* FEATURE: hide_email_address_taken forces use of email in forgot password form
This strengthens this site setting which is meant to be used to harden sites
that are experiencing abuse on forgot password routes.
Previously we would only deny letting people know if forgot password worked on not
New change also bans usage of username for forgot password when enabled
This column was dropped in a previous commit, in post migrations.
Unfortunatly that causes smoke tests to fail as there is a period between
migration and post migrations where records can not be inserted into the
table.
Old OnceOff job could perform pretty slowly on sites with millions of emails
New implementation operates in batches in a migration, minimizing locking.
The lazy route initialization cuts down boot time of rails.
On my local system it cuts out 200ms of boot time taking me from 3.2 to 3 seconds.
This is not a radically enormous amount of time, but paper cuts add up, and a faster boot in dev will make everyone happy.
TBD if we want to also include this in production.
Gem is heavily maintained by @amatsuda, last commit 3 days ago.
Previously due to "rowheader" role we would read out topic titles twice.
This adjusts it so we apply the heading role only to the topic link.
In turn this makes navigation through topic lists more accurate (h) only
lands you on topic links. It also reduces the amount of duplicate reading
NVDA does.
Before:
Topic title link new topic link support link b481 link 19h link 2 button...
After:
Topic title link
This reduces noise, up and down once you land on a topic link can give you
more context.
Previously we would need to launch unicorn separately this achieves
the same goal by making 2 modifications:
1. If -u is supplied ember-cli binary will launch and monitor ember cli and unicorn
2. We suppress 200 requests to keep console clean (we may consider moving to development rails logs)
Also cleans out output a bit by supplying silent flags to yarn.
Over the years we have found that a few communities never discovered tags.
Instead of having them default off we now have them default on, ensuring
that everyone finds out about them.
Co-authored-by: Dan Ungureanu <dan@ungureanu.me>
Previous to this change we would switch off MessageBus updating after 20
minutes.
This ensures that when the user becomes present again we turn on long polling.
Without long polling updates can be delayed for minutes.
* UX: alert screen readers when there is an issue saving a post
Adds a "alert" role to various popup-input-tips.
This means screen reader users can now tell why a post refuses to save.
Also ensures like icon in the "try the like button" has screen reader support
Previously auto focus would only work on modals that include buttons or
inputs.
To avoid a situation where information modals such as keyboard shortcuts
do not get focus, simply focus on the close button as a fallback.
Previously we would retry push notifications indefinitely for all errors
except for ExpiredSubscription
Under certain conditions other persistent errors may arise such as a persistent
rate limit.
If we track more than 3 errors in a period of time longer than a day we will
delete the subscription
Also performs a bit of internal cleanup to ensure protected methods really
are private.
Previously we had no role set for various topic links, nor did we have any
headers.
This teaches screen readers that topic links in topic lists are to be treated
as H2. We opted for this less radical change cause a change of the element
type would probably result in many broken themes.
Confirmed on NVDA you can very quickly breeze through topic lists now. Minor
edge case is pinned topics which can be a bit annoying due to multiple links.
NVDA does not detect HTML5 articles as regions. This explicitly sets a
region with an aria-label denoting post numbers making it much easier to
know where you are in a topic.
Note role: article which is more semantically correct is not respected by
NVDA d/D shortcut, hence the much more generic "region" role.
Under certain conditions admins would miss messages when posting action in
topics where they have permission.
This also fixes an error where we would sometimes explode when publishing to
an empty group.
Previously watched words ignored topic titles when applying auto tagging rules.
Also copy has been improved to reflect how the system behaves.
The text hints that we are only watching first post now
scopes are incredibly annoying to preload, simply adding :user_emails is not
enough.
Instead of relying on scopes simply iterate through user_emails which is
properly preloaded.
This removes 2 * N+1 when generating user reports.
We do not really support Discourse on Alpine linux / musl based distros.
We depend heavily on libc.
Besides this it looks like rubygems/bundler tends to handle this nuance incorrectly.
Remove it for now.
New version of mini_racer support M1 and ARM builds of Ruby
This should make developing on Discourse easier
Co-authored-by: Robin Ward <robin.ward@gmail.com>
Previously certain images may lead to convert / identify to run for unreasonable
amounts of time
This adds a maximum amount of time these commands can run prior to forcing
them to stop
Previously we used the raw data indexed to generate blurbs even for cases
when Chinese/Korean/Japanese text was used.
This caused superfluous spaces to show up in excerpts.
* DEV: upgrade mini_sql
Even though we are not planning on using this quite yet, mini_sql now supports
prepared statements.
Would like this upgrade merged so we can do some benchmarking.
Note, this will not work with pg_bouncer, but sites that are not using it
may benefit from the feature.
* implement multisite friendly prepared statements
In extreme circumstances when the uploads table is huge, the old version of
this migration could take a very long time.
The rewrite extracts the sha1 directly from the badges table and does an index
based match on the uploads table
Corrects flaky spec
Previously we were only clearing state after our spec ran, leaving possible
edge cases where `Discourse.plugins` had data.
Clean-up source of the plugin leak
1 pop was not enough to clear the plugin, plus make specs a bit more
deliberate
Added a new task to test if indexes are coherent with a blank database
This allows us to detect for cases where somehow indexes are out of sync
FIX_INDEXES=1 or `rake db:validate_indexes[fix]` to correct the issues it finds.
Detects:
- Badly named indexes that need to be renamed
- Missing indexes
- Extra indexes
Can correct all 3 with the fix option
This setting is confusing admins, it is now hidden by default
It only applies to major updates of the rendering engine or imports and
very infrequently needs tweaking
* FIX: optimise MoveNewSinceToTable
Avoids shuffling all ids around to the app (only use min / max)
Ensure the query for boundaries is ordered by user_id
This little helper script allows for easy ember cli development.
To see the options run `bin/ember-cli -h`
It allows you to proxy try.discourse.org with the `bin/ember-cli --try`
switch, which effectively allows for some development without a rails installed.
It passes on arguments to ember-cli so you can customize port and so on.
It makes the assumption that on local people are using `bin/unicorn` for
development. (it includes some extra discourse specific helpers)
These are a few small tweaks that slightly improve performance.
- we omitted 1 query from the post guardian which could cause an N+1
- cook_url has been sped up a bit
- url helper avoids re-creating sets for no reason
Inventory on S3 always lagged, over the past few weeks we are noticing that
1 day of lag is not enough.
We are increasing this to 2, to ensure that we do not get false positive
reports.
These 2 indexes optimise performance on profile pages.
The summary page displays:
1. A list of "Top Link" - links sorted by number of clicks posted by user
2. A list of "Top Replies" - replies made by a user that go the most hearts
These two areas could devolve into full index or table scans, new indexes are there to avoid this cost on large dbs
One minor downside is that storage requirements go a tiny bit up to maintain the new indexes
see: https://bundler.io/blog/2020/12/09/bundler-v2-2.html
New version of bundler includes explicit callouts for each platform supported
This means we need to bundle on all our supported platforms, the number should be fairly low as x86 linux covers most
Final sigma is not lower cased correctly in Ruby causing issues with routing.
This works around the issue by downcasing all usernames containing a sigma using JS.
Googlebot handles no-index headers very elegantly. It advises to leave as many routes as possible open and uses headers for high fidelity rules regarding indexes.
Discourse adds special `x-robot-tags` noindex headers to users, badges, groups, search and tag routes.
Following up on b52143feff we now have it so Googlebot gets special handling.
Rest of the crawlers get a far more aggressive disallow list to protect against excessive crawling.
You can now use `@me` to search for posts created by yourself, this is particularly handy if you have a long username.
`@me rainbow` will find all posts you created with the word rainbow.
Also cleans up test suite so it has no warnings.
The notification panel gets resized and the JS uses maxWidth of 320.
This tends to fight with the CSS causing notifications to "jump" a bit when a new one lands.
chromium may report float device pixel ratio below 1.5 that is still clearly retina:
```
window.devicePixelRatio
1.4999998807907104
```
We used to round this down to 1 and not provide these browsers with retina avatars.
New algorithm is much more forgiving, anything over 1.1 gets 2x images, anything over 2.1 gets 3x images.
Rapid concurrent SSO attempts is something that happens quite frequently
in the wild at large enough scale.
When this happens conditions such as adding a user to a group could possibly
fire concurrently causing a user to be added to the same group twice and
erroring out.
To avoid all concurrency issues here we protect with a coarse distributed
mutex. This heavily mitigates the risk around concurrent group additions and
concurrent updates to user related records.
disable_search_queue_threshold needs to be coerced to a float so it is not
treated as a string when sub second values are provided.
Longer term fix is to possibly provide hints in the config so we do the
coersion automatically. However this would be a far more complex change.
Previous to this change our anonymous rate limits acted as a throttle.
New implementation means we now also consider rate limited requests towards
the limit.
This means that if an anonymous user is hammering the server it will not be
able to get any requests through until it subsides with traffic.
Discourse used to break from convention by logging out all sessions on any
specific session logout.
This would leave users confused about why mobile is logged out when the user
logged out of desktop.
log_out_strict is too conservative for most and not the pattern the industry
has adopted (google/twitter/facebook all perform no strict logouts)
Most proxies out there will work with chunked encoding transfer. However
some proxies buffer, causing large delays which in turn force the message
bus client to disable chunked encoding. This wastes a request to the message
bus causing superfluous load on the server.
Also
- enableLongPolling is already default true in the client, no need to set it
- remove confusing comment about zepto
Previously if a onebox timed out we would not present the users in the log
with any information regarding the onebox. This makes it very difficult to
debug.
This adds url/topic/user in the debugging output.
This allows for an advanced feature where hitting control on click or
CTRL-SHIFT-ENTER will lead to a post being made but the browser not to
scroll to the end.
* FEATURE: CTRL-SHIFT-ENTER and SHIFT-Click do not scroll on post
This allows for an advanced feature where hitting control on click or
CTRL-SHIFT-ENTER will lead to a post being made but the browser not to
scroll to the end.
Previous to this change we had no protection to ensure we wait on a request
for more posts prior starting another request.
In outlier cases if 10 people post at the same time on a topic a flood of
requests could start.
To improve this situation we now ensure that we are done asking for new posts
prior to asking for the next batch.
Also addresses some style issues raised previously and moves init to top
of class.
* FEATURE: when we fail to ship topic timings attempt to retry
This change amends it so
1. Topic timings are treated as background requests and subject to more
aggressive rate limits.
2. If we notice an error when we ship timings we back off exponentially
The commit allows 405, 429, 500, 501, 502, 503 and 504 errors to be retried.
500+ errors usually happen when self hosters are rebuilding or some other
weird condition.
405 happens when site is in readonly.
429 happens when user is rate limited.
The retry cadence is hardcoded in AJAX_FAILURE_DELAYS, longest delay is
40 seconds, we may consider enlarging it.
After the last delay passes we give up and do not write timings to the
server.
* Address feedback
- Omit promise, no need to use promises in sendNextConsolidatedTiming
- Correct issue where >= -1 was used and > -1 was intended
- Use objects for consolidated timings instead of Array
- Stop using shift/unshift and instead use push / pop which are faster
* Move consolidated timing initialization to constructor
* Remove TODO and just console.warn if we have an issue
When the server gets overloaded and lots of requests start queuing server
will attempt to shed load by returning 429 errors on background requests.
The client can flag a request as background by setting the header:
`Discourse-Background` to `true`
Out-of-the-box we shed load when the queue time goes above 0.5 seconds.
The only request we shed at the moment is the request to load up a new post
when someone posts to a topic.
We can extend this as we go with a more general pattern on the client.
Previous to this change, rate limiting would "break" the post stream which
would make suggested topics vanish and users would have to scroll the page
to see more posts in the topic.
Server needs this protection for cases where tons of clients are navigated
to a topic and a new post is made. This can lead to a self inflicted denial
of service if enough clients are viewing the topic.
Due to the internal security design of Discourse it is hard for a large
number of clients to share a channel where we would pass the full post body
via the message bus.
It also renames (and deprecates) triggerNewPostInStream to triggerNewPostsInStream
This allows us to load a batch of new posts cleanly, so the controller can
keep track of a backlog
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
This code is a no-op on all sites, even though it looks rather dangerous
this migration has long run prior to people trying exploit it.
That said ... hygiene here ... is not good.
Remove this legacy, we do not want it, even in historical migrations.
* PERF: avoid lookbehinds when indexing search
Previously we used a `EmailCook.url_regexp` this regex used lookbehinds
Unfortunately certain strings could lead to pathological behavior causing
CPU to skyrocket and regex replace to take a very very long time.
EmailCook still needs a fix, but it is less urgent cause it already splits
to single lines. That said we will correct that as well in a seperate PR.
New implementation is far more naive and relies on the extra spaces search
indexer inserts.
To avoid blocking the sidekiq queue a limit of 10,000 digests per 30 minutes
is introduced.
This acts as a safety measure that makes sure we don't keep pouring oil on
a fire.
On multisites it is recommended to set the number way lower so sites do not
dominate the backlog. A reasonable default for multisites may be 100-500.
This can be controlled with the environment var
DISCOURSE_MAX_DIGESTS_ENQUEUED_PER_30_MINS_PER_SITE
The site setting `search_recent_posts_size` controls the window of posts
that we will search through before trying to search through the full index
If this number is too low then the search quality can suffer a lot as recent
posts may dominate search. If the number is too high then performance will
suffer.
This attempts to find a happy medium, 1 million posts will cover the majority
of forums out there and should perform adequately.
typographer can change " to ” leading to breakages in parser
At least codify this. Longer term we want to re-prioritize typographer so
it always runs after bbcode parsing.
Previously attributes such as `[test a='a"a' b="a'a"]` were not correctly
handled.
This amends the regex parser to ensure it correctly parses attributes
without breaking incorrectly on the first nested quote
Overall stubs lead to long term instability, this helps stabilize them.
Also fixed flaky spec around plugin hooks.
It was relying you `.posts` on system_user which could be loaded already
and invalid. Instead we now load it by hand.
In some cases Discourse admins may opt for sessions not to persist when a
browser is closed.
This is particularly useful in healthcare and education settings where
computers are shared among multiple workers.
By default `persistent_sessions` site setting is enabled, to opt out you
must disable the site setting.
This is a temporary fix since fonts may be cached without CORS
A larger scale change needed.
Our long term solution here is to have the SCSS compilation pipeline take
care of this by introducing some SCSS var that can be consumed in the various
stylesheets.
We do not want to write a file on boot
Previously in some cases the test suite could fail due to a bad entry in
redis from previous tests
This ensures the correct cache is expired when needed
Additionally improves performance of the redis check
After thinking about it, I worry that this will potentially leave a site
setting set when people hit ctrl-c ... feels a tiny bit risky, so leaving
it out.
- Introduces uploads:delete_missing_s3 which can be used to "give up" and
delete broken records from the database
- Fixes a bug in fix_missing_s3 - crashing on deleted posts
- Adds more info to analyze_missing_s3
Very large batches can take an enormous amount of time due to churn
Limiting to 200k changes at a time gives us a far larger chance of finishing
the job without timing out or deadlocking.
A giant transaction in a post migration can be very risky.
This splits the large amount of work this migration needs to do into 2 parts:
1. A re-runnable cleanup job prior to transaction
2. A minimally sized transaction to add the database constraint
This avoids large amounts of churn on the table
There is an fk to user_profile that can make destroying uploads fail
if they happen to be set as user profile.
This ensures we clear this information when destroying uploads.
There are more relationships, but this makes some more progress.
Sam
