Profiling

To make it easier to track down performance problems GitLab comes with a set of profiling tools, some of these are available by default while others need to be explicitly enabled.

Profiling a URL

There is a Gitlab::Profiler.profile method, and corresponding bin/profile-url script, that enable profiling a GET or POST request to a specific URL, either as an anonymous user (the default) or as a specific user.

The first argument to the profiler is either a full URL (including the instance hostname) or an absolute path, including the leading slash.

By default the report dump will be stored in a temporary file, which can be interacted with using the Stackprof API.

When using the script, command-line documentation is available by passing no arguments.

When using the method in an interactive console session, any changes to the application code within that console session is reflected in the profiler output.

For example:

Gitlab::Profiler.profile('/my-user')
# Returns the location of the temp file where the report dump is stored
class UsersController; def show; sleep 100; end; end
Gitlab::Profiler.profile('/my-user')
# Returns the location of the temp file where the report dump is stored
# where 100 seconds is spent in UsersController#show

For routes that require authorization you must provide a user to Gitlab::Profiler. You can do this like so:

Gitlab::Profiler.profile('/gitlab-org/gitlab-test', user: User.first)

Passing a logger: keyword argument to Gitlab::Profiler.profile sends ActiveRecord and ActionController log output to that logger. Further options are documented with the method source.

Gitlab::Profiler.profile('/gitlab-org/gitlab-test', user: User.first, logger: Logger.new($stdout))

Pass in a profiler_options hash to configure the output file (out) of the sampling data. For example:

Gitlab::Profiler.profile('/gitlab-org/gitlab-test', user: User.first, profiler_options: { out: 'tmp/profile.dump' })

Reading a `GitLab::Profiler` report

You can get a summary of where time was spent by running Stackprof against the sampling data. For example:

stackprof tmp/profile.dump

Example sampling data:

==================================
  Mode: wall(1000)
  Samples: 8745 (6.92% miss rate)
  GC: 1399 (16.00%)
==================================
     TOTAL    (pct)     SAMPLES    (pct)     FRAME
      1022  (11.7%)        1022  (11.7%)     Sprockets::PathUtils#stat
       957  (10.9%)         957  (10.9%)     (marking)
       493   (5.6%)         493   (5.6%)     Sprockets::PathUtils#entries
       576   (6.6%)         471   (5.4%)     Mustermann::AST::Translator#decorator_for
       439   (5.0%)         439   (5.0%)     (sweeping)
       630   (7.2%)         241   (2.8%)     Sprockets::Cache::FileStore#get
       208   (2.4%)         208   (2.4%)     ActiveSupport::FileUpdateChecker#watched
       206   (2.4%)         206   (2.4%)     Digest::Instance#file
       544   (6.2%)         176   (2.0%)     Sprockets::Cache::FileStore#safe_open
       176   (2.0%)         176   (2.0%)     ActiveSupport::FileUpdateChecker#max_mtime
       268   (3.1%)         147   (1.7%)     ActiveRecord::ConnectionAdapters::PostgreSQLAdapter#exec_no_cache
       140   (1.6%)         140   (1.6%)     ActiveSupport::BacktraceCleaner#add_gem_filter
       116   (1.3%)         116   (1.3%)     Bootsnap::CompileCache::ISeq.storage_to_output
       160   (1.8%)         113   (1.3%)     Gem::Version#<=>
       109   (1.2%)         109   (1.2%)     block in <main>
       108   (1.2%)         108   (1.2%)     Gem::Version.new
       131   (1.5%)         105   (1.2%)     Sprockets::EncodingUtils#unmarshaled_deflated
      1166  (13.3%)          82   (0.9%)     Mustermann::RegexpBased#initialize
        82   (0.9%)          78   (0.9%)     FileUtils.touch
        72   (0.8%)          72   (0.8%)     Sprockets::Manifest.compile_match_filter
        71   (0.8%)          70   (0.8%)     Grape::Router#compile!
        91   (1.0%)          65   (0.7%)     ActiveRecord::ConnectionAdapters::PostgreSQL::DatabaseStatements#query
        93   (1.1%)          64   (0.7%)     ActionDispatch::Journey::Path::Pattern::AnchoredRegexp#accept
        59   (0.7%)          59   (0.7%)     Mustermann::AST::Translator.dispatch_table
        62   (0.7%)          59   (0.7%)     Rails::BacktraceCleaner#initialize
      2492  (28.5%)          49   (0.6%)     Sprockets::PathUtils#stat_directory
       242   (2.8%)          49   (0.6%)     Gitlab::Instrumentation::RedisBase.add_call_details
        47   (0.5%)          47   (0.5%)     URI::RFC2396_Parser#escape
        46   (0.5%)          46   (0.5%)     #<Class:0x00000001090c2e70>#__setobj__
        44   (0.5%)          44   (0.5%)     Sprockets::Base#normalize_logical_path

You can also generate flamegraphs:

stackprof --d3-flamegraph tmp/profile.dump > flamegraph.html

See the Stackprof documentation for more details.

Speedscope flamegraphs

You can generate a flamegraph for a particular URL by selecting a flamegraph sampling mode button in the performance bar or by adding the performance_bar=flamegraph parameter to the request.

Find more information about the views in the Speedscope docs.

Find more information about different sampling modes in the Stackprof docs.

This is enabled for all users that can access the performance bar.

Bullet

Bullet is a Gem that can be used to track down N+1 query problems. It logs query problems to the Rails log and the browser console. The Bullet section is displayed on the performance bar.

Bullet is enabled only in development mode by default. However, logging is disabled, because Bullet logging is noisy. To configure Bullet and its logging:

To manually enable or disable Bullet on an environment, add these lines to config/gitlab.yml, changing the enabled value as needed:
```
bullet:
  enabled: false
```
To enable Bullet logging, set the ENABLE_BULLET environment variable to a non-empty value before starting GitLab:
```
ENABLE_BULLET=true bundle exec rails s
```

As a follow-up to finding N+1 queries with Bullet, consider writing a QueryRecoder test to prevent a regression.

System stats

During or after profiling, you may want to get detailed information about the Ruby virtual machine process, such as memory consumption, time spent on CPU, or garbage collector statistics. These are easy to produce individually through various tools, but for convenience, a summary endpoint has been added that exports this data as a JSON payload:

curl localhost:3000/-/metrics/system | jq

Example output:

{
  "version": "ruby 2.7.2p137 (2020-10-01 revision a8323b79eb) [x86_64-linux-gnu]",
  "gc_stat": {
    "count": 118,
    "heap_allocated_pages": 11503,
    "heap_sorted_length": 11503,
    "heap_allocatable_pages": 0,
    "heap_available_slots": 4688580,
    "heap_live_slots": 3451712,
    "heap_free_slots": 1236868,
    "heap_final_slots": 0,
    "heap_marked_slots": 3451450,
    "heap_eden_pages": 11503,
    "heap_tomb_pages": 0,
    "total_allocated_pages": 11503,
    "total_freed_pages": 0,
    "total_allocated_objects": 32679478,
    "total_freed_objects": 29227766,
    "malloc_increase_bytes": 84760,
    "malloc_increase_bytes_limit": 32883343,
    "minor_gc_count": 88,
    "major_gc_count": 30,
    "compact_count": 0,
    "remembered_wb_unprotected_objects": 114228,
    "remembered_wb_unprotected_objects_limit": 228456,
    "old_objects": 3185330,
    "old_objects_limit": 6370660,
    "oldmalloc_increase_bytes": 21838024,
    "oldmalloc_increase_bytes_limit": 119181499
  },
  "memory_rss": 1326501888,
  "memory_uss": 1048563712,
  "memory_pss": 1139554304,
  "time_cputime": 82.885264633,
  "time_realtime": 1610459445.5579069,
  "time_monotonic": 24001.23145713,
  "worker_id": "puma_0"
}

NOTE: This endpoint is only available for Rails web workers. Sidekiq workers cannot be inspected this way.

Settings that impact performance

Application settings

development environment by default works with hot-reloading enabled, this makes Rails to check file changes every request, and create a potential contention lock, as hot reload is single threaded.
development environment can load code lazily once the request is fired which results in first request to always be slow.

To disable those features for profiling/benchmarking set the RAILS_PROFILE environment variable to true before starting GitLab. For example when using GDK:

create a file env.runit in the root GDK directory
add export RAILS_PROFILE=true to your env.runit file
restart GDK with gdk restart

This environment variable is only applicable for the development mode.

GC settings

Ruby's garbage collector (GC) can be tuned via a variety of environment variables that will directly impact application performance.

The following table lists these variables along with their default values.

Environment variable	Default value
`RUBY_GC_HEAP_INIT_SLOTS`	`10000`
`RUBY_GC_HEAP_FREE_SLOTS`	`4096`
`RUBY_GC_HEAP_FREE_SLOTS_MIN_RATIO`	`0.20`
`RUBY_GC_HEAP_FREE_SLOTS_GOAL_RATIO`	`0.40`
`RUBY_GC_HEAP_FREE_SLOTS_MAX_RATIO`	`0.65`
`RUBY_GC_HEAP_GROWTH_FACTOR`	`1.8`
`RUBY_GC_HEAP_GROWTH_MAX_SLOTS`	`0 (disable)`
`RUBY_GC_HEAP_OLDOBJECT_LIMIT_FACTOR`	`2.0`
`RUBY_GC_MALLOC_LIMIT(_MIN)`	`(16 * 1024 * 1024 /* 16MB */)`
`RUBY_GC_MALLOC_LIMIT_MAX`	`(32 * 1024 * 1024 /* 32MB */)`
`RUBY_GC_MALLOC_LIMIT_GROWTH_FACTOR`	`1.4`
`RUBY_GC_OLDMALLOC_LIMIT(_MIN)`	`(16 * 1024 * 1024 /* 16MB */)`
`RUBY_GC_OLDMALLOC_LIMIT_MAX`	`(128 * 1024 * 1024 /* 128MB */)`
`RUBY_GC_OLDMALLOC_LIMIT_GROWTH_FACTOR`	`1.2`

(Source)

GitLab may decide to change these settings to speed up application performance, lower memory requirements, or both.

You can see how each of these settings affect GC performance, memory use and application start-up time for an idle instance of GitLab by running the scripts/perf/gc/collect_gc_stats.rb script. It will output GC stats and general timing data to standard out as CSV.

An example of investigating performance issues

The Pipeline Authoring team has worked on solving the pipeline creation performance issues and used both the existing profiling methods such as stackprof flamegraphs and memory_profiler and a new method ruby-prof.

Using stackprof flamegraphs

Performance bar is a great tool to get a stackprof report and see a flamegraph via a single click;

However, it's not available for other than GET requests.

To get a flamegraph for a POST request, we use the performance_bar=flamegraph parameter with the API request. In our case, we want to see the flamegraph for the pipeline creation endpoint of a merge request.

Normally, we could use the following command to get a stackprof report as a JSON file but our user control of Gitlab::PerformanceBar.allowed_for_user?(request.env['warden']&.user) allows only users authenticated via the web interface.

# This will not work on production

curl --request POST \
  --output flamegraph.json \
  --header 'Content-Type: application/json' \
  --header 'PRIVATE-TOKEN: :token' \
  "https://gitlab.example.com/api/v4/projects/:id/merge_requests/:iid/pipelines?performance_bar=flamegraph"

To get around this, we copy the request as curl and use it in the terminal.

We'll have a curl command like this:

curl "https://gitlab.com/api/v4/projects/:id/merge_requests/:iid/pipelines" \
  -H 'accept: application/json, text/plain, */*' \
  -H 'content-type: application/json' \
  -H 'cookie: xyz' \
  -H 'x-csrf-token: xyz' \
  --data-raw '{"async":true}'

Notice the async parameter in the request body. We need to remove it to get the actual performance of the pipeline creation endpoint.
We need to add the performance_bar=flamegraph parameter to the request.
We need to add the --output flamegraph.json parameter to save the JSON response to a file.
Lastly, we need to accept the JSON response only.

curl "https://gitlab.com/api/v4/projects/:id/merge_requests/:iid/pipelines?performance_bar=flamegraph" \
  -X POST \
  -o flamegraph.json \
  -H 'accept: application/json' \
  -H 'content-type: application/json' \
  -H 'cookie: xyz' \
  -H 'x-csrf-token: xyz'

Then, we use the flamegraph.json file on the https://www.speedscope.app/ website to see the flamegraph.

As an example, when investigating into this speedscope flamegraph, we saw that the kubernetes_variables method was taking a lot of time and created an issue.

Using `ruby-prof`

Another method to see where to spend most of the time is to use ruby-prof. It's not an included gem in the Gemfile, so we need to add it to the Gemfile and run bundle install first.

To investigate the problem, we need to have a replica repository. To do this, we could mirror the repository from the production instance to the development instance. Then, we can run the ruby-prof profiler to see where the time is spent.

# RAILS_PROFILE=true GITALY_DISABLE_REQUEST_LIMITS=true rails console

require 'ruby-prof'

ActiveRecord::Base.logger = nil
project = Project.find_by_full_path('root/gitlab-mirror')
user = project.first_owner
merge_request = project.merge_requests.find_by_iid(1)

profile = RubyProf::Profile.new
profile.exclude_common_methods! # see https://github.com/ruby-prof/ruby-prof/blob/1.7.0/lib/ruby-prof/exclude_common_methods.rb

profile.start

Gitlab::SafeRequestStore.ensure_request_store do
  Ci::CreatePipelineService
    .new(project, user, ref: merge_request.source_branch)
    .execute(:merge_request_event, merge_request: merge_request)
    .payload
end; nil

result = profile.stop

callstack_printer = RubyProf::CallStackPrinter.new(result)
File.open('tmp/ruby-prof-callstack-report.html', 'w') do |file|
  callstack_printer.print(file)
end

::Ci::DestroyPipelineService.new(project, user).execute(Ci::Pipeline.last)

Here, we can see that we call Ci::GenerateKubeconfigService ~2k times. This is a good indicator that we need to investigate this.

Using `memory_profiler`

memory_profiler is a tool to profile memory usage. This is also important because high memory usage can lead to performance issues.

As we did with stackprof, we could also use curl with the performance_bar parameter.

curl "https://gitlab.com/api/v4/projects/:id/merge_requests/:iid/pipelines?performance_bar=memory" \
  -X POST \
  -o flamegraph.json \
  -H 'accept: application/json' \
  -H 'content-type: application/json' \
  -H 'cookie: xyz' \
  -H 'x-csrf-token: xyz'

However, this will not work on production because we have 60-second timeouts for the requests. So, we need to use the development environment to get the memory profile. More information can be found in the memory profiler documentation.

# RAILS_PROFILE=true GITALY_DISABLE_REQUEST_LIMITS=true rails console

require 'memory_profiler'

ActiveRecord::Base.logger = nil
project = Project.find_by_full_path('root/gitlab-mirror')
user = project.first_owner
merge_request = project.merge_requests.find_by_iid(1)

# Warmup
Ci::CreatePipelineService
  .new(project, user, ref: merge_request.source_branch)
  .execute(:merge_request_event, merge_request: merge_request); nil

report = MemoryProfiler.report do
  Gitlab::SafeRequestStore.ensure_request_store do
    Ci::CreatePipelineService
      .new(project, user, ref: merge_request.source_branch)
      .execute(:merge_request_event, merge_request: merge_request); nil
  end
end; nil

output = File.open('tmp/memory-profile-report.txt', 'w')
report.pretty_print(output, detailed_report: true, scale_bytes: true, normalize_paths: true)

Result;

#
# Note: I redacted some parts related to the gems and the Rails framework.
#       also, the output is shortened for readability.
#

Total allocated: 1.30 GB (12974240 objects)
Total retained:  29.67 MB (335085 objects)

allocated memory by gem
-----------------------------------
 675.48 MB  gitlab/lib

...

allocated memory by file
-----------------------------------
 253.68 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb
 143.58 MB  gitlab/lib/gitlab/ci/variables/collection.rb
  51.66 MB  gitlab/lib/gitlab/config/entry/configurable.rb
  20.89 MB  gitlab/lib/gitlab/ci/pipeline/expression/lexeme/base.rb

...

allocated memory by location
-----------------------------------
 107.12 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb:64
  70.22 MB  gitlab/lib/gitlab/ci/variables/collection.rb:28
  57.66 MB  gitlab/lib/gitlab/ci/variables/collection.rb:82
  45.70 MB  gitlab/lib/gitlab/config/entry/configurable.rb:67
  42.35 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb:17
  42.35 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb:80
  41.32 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb:76
  20.10 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb:72

...

In this example, we see where we can optimize the memory usage by looking at the allocated memory by file and location.

And in a recent work, we found a way to improve the memory usage and got this result;

#
# Note: I redacted some parts related to the gems and the Rails framework.
#       also, the output is shortened for readability.
#

Total allocated: 1.08 GB (11171148 objects)
Total retained:  29.67 MB (335082 objects)

allocated memory by gem
-----------------------------------
 495.88 MB  gitlab/lib

...

allocated memory by file
-----------------------------------
 112.44 MB  gitlab/lib/gitlab/ci/variables/collection.rb
 105.24 MB  gitlab/lib/gitlab/ci/variables/collection/item.rb
  51.66 MB  gitlab/lib/gitlab/config/entry/configurable.rb
  20.89 MB  gitlab/lib/gitlab/ci/pipeline/expression/lexeme/base.rb

...

Total memory reduction for this example pipeline; ~200 MB.