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.

Note: 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.

Note: 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

If you want to see what arguments are being passed to Sidekiq jobs you can set the SIDEKIQ_LOG_ARGUMENTS environment variable to 1 (true).

Example:

gitlab_rails['env'] = {"SIDEKIQ_LOG_ARGUMENTS" => "1"}

Please note: It is not recommend to enable this setting in production because some Sidekiq jobs (such as sending a password reset email) take secret arguments (for example the password reset token).

When using Sidekiq JSON logging, 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 "...".

Thread dump

Send the Sidekiq process ID the TTIN signal and it will output thread backtraces in the log file.

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:

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:

# 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:

sudo perf record -p <sidekiq_pid>

Let this run for 30-60 seconds and then press Ctrl-C. Then view the perf report:

$ 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.

NOTE: Note: 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:

gdb -p <sidekiq_pid>

Then gather information on all the threads:

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:

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:

set pagination off
thread apply all bt

Once you're done debugging with gdb, be sure to detach from the process and exit:

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.

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 for the query details.

Managing Sidekiq queues

It is possible to use Sidekiq API to perform a number of troubleshooting 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 this commands should be run using gitlab-rails console.

View the queue size

Sidekiq::Queue.new("pipeline_processing:build_queue").size

Enumerate all enqueued jobs

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

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:

queue = Sidekiq::Queue.new('<queue name>')
queue.each { |job| job.delete if <condition>}

NOTE: Note: This will remove jobs that are queued but not started, running jobs will not be killed. 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.

NOTE: Note: 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:

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' }
# 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)

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.

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):

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 :)