts containing
a huge amount of duplicate terms. (Eg: a post containing "sam sam sam sam
sam sam sam sam", will index as "sam sam sam sam sam sam", only including
the word up to 6 times.) This corrects a flaw where title weighting could
be ignored.
- Prioritize exact matches of words in titles. Our search always performs
a prefix match. However we want to give special weight to exact title matches
meaning that a search for "sum" will find topics such as "the sum of us" vs
"summer in spring".
- Pick up fixes to our search algorithm which are missing from old indexes.
Specifically pick up the fix that indexes URLs properly. (`https://happy.com`
was stemmed to `happi` in keywords and then was not searchable)
see also:
https://meta.discourse.org/t/refinements-to-search-being-tested-on-meta/254158
Indexing will take a while and work in batches, in the background.
16 bit images were not returning the correct dominant color due truncation
The routine expected an 8bit color eg: #FFAA00, but ended up getting a 16bit one eg: #FFFAAA000. This caused a truncation, which leads to wildly off colors.
* FIX: do not notify admins on suppressed categories
Avoid notifying admins on categories where they are not explicitly members
in cases where SiteSetting.suppress_secured_categories_from_admin is
enabled.
This helps keep notification stream clean and avoids admins mistakenly
being invited to discussions that should be suppressed
Previously after uploads completed post raw would drift.
If you autocompleted text after the upload stub got replaced it would
insert in the wrong position.
Previously due to an error archived topics were more prominent in search
than closed topics.
This amends our internal logic to ensure archived topics are bumped down
the list.
If a post contains domain with a word that stems to a non prefix single
words will not match it.
For example: in happy.com, `happy` stems to `happi`. Thus searches for happy
will not find URLs with it included.
This bloats the index a tiny bit, but impact is limited.
Will require a full reindex of search to take effect.
When we are done refining search we can consider a full version bump.
Previously to_tsquery would split terms and join with &
In PG 14 terms are split and use <-> which means followed directly by.
In PG 13:
discourse_test=# SELECT to_tsquery('english', '''hello world''');
to_tsquery
---------------------
'hello' & 'world'
(1 row)
In PG 14:
discourse_test=# SELECT to_tsquery('english', '''hello world''');
to_tsquery
---------------------
'hello' <-> 'world'
(1 row)
Change is very unobtrosive, we simply amend our to_tsquery to behave like
it used to behave and make no use of the `<->` operator
More detail at: https://akorotkov.github.io/blog/2021/05/22/pg-14-query-parsing/
Note that plainto_tsquery used elsewhere in Discourse keeps the exact
same function.
This also corrects a faulty test that was passing by a fluke on older
version of PG
Previous regex did not allow for cases where a lexeme contains a : (colon)
This can happen when parsing URLs. New algorithm allows for this.
Test was amended to more clearly call out index problems
The new `prioritize_exact_search_match` can be used to force the search
algorithm to prioritize exact term matches in title when ranking results.
This is scoped narrowly to titles for cases such as a topic titled:
"organisation chart" and a search of "org chart".
If we scoped this wider, all discussion about "org chart" would float to
the top and leave a very common title de-prioritized.
This is a hidden site setting and it has some performance impact due
to double ranking.
That said, performance impact is somewhat mitigated cause ranking on
title alone is a very cheap operation.
* FEATURE: allow restricting duplication in search index
This introduces the site setting `max_duplicate_search_index_terms`.
Using this number we limit the amount of duplication in our search index.
This allows us to more correctly weight title searches, so bloated posts
don't unfairly bump to the top of search results.
This feature is completely disabled by default and behind a site setting
We will experiment with it first. Note entire search index must be rebuilt
for it to take effect.
---------
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
Many users seems surprised by prefix matching in search leading to
unexpected results.
Over the years we always would return results starting with a search term
and not expect exact matches.
Meaning a search for `abra` would find `abracadabra`
This introduces the Site Setting `enable_search_prefix_matching` which
defaults to true. (behavior unchanged)
We plan to experiment on select sites with exact matches to see if the
results are less surprising
In "GlobalSetting.redirect_avatar_requests" mode, when the application gets
an avatar request it returns a "redirect" to the S3 CDN.
This shields the application from caching avatars and downloading from S3.
However clients will make 2 requests per avatar. (one to get redirect,
second to get avatar)
A one hour cache on a redirect means there may be an increase in CDN
traffic, given more clients will ask for the redirect every hour.
This may also lead to an increase in origin requests to the application.
To mitigate lets cache the CDN URL for 1 day.
The downside is that any changes to S3 CDN need extra care to allow for
the extra 1 day delay. (leave data around for 1 extra day)
This will give us some aggregate stats on the defer queue performance.
It is limited to 100 entries (for safety) which is stored in an LRU cache.
Scheduler::Defer.stats can then be used to get an array that denotes:
- number of runs and completions (queued, finished)
- error count (errors)
- total duration (duration)
We can look later at exposing these metrics to gain visibility on the reason
the defer queue is clogged.
Previously we would unconditionally fetch all images via HTTP to grab
original sizing from cooked post processor in 2 different spots.
This was wasteful as we already calculate and cache this info in upload records.
This also simplifies some specs and reduces use of mocks.
The hidden site setting `suppress_secured_categories_from_admin` will
suppress visibility of categories without explicit access from admins
in a few key areas (category drop downs and topic lists)
It is not intended to be a security wall since admins can amend any site
setting. Instead it is feature that allows hiding the categories from the
UI.
Admins will still be able to see topics in categories without explicit
access using direct URLs or flags.
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
* FIX: allow tl4 to bulk select
- Also allows tl4 to perform batch tagging
---
Long term this needs to be rewritten to account for "bulk action" permission
given from the server.
Co-authored-by: Martin Brennan <martin@discourse.org>
RS256 was added for Windows Hello and as a side effect we speculatively added
RS384 and RS512. These ciphers were not tested and are now failing on solo
keys. It may be the case that the ciphers are not configured correctly on
our side. It may be the case that this is a Solo key bug.
Regardless, we are removing the ciphers and will only consider adding them
again if absolutely needed.
Certain HTML can be rejected by nokogumbo, specifically cases where there
are enormous amounts of attributes
This ensures that malformed HTML is simply skipped instead of leaking out
an exception and terminating downstream processes.
normalized_encode in addressable has a number of issues, including https://github.com/sporkmonger/addressable/issues/472
To temporaily work around those issues for the majority of cases, we try parsing with `::URI`. If that fails (e.g. due to non-ascii characters) then we will fall back to addressable.
Hopefully we can simplify this back to `Addressable::URI.normalized_encode` in the future.
This commit also adds support for unicode domain names and emoji domain names with escape_uri.
This removes an unneeded hack checking for pre-signed urls, which are now handled by the general case due to starting off valid and only being minimally normalized. Previous test case continues to pass.
UrlHelper.s3_presigned_url? which was somewhat wide was removed.
Our internal implementation of #perform on jobs performs remapping.
This happens cause we do "exception aggregation".
Scheduled jobs run on every site in the multisite cluster, and we report
one error per site that failed. During this aggregation we reshape the
context from the original object shape returned by mini_scheduler
The new integration test ensures this interface will remain stable even if
decoupled parts of the code change shapes.
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
* FIX: don't memoize site setting in guardian
Memoizing site settings can make tests more fragile and harder to debug
Co-authored-by: Jarek Radosz <jradosz@gmail.com>
* FEATURE: track stats around failing scheduled jobs
Discourse.job_exception_stats can now be used to gather stats around how
many regular scheduled jobs failed in the current process.
This will be consumed by the Prometheus plugin and potentially other
monitoring plugins.
* FIX: bots could generate errors when slug generation method is encoded
When slug generation method is encoded (non default) then bots could
cause errors in the logs for urls containing special chars.
ó for example in a URL can be requested in a valid ASCII-8BIT string, and
later when joined to UTF-8 would result in encoding issues.
Fix here ensures we force encoding correctly for outlier cases.
Browser tend to always encode these chars, hence we did not notice this.
Co-authored-by: Jarek Radosz <jradosz@gmail.com>
Certain rogue bots such as Yandex may send across invalid CSP reports
when CSP report collection is enabled.
This ensures that invalid reports will not cause log floods and simply
returns a 422 error.
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
Sam
