In our production environment, we have been seeing Sidekiq processes
getting stuck randomly when a USR1 signal is sent to the Unicorn master
process. We have not been able to identify the root cause of why the
Sidekiq process gets stuck. We however noticed that when the Unicorn
master process receives a USR1 signal, it will reopen the logs for the
Unicorn master process first before sending a USR1 signal for the
Unicorn worker processes to reopen the logs. We figured that we should
do the same for the Sidekiq process as well when a USR1 signal.
In this commit, we introduce an arbitrary delay of 1 second before we
the Sidekiq process reopens its log files so as to allow enough time for the Unicorn
master to finish reopening it logs first.
We also do not send reopen logs for the Sidekiq process if the `DISCOURSE_LOG_SIDEKIQ`
env is not present because there is no need to reopen any logs.
Introduce a new log line for Sidekiq jobs that consume more than
`DISCOURSE_LIVE_SLOTS_SIDEKIQ_LIMIT` live slots. This is useful to
track down jobs that may leak memory.
This is enabled only when Sidekiq's job instrumenter is enabled (set
`DISCOURSE_LOG_SIDEKIQ` to `1`).
Why this change?
On CI, we have been seeing the "handles job concurrency" job timing out
on CI after 45 seconds. Upon closer inspection of `Jobs::Base#perform`
when cluster concurrency has been set, we see that a thread is spun up
to extend the expiring of a redis key by 120 seconds every 60 seconds
while the job is still being executed. The thread looks like this before
the fix:
```
keepalive_thread =
Thread.new do
while parent_thread.alive? && !finished
Discourse.redis.without_namespace.expire(cluster_concurrency_redis_key, 120)
sleep 60
end
end
```
In an ensure block of `Jobs::Base#perform`, the thread is stop by doing
something like this:
```
finished = true
keepalive_thread.wakeup
keepalive_thread.join
```
If the thread is sleeping, `keepalive_thread.wakeup` will stop the
`sleep` method and run the next iteration causing the thread to
complete. However, there is a timing issue at play here. If
`keepalive_thread.wakeup` is called at a time when the thread is not
sleeping, it will have no effect and the thread may end up sleeping for
60 seconds which is longer than our timeout on CI of 45 seconds.
What does this change do?
1. Change `sleep 60` to sleep in intervals of 1 second checking if the
job has been finished each time.
2. Add `use_redis_snapshotting` to `Jobs::Base` spec since Redis is
involved in scheduling and we want to ensure we don't leak Redis
keys.
3. Add `ConcurrentJob.stop!` and `thread.join` to `ensure` block in "handles job concurrency"
test since a failing expectation will cause us to not clean up the
thread we created in the test.
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.
```
class Jobs::DummyDelayedJob < Jobs::Base
def execute(args = {})
end
end
RSpec.describe "Jobs.run_immediately!" do
before { Jobs.run_immediately! }
it "explodes" do
current_user = Fabricate(:user)
Jobs.enqueue_in(1.seconds, :dummy_delayed_job)
sign_in(current_user)
end
end
```
The test above will fail with the following error if `ActiveRecord::Base.connection_handler.clear_active_connections!` is called before the configured Capybara server checks out a connection from the connection pool.
```
ActiveRecord::ActiveRecordError:
Cannot expire connection, it is owned by a different thread: #<Thread:0x00007f437391df58@puma srv tp 001 /home/tgxworld/.asdf/installs/ruby/3.1.3/lib/ruby/gems/3.1.0/gems/puma-6.0.2/lib/puma/thread_pool.rb:106 sleep_forever>. Current thread: #<Thread:0x00007f437d6cfc60 run>.
```
We're not exactly sure if this is an ActiveRecord bug or not but we've
invested too much time into investigating this problem. Fundamentally,
we also no longer understand why `ActiveRecord::Base.connection_handler.clear_active_connections!` is being called in an ensure block
within `Jobs::Base#perform` which was added in
ceddb6e0da 10 years ago. This
commit moves the logic for running jobs immediately out of the
`Jobs::Base#perform` method into another `Jobs::Base#perform_immediately` method such that
`ActiveRecord::Base.connection_handler.clear_active_connections!` is not
called. This change will only impact the test environment.
This allows text editors to use correct syntax coloring for the heredoc sections.
Heredoc tag names we use:
languages: SQL, JS, RUBY, LUA, HTML, CSS, SCSS, SH, HBS, XML, YAML/YML, MF, ICS
other: MD, TEXT/TXT, RAW, EMAIL
This commit introduces our own handling and warning for Sidekiq's new 'non-json-serializable' warning. This decouples us from Sidekiq's own deprecation cycle, and allows us to use our own deprecation system. It also means that the dump/parse happens in test mode, which will help us to catch occurrences before they reach production.
The latest version of Sidekiq introduced a warning when jobs are queued with arguments which 'do not stringify to JSON safely'. In the vast majority of cases, this is because a hash is passed with symbols as keys. When those args are passed to the job, the keys will be stringified.
Our job wrapper already takes care of this issue by calling '.with_indifferent_access' on the args before passing them to `#execute`, so we don't need to change anything about our use. All we need to do is satisfy Sidekiq's warning system by 'stringifying' all the keys before enqueuing the job.
This reverts commit f5cf647e57.
The gem breaks usage of Rails URL helpers when used outside views and
controllers, for example in
88ecb83382/app/models/upload.rb (L239-L242)
the `upload_short_path` method call fails with an undefined method
exception when this gem is enabled.
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.
When jobs are enqueued inside a transaction, it's possible that they will be executed before the necessary data is available in the database. This commit ensures all jobs are enqueued in an ActiveRecord after_commit hook.
One potential downside here is if the job fails to enqueue, the transaction will no longer be aborted. However, the chance of that happening is reasonably low, and the impact is significantly lower than the current issue where jobs are scheduled before their data is ready.
When running jobs in tests, we use `Jobs.run_immediately!`. This means that jobs are run synchronously when they are enqueued. Jobs sometimes enqueue other jobs, which are also executed synchronously. This means that the outermost job will block until the inner jobs have finished executing. In some cases (e.g. process_post with hotlinked images) this can lead to a deadlock.
This commit changes the behavior slightly. Now we will never run jobs inside other jobs. Instead, we will queue them up and run them sequentially in the order they were enqueued. As a whole, they are still executed synchronously. Consider the example
```ruby
class Jobs::InnerJob < Jobs::Base
def execute(args)
puts "Running inner job"
end
end
class Jobs::OuterJob < Jobs::Base
def execute(args)
puts "Starting outer job"
Jobs.enqueue(:inner_job)
puts "Finished outer job"
end
end
Jobs.enqueue(:outer_job)
puts "All jobs complete"
```
The old behavior would result in:
```
Starting outer job
Running inner job
Finished outer job
All jobs complete
```
The new behavior will result in:
```
Starting outer job
Finished outer job
Running inner job
All jobs complete
```
Previously we had many places in the app that called `hostname` to get
hostname of a server. This commit replaces the pattern in 2 ways
1. We cache the result in `Discourse.os_hostname` so it is only ever called once
2. We prefer to use Socket.gethostname which avoids making a shell command
This improves performance as we are not spawning hostname processes throughout
the app lifetime
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.
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
Includes support for flags, reviewable users and queued posts, with REST API
backwards compatibility.
Co-Authored-By: romanrizzi <romanalejandro@gmail.com>
Co-Authored-By: jjaffeux <j.jaffeux@gmail.com>
It is not a setting, and only relevant in specs. The new API is:
```
Jobs.run_later! # jobs will be thrown on the queue
Jobs.run_immediately! # jobs will run right away, avoid the queue
```
- Open the log file in "append" mode. This avoids issues if the file does not exist (and matches standard rails log behavior)
- Correctly parse the interval logging environment variable
By default, this does nothing. Two environment variables are available:
- `DISCOURSE_LOG_SIDEKIQ`
Set to `"1"` to enable logging. This will log all completed jobs to `log/rails/sidekiq.log`, along with various db/redis/network statistics. This is useful to track down poorly performing jobs.
- `DISCOURSE_LOG_SIDEKIQ_INTERVAL`
(seconds) Check running jobs periodically, and log their current duration. They will appear in the logs with `status:pending`. This is useful to track down jobs which take a long time, then crash sidekiq before completing.
This allows us to run regular rebakes without starving the normal queue.
It additionally adds the ability to specify queue with `Jobs.enqueue` so
we can specifically queue a job with lower priority using the `queue` arg.
If sidekiq is paused or Discourse is in readonly continue to queue
heartbeats
If we do not do that then a master process can end up reaping sidekiq
workers and causing various badness
This also impacts restore which can do weird stuff TM in cases like this
Bianca Nenciu
Alan Guo Xiang Tan
Renato Atilio
Daniel Waterworth
Sam
David Taylor
Martin Brennan
Jarek Radosz
Osama Sayegh
Dan Ungureanu
Krzysztof Kotlarek
Gerhard Schlager
Robin Ward
Neil Lalonde
Guo Xiang Tan
Régis Hanol