360 lines
16 KiB
Markdown
360 lines
16 KiB
Markdown
---
|
||
stage: Enablement
|
||
group: Distribution
|
||
info: To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments
|
||
---
|
||
|
||
# Troubleshooting Sidekiq
|
||
|
||
Sidekiq is the background job processor GitLab uses to asynchronously run
|
||
tasks. When things go wrong it can be difficult to troubleshoot. These
|
||
situations also tend to be high-pressure because a production system job queue
|
||
may be filling up. Users will notice when this happens because new branches
|
||
may not show up and merge requests may not be updated. The following are some
|
||
troubleshooting steps that will help you diagnose the bottleneck.
|
||
|
||
GitLab administrators/users should consider working through these
|
||
debug steps with GitLab Support so the backtraces can be analyzed by our team.
|
||
It may reveal a bug or necessary improvement in GitLab.
|
||
|
||
In any of the backtraces, be wary of suspecting cases where every
|
||
thread appears to be waiting in the database, Redis, or waiting to acquire
|
||
a mutex. This **may** mean there's contention in the database, for example,
|
||
but look for one thread that is different than the rest. This other thread
|
||
may be using all available CPU, or have a Ruby Global Interpreter Lock,
|
||
preventing other threads from continuing.
|
||
|
||
## Log arguments to Sidekiq jobs
|
||
|
||
[In GitLab 13.6 and later](https://gitlab.com/gitlab-org/gitlab/-/merge_requests/44853)
|
||
some arguments passed to Sidekiq jobs are logged by default.
|
||
To avoid logging sensitive information (for instance, password reset tokens),
|
||
GitLab logs numeric arguments for all workers, with overrides for some specific
|
||
workers where their arguments are not sensitive.
|
||
|
||
Example log output:
|
||
|
||
```json
|
||
{"severity":"INFO","time":"2020-06-08T14:37:37.892Z","class":"AdminEmailsWorker","args":["[FILTERED]","[FILTERED]","[FILTERED]"],"retry":3,"queue":"admin_emails","backtrace":true,"jid":"9e35e2674ac7b12d123e13cc","created_at":"2020-06-08T14:37:37.373Z","meta.user":"root","meta.caller_id":"Admin::EmailsController#create","correlation_id":"37D3lArJmT1","uber-trace-id":"2d942cc98cc1b561:6dc94409cfdd4d77:9fbe19bdee865293:1","enqueued_at":"2020-06-08T14:37:37.410Z","pid":65011,"message":"AdminEmailsWorker JID-9e35e2674ac7b12d123e13cc: done: 0.48085 sec","job_status":"done","scheduling_latency_s":0.001012,"redis_calls":9,"redis_duration_s":0.004608,"redis_read_bytes":696,"redis_write_bytes":6141,"duration_s":0.48085,"cpu_s":0.308849,"completed_at":"2020-06-08T14:37:37.892Z","db_duration_s":0.010742}
|
||
{"severity":"INFO","time":"2020-06-08T14:37:37.894Z","class":"ActiveJob::QueueAdapters::SidekiqAdapter::JobWrapper","wrapped":"ActionMailer::MailDeliveryJob","queue":"mailers","args":["[FILTERED]"],"retry":3,"backtrace":true,"jid":"e47a4f6793d475378432e3c8","created_at":"2020-06-08T14:37:37.884Z","meta.user":"root","meta.caller_id":"AdminEmailsWorker","correlation_id":"37D3lArJmT1","uber-trace-id":"2d942cc98cc1b561:29344de0f966446d:5c3b0e0e1bef987b:1","enqueued_at":"2020-06-08T14:37:37.885Z","pid":65011,"message":"ActiveJob::QueueAdapters::SidekiqAdapter::JobWrapper JID-e47a4f6793d475378432e3c8: start","job_status":"start","scheduling_latency_s":0.009473}
|
||
{"severity":"INFO","time":"2020-06-08T14:39:50.648Z","class":"NewIssueWorker","args":["455","1"],"retry":3,"queue":"new_issue","backtrace":true,"jid":"a24af71f96fd129ec47f5d1e","created_at":"2020-06-08T14:39:50.643Z","meta.user":"root","meta.project":"h5bp/html5-boilerplate","meta.root_namespace":"h5bp","meta.caller_id":"Projects::IssuesController#create","correlation_id":"f9UCZHqhuP7","uber-trace-id":"28f65730f99f55a3:a5d2b62dec38dffc:48ddd092707fa1b7:1","enqueued_at":"2020-06-08T14:39:50.646Z","pid":65011,"message":"NewIssueWorker JID-a24af71f96fd129ec47f5d1e: start","job_status":"start","scheduling_latency_s":0.001144}
|
||
```
|
||
|
||
When using [Sidekiq JSON logging](../logs.md#sidekiqlog),
|
||
arguments logs are limited to a maximum size of 10 kilobytes of text;
|
||
any arguments after this limit will be discarded and replaced with a
|
||
single argument containing the string `"..."`.
|
||
|
||
You can set `SIDEKIQ_LOG_ARGUMENTS` [environment variable](https://docs.gitlab.com/omnibus/settings/environment-variables.html)
|
||
to `0` (false) to disable argument logging.
|
||
|
||
Example:
|
||
|
||
```ruby
|
||
gitlab_rails['env'] = {"SIDEKIQ_LOG_ARGUMENTS" => "0"}
|
||
```
|
||
|
||
In GitLab 13.5 and earlier, set `SIDEKIQ_LOG_ARGUMENTS` to `1` to start logging arguments passed to Sidekiq.
|
||
|
||
## Thread dump
|
||
|
||
Send the Sidekiq process ID the `TTIN` signal and it will output thread
|
||
backtraces in the log file.
|
||
|
||
```shell
|
||
kill -TTIN <sidekiq_pid>
|
||
```
|
||
|
||
Check in `/var/log/gitlab/sidekiq/current` or `$GITLAB_HOME/log/sidekiq.log` for
|
||
the backtrace output. The backtraces will be lengthy and generally start with
|
||
several `WARN` level messages. Here's an example of a single thread's backtrace:
|
||
|
||
```plaintext
|
||
2016-04-13T06:21:20.022Z 31517 TID-orn4urby0 WARN: ActiveRecord::RecordNotFound: Couldn't find Note with 'id'=3375386
|
||
2016-04-13T06:21:20.022Z 31517 TID-orn4urby0 WARN: /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/activerecord-4.2.5.2/lib/active_record/core.rb:155:in `find'
|
||
/opt/gitlab/embedded/service/gitlab-rails/app/workers/new_note_worker.rb:7:in `perform'
|
||
/opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/sidekiq-4.0.1/lib/sidekiq/processor.rb:150:in `execute_job'
|
||
/opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/sidekiq-4.0.1/lib/sidekiq/processor.rb:132:in `block (2 levels) in process'
|
||
/opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/sidekiq-4.0.1/lib/sidekiq/middleware/chain.rb:127:in `block in invoke'
|
||
/opt/gitlab/embedded/service/gitlab-rails/lib/gitlab/sidekiq_middleware/memory_killer.rb:17:in `call'
|
||
/opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/sidekiq-4.0.1/lib/sidekiq/middleware/chain.rb:129:in `block in invoke'
|
||
/opt/gitlab/embedded/service/gitlab-rails/lib/gitlab/sidekiq_middleware/arguments_logger.rb:6:in `call'
|
||
...
|
||
```
|
||
|
||
In some cases Sidekiq may be hung and unable to respond to the `TTIN` signal.
|
||
Move on to other troubleshooting methods if this happens.
|
||
|
||
## Process profiling with `perf`
|
||
|
||
Linux has a process profiling tool called `perf` that is helpful when a certain
|
||
process is eating up a lot of CPU. If you see high CPU usage and Sidekiq won't
|
||
respond to the `TTIN` signal, this is a good next step.
|
||
|
||
If `perf` is not installed on your system, install it with `apt-get` or `yum`:
|
||
|
||
```shell
|
||
# Debian
|
||
sudo apt-get install linux-tools
|
||
|
||
# Ubuntu (may require these additional Kernel packages)
|
||
sudo apt-get install linux-tools-common linux-tools-generic linux-tools-`uname -r`
|
||
|
||
# Red Hat/CentOS
|
||
sudo yum install perf
|
||
```
|
||
|
||
Run `perf` against the Sidekiq PID:
|
||
|
||
```shell
|
||
sudo perf record -p <sidekiq_pid>
|
||
```
|
||
|
||
Let this run for 30-60 seconds and then press Ctrl-C. Then view the `perf` report:
|
||
|
||
```shell
|
||
$ sudo perf report
|
||
|
||
# Sample output
|
||
Samples: 348K of event 'cycles', Event count (approx.): 280908431073
|
||
97.69% ruby nokogiri.so [.] xmlXPathNodeSetMergeAndClear
|
||
0.18% ruby libruby.so.2.1.0 [.] objspace_malloc_increase
|
||
0.12% ruby libc-2.12.so [.] _int_malloc
|
||
0.10% ruby libc-2.12.so [.] _int_free
|
||
```
|
||
|
||
Above you see sample output from a `perf` report. It shows that 97% of the CPU is
|
||
being spent inside Nokogiri and `xmlXPathNodeSetMergeAndClear`. For something
|
||
this obvious you should then go investigate what job in GitLab would use
|
||
Nokogiri and XPath. Combine with `TTIN` or `gdb` output to show the
|
||
corresponding Ruby code where this is happening.
|
||
|
||
## The GNU Project Debugger (`gdb`)
|
||
|
||
`gdb` can be another effective tool for debugging Sidekiq. It gives you a little
|
||
more interactive way to look at each thread and see what's causing problems.
|
||
|
||
Attaching to a process with `gdb` will suspends the normal operation
|
||
of the process (Sidekiq will not process jobs while `gdb` is attached).
|
||
|
||
Start by attaching to the Sidekiq PID:
|
||
|
||
```shell
|
||
gdb -p <sidekiq_pid>
|
||
```
|
||
|
||
Then gather information on all the threads:
|
||
|
||
```plaintext
|
||
info threads
|
||
|
||
# Example output
|
||
30 Thread 0x7fe5fbd63700 (LWP 26060) 0x0000003f7cadf113 in poll () from /lib64/libc.so.6
|
||
29 Thread 0x7fe5f2b3b700 (LWP 26533) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
28 Thread 0x7fe5f2a3a700 (LWP 26534) 0x0000003f7ce0ba5e in pthread_cond_timedwait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
27 Thread 0x7fe5f2939700 (LWP 26535) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
26 Thread 0x7fe5f2838700 (LWP 26537) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
25 Thread 0x7fe5f2737700 (LWP 26538) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
24 Thread 0x7fe5f2535700 (LWP 26540) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
23 Thread 0x7fe5f2434700 (LWP 26541) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
22 Thread 0x7fe5f2232700 (LWP 26543) 0x0000003f7ce0b68c in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
|
||
21 Thread 0x7fe5f2131700 (LWP 26544) 0x00007fe5f7b570f0 in xmlXPathNodeSetMergeAndClear ()
|
||
from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
...
|
||
```
|
||
|
||
If you see a suspicious thread, like the Nokogiri one above, you may want
|
||
to get more information:
|
||
|
||
```plaintext
|
||
thread 21
|
||
bt
|
||
|
||
# Example output
|
||
#0 0x00007ff0d6afe111 in xmlXPathNodeSetMergeAndClear () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#1 0x00007ff0d6b0b836 in xmlXPathNodeCollectAndTest () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#2 0x00007ff0d6b09037 in xmlXPathCompOpEval () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#3 0x00007ff0d6b09017 in xmlXPathCompOpEval () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#4 0x00007ff0d6b092e0 in xmlXPathCompOpEval () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#5 0x00007ff0d6b0bc37 in xmlXPathRunEval () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#6 0x00007ff0d6b0be5f in xmlXPathEvalExpression () from /opt/gitlab/embedded/service/gem/ruby/2.1.0/gems/nokogiri-1.6.7.2/lib/nokogiri/nokogiri.so
|
||
#7 0x00007ff0d6a97dc3 in evaluate (argc=2, argv=0x1022d058, self=<value optimized out>) at xml_xpath_context.c:221
|
||
#8 0x00007ff0daeab0ea in vm_call_cfunc_with_frame (th=0x1022a4f0, reg_cfp=0x1032b810, ci=<value optimized out>) at vm_insnhelper.c:1510
|
||
```
|
||
|
||
To output a backtrace from all threads at once:
|
||
|
||
```plaintext
|
||
set pagination off
|
||
thread apply all bt
|
||
```
|
||
|
||
Once you're done debugging with `gdb`, be sure to detach from the process and
|
||
exit:
|
||
|
||
```plaintext
|
||
detach
|
||
exit
|
||
```
|
||
|
||
## Sidekiq kill signals
|
||
|
||
TTIN was described above as the signal to print backtraces for logging, however
|
||
Sidekiq responds to other signals as well. For example, TSTP and TERM can be used
|
||
to gracefully shut Sidekiq down, see
|
||
[the Sidekiq Signals docs](https://github.com/mperham/sidekiq/wiki/Signals#ttin).
|
||
|
||
## Check for blocking queries
|
||
|
||
Sometimes the speed at which Sidekiq processes jobs can be so fast that it can
|
||
cause database contention. Check for blocking queries when backtraces above
|
||
show that many threads are stuck in the database adapter.
|
||
|
||
The PostgreSQL wiki has details on the query you can run to see blocking
|
||
queries. The query is different based on PostgreSQL version. See
|
||
[Lock Monitoring](https://wiki.postgresql.org/wiki/Lock_Monitoring) for
|
||
the query details.
|
||
|
||
## Managing Sidekiq queues
|
||
|
||
It is possible to use [Sidekiq API](https://github.com/mperham/sidekiq/wiki/API)
|
||
to perform a number of troubleshooting steps on Sidekiq.
|
||
|
||
These are the administrative commands and it should only be used if currently
|
||
admin interface is not suitable due to scale of installation.
|
||
|
||
All these commands should be run using `gitlab-rails console`.
|
||
|
||
### View the queue size
|
||
|
||
```ruby
|
||
Sidekiq::Queue.new("pipeline_processing:build_queue").size
|
||
```
|
||
|
||
### Enumerate all enqueued jobs
|
||
|
||
```ruby
|
||
queue = Sidekiq::Queue.new("chaos:chaos_sleep")
|
||
queue.each do |job|
|
||
# job.klass # => 'MyWorker'
|
||
# job.args # => [1, 2, 3]
|
||
# job.jid # => jid
|
||
# job.queue # => chaos:chaos_sleep
|
||
# job["retry"] # => 3
|
||
# job.item # => {
|
||
# "class"=>"Chaos::SleepWorker",
|
||
# "args"=>[1000],
|
||
# "retry"=>3,
|
||
# "queue"=>"chaos:chaos_sleep",
|
||
# "backtrace"=>true,
|
||
# "queue_namespace"=>"chaos",
|
||
# "jid"=>"39bc482b823cceaf07213523",
|
||
# "created_at"=>1566317076.266069,
|
||
# "correlation_id"=>"c323b832-a857-4858-b695-672de6f0e1af",
|
||
# "enqueued_at"=>1566317076.26761},
|
||
# }
|
||
|
||
# job.delete if job.jid == 'abcdef1234567890'
|
||
end
|
||
```
|
||
|
||
### Enumerate currently running jobs
|
||
|
||
```ruby
|
||
workers = Sidekiq::Workers.new
|
||
workers.each do |process_id, thread_id, work|
|
||
# process_id is a unique identifier per Sidekiq process
|
||
# thread_id is a unique identifier per thread
|
||
# work is a Hash which looks like:
|
||
# {"queue"=>"chaos:chaos_sleep",
|
||
# "payload"=>
|
||
# { "class"=>"Chaos::SleepWorker",
|
||
# "args"=>[1000],
|
||
# "retry"=>3,
|
||
# "queue"=>"chaos:chaos_sleep",
|
||
# "backtrace"=>true,
|
||
# "queue_namespace"=>"chaos",
|
||
# "jid"=>"b2a31e3eac7b1a99ff235869",
|
||
# "created_at"=>1566316974.9215662,
|
||
# "correlation_id"=>"e484fb26-7576-45f9-bf21-b99389e1c53c",
|
||
# "enqueued_at"=>1566316974.9229589},
|
||
# "run_at"=>1566316974}],
|
||
end
|
||
```
|
||
|
||
### Remove Sidekiq jobs for given parameters (destructive)
|
||
|
||
The general method to kill jobs conditionally is the following command, which
|
||
will remove jobs that are queued but not started. Running jobs will not be killed.
|
||
|
||
```ruby
|
||
queue = Sidekiq::Queue.new('<queue name>')
|
||
queue.each { |job| job.delete if <condition>}
|
||
```
|
||
|
||
Have a look at the section below for cancelling running jobs.
|
||
|
||
In the method above, `<queue-name>` is the name of the queue that contains the job(s) you want to delete and `<condition>` will decide which jobs get deleted.
|
||
|
||
Commonly, `<condition>` references the job arguments, which depend on the type of job in question. To find the arguments for a specific queue, you can have a look at the `perform` function of the related worker file, commonly found at `/app/workers/<queue-name>_worker.rb`.
|
||
|
||
For example, `repository_import` has `project_id` as the job argument, while `update_merge_requests` has `project_id, user_id, oldrev, newrev, ref`.
|
||
|
||
Arguments need to be referenced by their sequence ID using `job.args[<id>]` because `job.args` is a list of all arguments provided to the Sidekiq job.
|
||
|
||
Here are some examples:
|
||
|
||
```ruby
|
||
queue = Sidekiq::Queue.new('update_merge_requests')
|
||
# In this example, we want to remove any update_merge_requests jobs
|
||
# for the Project with ID 125 and ref `ref/heads/my_branch`
|
||
queue.each { |job| job.delete if job.args[0] == 125 and job.args[4] == 'ref/heads/my_branch' }
|
||
```
|
||
|
||
```ruby
|
||
# Cancelling jobs like: `RepositoryImportWorker.new.perform_async(100)`
|
||
id_list = [100]
|
||
|
||
queue = Sidekiq::Queue.new('repository_import')
|
||
queue.each do |job|
|
||
job.delete if id_list.include?(job.args[0])
|
||
end
|
||
```
|
||
|
||
### Remove specific job ID (destructive)
|
||
|
||
```ruby
|
||
queue = Sidekiq::Queue.new('repository_import')
|
||
queue.each do |job|
|
||
job.delete if job.jid == 'my-job-id'
|
||
end
|
||
```
|
||
|
||
## Canceling running jobs (destructive)
|
||
|
||
> Introduced in GitLab 12.3.
|
||
|
||
This is highly risky operation and use it as last resort.
|
||
Doing that might result in data corruption, as the job
|
||
is interrupted mid-execution and it is not guaranteed
|
||
that proper rollback of transactions is implemented.
|
||
|
||
```ruby
|
||
Gitlab::SidekiqDaemon::Monitor.cancel_job('job-id')
|
||
```
|
||
|
||
> This requires the Sidekiq to be run with `SIDEKIQ_MONITOR_WORKER=1`
|
||
> environment variable.
|
||
|
||
To perform of the interrupt we use `Thread.raise` which
|
||
has number of drawbacks, as mentioned in [Why Ruby’s Timeout is dangerous (and Thread.raise is terrifying)](https://jvns.ca/blog/2015/11/27/why-rubys-timeout-is-dangerous-and-thread-dot-raise-is-terrifying/):
|
||
|
||
> This is where the implications get interesting, and terrifying. This means that an exception can get raised:
|
||
>
|
||
> - during a network request (ok, as long as the surrounding code is prepared to catch Timeout::Error)
|
||
> - during the cleanup for the network request
|
||
> - during a rescue block
|
||
> - while creating an object to save to the database afterwards
|
||
> - in any of your code, regardless of whether it could have possibly raised an exception before
|
||
>
|
||
> Nobody writes code to defend against an exception being raised on literally any line. That’s not even possible. So Thread.raise is basically like a sneak attack on your code that could result in almost anything. It would probably be okay if it were pure-functional code that did not modify any state. But this is Ruby, so that’s unlikely :)
|