It seems like the overhead of GPU acceleration is not worth it and is
slowing down our system tests. Locally the following command completes
in `2 minutes 8.6 seconds` with GPU disabled as compared to `2 minutes 45.4 seconds` with GPU enabled.
## What is the problem?
MessageBus by default uses long polling which keeps a connection
open for 25 seconds by default. The problem here is that Capybara does not know about these
connections being kept opened by MessageBus and hence does not know how
to stop these connections at the end of each test. As a result, the long polling MessageBus connections are kept opened by the browser and we hit chrome's limit of 6 concurrent requests per host, new request made in the browser is marked as "pending" until a request is freed up. Since we keep a MessageBus long polling connection opened for 25 seconds, our finders in Capybara end up hitting Capybara's wait time out causing the tests to fail.
## What is the fix?
Since we can't rely on Capybara to close all the existing Capybara
connections, we manually execute a script to stop all MessageBus
connections after each system test.
```
for i in {1..10}; do
echo "Running iteration $i"
PARALLEL_TEST_PROCESSORS=8 CAPYBARA_DEFAULT_MAX_WAIT_TIME=10 bin/turbo_rspec --seed=34908 --profile --verbose --format documentation spec/system
if [ $? -ne 0 ]; then
echo "Error encountered on iteration $i"
exit 1
fi
done
echo "All 10 iterations completed successfully"
```
Without the fix, the script fails consistently in the first few iterations. Running in non-headless mode with the "network" tab opened will reveal the requests that are marked as pending.
Rescuing them still makes timing-out tests fail but doesn't break `after` spec cleanup (which could trigger more errors) Using custom error class to avoid any other possible timeout-catching code.
Also:
* remove an unnecessary `.select { |x| x.size > 0 }`
* fix a typo in a test title
Why is this change required?
By default, `RSpec` comes with a `--profile=[COUNT]` option as well but
enabling that option means that the entire test suite needs to be
executed. This does not work so well for `turbo_rspec` which splits our
test files into various "buckets" for the tests to be executed in
multiple processes. Therefore, this commit adds a similar
`--profile=[COUNT]` option to `turbo_rspec` but will only profile the
tests being executed. Examples:
`LOAD_PLUGINS=1 bin/turbo_rspec --profile plugins/*/spec/system`
or
`LOAD_PLUGINS=1 bin/turbo_rspec --profile=20 plugins/*/spec/system`
What is this change?
This change is an attempt to avoid flakiness in our tests due to
animations being enabled in our tests. An example of flakiness caused by
animations is when the `find(selector).click` pattern is used. When
`find(selector)` returns the node, its position may have changed if the
element is still moving. However, the `click` method will end up
clicking on the old position.
Either way, there is no need for us to make system tests even more
complicated by enabling animations.
Usage:
```
CHROME_DEV_TOOLS=bottom bundle exec rspec /path/to/system/spec
```
This commit also regroups common chrome options under `apply_base_chrome_options`, and removes the size of the mobile window which was incorrect. browser_log param is also passed to mobile chrome options.
New headless shares the same implementation as the chrome browser
instead of being a separate implementation of its own.
See https://developer.chrome.com/articles/new-headless/ for more
details
Co-authored-by: Rafael dos Santos Silva <xfalcox@gmail.com>
This commit also includes two changes to the rails helper which make tests more consistent on different devices. With this change the failure was reproducible locally and not only on CI:
```
options.add_argument("--force-device-scale-factor=1")
```
The fix itself is quite simple and attempts to find safe click coordinates, the previous solution could fail depending on the size of the sidebar.
* Color for turbo_rspec in CI (`progress` and `documentation` formats)
* Show "DONE" only when `documentation` formatter is used
* Fix formatting
* Collapse RSpec commands
* Add line wrapping to the `progress` formatter (to mitigate GH Actions issue)
This pull request is a full overhaul of the chat-composer and contains various improvements to the thread panel. They have been grouped in the same PR as lots of improvements/fixes to the thread panel needed an improved composer. This is meant as a first step.
### New features included in this PR
- A resizable side panel
- A clear dropzone area for uploads
- A simplified design for image uploads, this is only a first step towards more redesign of this area in the future
### Notable fixes in this PR
- Correct placeholder in thread panel
- Allows to edit the last message of a thread with arrow up
- Correctly focus composer when replying to a message
- The reply indicator is added instantly in the channel when starting a thread
- Prevents a large variety of bug where the composer could bug and prevent sending message or would clear your input while it has content
### Technical notes
To achieve this PR, three important changes have been made:
- `<ChatComposer>` has been fully rewritten and is now a glimmer component
- The chat composer now takes a `ChatMessage` as input which can directly be used in other operations, it simplifies a lot of logic as we are always working a with a `ChatMessage`
- `TextareaInteractor` has been created to wrap the existing `TextareaTextManipulation` mixin, it will make future migrations easier and allow us to have a less polluted `<ChatComposer>`
Note ".chat-live-pane" has been renamed ".chat-channel"
Design for upload dropzone is from @chapoi
Similar spirit to e195e6f614,
this moves the Bookmarkable registration to DiscoursePluginRegistry
so plugins which are not enabled do not register additional
bookmarkable classes.
This commit introduces the skeleton of the chat thread UI. The
structure of the components looks like this. Its done this way
so the side panel can be used for other things as well if we wish,
not just for threads:
```
.main-chat-outlet
<ChatLivePane />
<ChatSidePanel>
<-- rendered with {{outlet}} -->
<ChatThread />
</ChatSidePanel>
```
Later on the `ChatThreadList` will be rendered here as well.
Now, when you go to a channel you can open a thread by clicking
on either the Open Thread message action button or by clicking on
the reply indicator. This will take you to a route like `chat/c/:slug/:channelId/t/:threadId`.
This works on mobile as well.
This commit includes basic serializers and routes for threads,
as well as a new `ChatThreadsManager` service in JS that caches
threads for a channel the same way the channel threads manager does.
The chat messages inside the thread are intentionally left out
until a later PR.
**NOTE: These changes are gated behind the site setting enable_experimental_chat_threaded_discussions
and the threading_enabled boolean on a ChatChannel**
* DEV: Rnemae channel path to just c
Also swap the channel id and channel slug params to be consistent with core.
* linting
* channel_path
* Drop slugify helper and channel route without slug
* Request slug and route models through the channel model if possible
* DEV: Pass messageId as a dynamic segment instead of a query param
* Ensure change is backwards-compatible
* drop query param from oneboxes
* Correctly extract channelId from routes
* Better route organization using siblings for regular and near-message
* Ensures sessions are unique even when using parallelism
* prevents didReceiveAttrs to clear input mid test
* we disable animations in capybara so sometimes the message was barely showing
* adds wait
* ensures finished loading
* is it causing more harm than good?
* this check is slowing things for no reason
* actually target the button
* more resilient select chat message
* apply similar fix to bookmark
* fix
---------
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
In dev/prod, these are absorbed by unicorn. Most commonly, they occur when a client interrupts a message-bus long-polling request.
Also reverts the EPIPE workaround introduced in 011c9b9973
Having this set to ALL pollutes the JS system spec
logs with a bunch of unnecessary noise like this:
> "PresenceChannel '/chat-user/core/1' dropped message (received 315, expecting 246), resyncing..."
Or:
> "DEPRECATION: The \u003Cdiscourse@component:plugin-connector::ember1112>#save computed property was just overridden. This removes the computed property and replaces it with a plain value, and has been deprecated.
Now, we will only log errors. To configure this set
the `SELENIUM_BROWSER_LOG_LEVEL` env var.
Our working theory is that system tests on Github run on much less
powerful hardware as compared to running the tests on our work machines.
Hopefully, increasing the wait time now will help reduce some flakes
that we're seeing on Github.
It should fix flakeys we have due to using_session. This commit is also fixing tests which were failing constantly with treadsafe enabled.
A test has also bene skipped as the issue couldn't be found so far.
More info: https://github.com/teamcapybara/capybara#threadsafe-mode
Previously, browser logs would be printed to STDOUT halfway through the test run. This commit changes the behaviour so that the logs are included in the failure summary along with other rspec failure information.
Currently the `turbo:spec` task will fail when encountering system
tests as Capypara tries to use the same port for each process.
This simple change uses the same strategy as for databases, by just
incrementing the port number by `TEST_ENV_NUMBER` for each process.
We are all in on system specs, so this commit moves all the chat quoting acceptance tests (some of which have been skipped for a while) into system specs.
Follow up to a review in #18937, this commit changes the HashtagAutocompleteService to no longer use class variables to register hashtag data sources or types in context priority order. This is to address multisite concerns, where one site could e.g. have chat disabled and another might not. The filtered plugin registers I added will not be included if the plugin is disabled.
This ensures that all system tests are starting from a clean state and
not leak state between requests. Note that we have to simplify flush the
Redis db here because it is not pratical to manually clean up Redis keys
in system tests.
Will make your test run in an emulated iPhone 12 Pro view. It means you can now use `click(delay: 0.5)` to emulate some long press or that `mobile_view=1` will be set automatically.
Usage:
```
it "works", mobile: true do
visit("/")
end
```
Note: `window-size=390,950` is different than native iPhone 12 Pro size, but due to minimum browser size and the automated browser alert at the top of the view, this was the best size I could find.
This commit fleshes out and adds functionality for the new `#hashtag` search and
lookup system, still hidden behind the `enable_experimental_hashtag_autocomplete`
feature flag.
**Serverside**
We have two plugin API registration methods that are used to define data sources
(`register_hashtag_data_source`) and hashtag result type priorities depending on
the context (`register_hashtag_type_in_context`). Reading the comments in plugin.rb
should make it clear what these are doing. Reading the `HashtagAutocompleteService`
in full will likely help a lot as well.
Each data source is responsible for providing its own **lookup** and **search**
method that returns hashtag results based on the arguments provided. For example,
the category hashtag data source has to take into account parent categories and
how they relate, and each data source has to define their own icon to use for the
hashtag, and so on.
The `Site` serializer has two new attributes that source data from `HashtagAutocompleteService`.
There is `hashtag_icons` that is just a simple array of all the different icons that
can be used for allowlisting in our markdown pipeline, and there is `hashtag_context_configurations`
that is used to store the type priority orders for each registered context.
When sending emails, we cannot render the SVG icons for hashtags, so
we need to change the HTML hashtags to the normal `#hashtag` text.
**Markdown**
The `hashtag-autocomplete.js` file is where I have added the new `hashtag-autocomplete`
markdown rule, and like all of our rules this is used to cook the raw text on both the clientside
and on the serverside using MiniRacer. Only on the server side do we actually reach out to
the database with the `hashtagLookup` function, on the clientside we just render a plainer
version of the hashtag HTML. Only in the composer preview do we do further lookups based
on this.
This rule is the first one (that I can find) that uses the `currentUser` based on a passed
in `user_id` for guardian checks in markdown rendering code. This is the `last_editor_id`
for both the post and chat message. In some cases we need to cook without a user present,
so the `Discourse.system_user` is used in this case.
**Chat Channels**
This also contains the changes required for chat so that chat channels can be used
as a data source for hashtag searches and lookups. This data source will only be
used when `enable_experimental_hashtag_autocomplete` is `true`, so we don't have
to worry about channel results suddenly turning up.
------
**Known Rough Edges**
- Onebox excerpts will not render the icon svg/use tags, I plan to address that in a follow up PR
- Selecting a hashtag + pressing the Quote button will result in weird behaviour, I plan to address that in a follow up PR
- Mixed hashtag contexts for hashtags without a type suffix will not work correctly, e.g. #ux which is both a category and a channel slug will resolve to a category when used inside a post or within a [chat] transcript in that post. Users can get around this manually by adding the correct suffix, for example ::channel. We may get to this at some point in future
- Icons will not show for the hashtags in emails since SVG support is so terrible in email (this is not likely to be resolved, but still noting for posterity)
- Additional refinements and review fixes wil
This commit adds a new `/hashtag/search` endpoint and both
relevant JS and ruby plugin APIs to handle plugins adding their
own data sources and priority orders for types of things to search
when `#` is pressed.
A `context` param is added to `setupHashtagAutocomplete` which
a corresponding chat PR https://github.com/discourse/discourse-chat/pull/1302
will now use.
The UI calls `registerHashtagSearchParam` for each context that will
require a `#` search (e.g. the topic composer), for each type of record that
the context needs to search for, as well as a priority order for that type. Core
uses this call to add the `category` and `tag` data sources to the topic composer.
The `register_hashtag_data_source` ruby plugin API call is for plugins to
add a new data source for the hashtag searching endpoint, e.g. discourse-chat
may add a `channel` data source.
This functionality is hidden behind the `enable_experimental_hashtag_autocomplete`
flag, except for the change to `setupHashtagAutocomplete` since only core and
discourse-chat are using that function. Note this PR does **not** include required
changes for hashtag lookup or new styling.
This commit introduces rails system tests run with chromedriver, selenium,
and headless chrome to our testing toolbox.
We use the `webdrivers` gem and `selenium-webdriver` which is what
the latest Rails uses so the tests run locally and in CI out of the box.
You can use `SELENIUM_VERBOSE_DRIVER_LOGS=1` to show extra
verbose logs of what selenium is doing to communicate with the system
tests.
By default JS logs are verbose so errors from JS are shown when
running system tests, you can disable this with
`SELENIUM_DISABLE_VERBOSE_JS_LOGS=1`
You can use `SELENIUM_HEADLESS=0` to run the system
tests inside a chrome browser instead of headless, which can be useful to debug things
and see what the spec sees. See note above about `bin/ember-cli` to avoid
surprises.
I have modified `bin/turbo_rspec` to exclude `spec/system` by default,
support for parallel system specs is a little shaky right now and we don't
want them slowing down the turbo by default either.
### PageObjects and System Tests
To make querying and inspecting parts of the page easier
and more reusable inbetween system tests, we are using the
concept of [PageObjects](https://www.selenium.dev/documentation/test_practices/encouraged/page_object_models/) in
our system tests. A "Page" here is generally corresponds to
an overarching ember route, e.g. "Topic" for `/t/324345/some-topic`,
and this contains logic for querying components within the topic
such as "Posts".
I have also split "Modals" into their own entity. Further down the
line we may want to explore creating independent "Component"
contexts.
Capybara DSL should be included in each PageObject class,
reference for this can be found at https://rubydoc.info/github/teamcapybara/capybara/master#the-dsl
For system tests, since they are so slow, we want to focus on
the "happy path" and not do every different possible context
and branch check using them. They are meant to be overarching
tests that check a number of things are correct using the full stack
from JS and ember to rails to ruby and then the database.
### CI Setup
Whenever a system spec fails, a screenshot
is taken and a build artifact is produced _after the entire CI run is complete_,
which can be downloaded from the Actions UI in the repo.
Most importantly, a step to build the Ember app using Ember CLI
is needed, otherwise the JS assets cannot be found by capybara:
```
- name: Build Ember CLI
run: bin/ember-cli --build
```
A new `--build` argument has been added to `bin/ember-cli` for this
case, which is not needed locally if you already have the discourse
rails server running via `bin/ember-cli -u` since the whole server is built and
set up by default.
Co-authored-by: David Taylor <david@taylorhq.com>
The cache was causing state to leak between tests since the `WatchedWord` record in the DB would have been rolled back but `WordWatcher` still had the word in the cache.
This pull request follows on from https://github.com/discourse/discourse/pull/16308. This one does the following:
* Changes `BookmarkQuery` to allow for querying more than just Post and Topic bookmarkables
* Introduces a `Bookmark.register_bookmarkable` method which requires a model, serializer, fields and preload includes for searching. These registered `Bookmarkable` types are then used when validating new bookmarks, and also when determining which serializer to use for the bookmark list. The `Post` and `Topic` bookmarkables are registered by default.
* Adds new specific types for Post and Topic bookmark serializers along with preloading of associations in `UserBookmarkList`
* Changes to the user bookmark list template to allow for more generic bookmarkable types alongside the Post and Topic ones which need to display in a particular way
All of these changes are gated behind the `use_polymorphic_bookmarks` site setting, apart from the .hbs changes where I have updated the original `UserBookmarkSerializer` with some stub methods.
Following this PR will be several plugin PRs (for assign, chat, encrypt) that will register their own bookmarkable types or otherwise alter the bookmark serializers in their own way, also gated behind `use_polymorphic_bookmarks`.
This commit also removes `BookmarkQuery.preloaded_custom_fields` and the functionality surrounding it. It was added in 0cd502a558 but only used by one plugin (discourse-assign) where it has since been removed, and is now used by no plugins. We don't need it anymore.
Previously cached counting made redis calls in main thread and performed
the flush in main thread.
This could lead to pathological states in extreme heavy load.
This refactor reduces load and cleans up the interface
Also:
* Remove an unused method (#fill_email)
* Replace a method that was used just once (#generate_username) with `SecureRandom.alphanumeric`
* Remove an obsolete dev puma `tmp/restart` file logic
* File.exists? is deprecated and removed in Ruby 3.2 in favor of
File.exist?
* Dir.exists? is deprecated and removed in Ruby 3.2 in favor of
Dir.exist?
OmniAuth test mode is disabled by default, so that we can integration-test the omniauth strategies. Sometimes, we manually enable test mode for specific specs. This commit ensures that test_mode is always disabled again after each spec.
We can fake redis transactions so that `fab!` works for redis and PG
data, but it's too slow to be used indiscriminately. Instead, you can
opt into it with the `use_redis_snapshotting` helper.
Insofar as snapshotting allows us to `fab!` more things, it provides a
speedup.
