GitLab Application Service Level Indicators (SLIs)

History

It is possible to define Service Level Indicators(SLIs) directly in the Ruby codebase. This keeps the definition of operations and their success close to the implementation and allows the people building features to easily define how these features should be monitored.

Existing SLIs

  1. rails_request
  2. global_search_apdex
  3. global_search_error_rate
  4. global_search_indexing_apdex
  5. sidekiq_execution

Defining a new SLI

An SLI can be defined with the Gitlab::Metrics::Sli::Apdex or Gitlab::Metrics::Sli::ErrorRate class. When you define an SLI, two Prometheus counters are emitted from the Rails application. Both counters work in broadly the same way and contain a total operation count. Apdex uses a success rate to calculate a success ratio, and ErrorRate uses an error rate to calculate an error ratio.

The following metrics are defined:

  • Gitlab::Metrics::Sli::Apdex.new('foo') defines:
    • gitlab_sli_foo_apdex_total for the total number of measurements.
    • gitlab_sli_foo_apdex_success_total for the number of successful measurements.
  • Gitlab::Metrics::Sli::ErrorRate.new('foo') defines:
    • gitlab_sli_foo_total for the total number of measurements.
    • gitlab_sli_foo_error_total for the number of error measurements. Because this metric is an error rate, errors are divided by the total number.

As shown in this example, they can share a base name (foo in this example). We recommend this when they refer to the same operation.

You should use Apdex to measure the performance of successful operations. You don’t have to measure the performance of a failing request because that performance should be tracked with ErrorRate. For example, you can measure whether a request is performing within a specified latency threshold.

You should use ErrorRate to measure the rate of unsuccessful operations. For example, you can measure whether a failed request returns an HTTP status greater than or equal to 500.

Before the first scrape, it is important to have initialized the SLI with all possible label-combinations. This avoid confusing results when using these counters in calculations.

To initialize an SLI, use the .initialize_sli class method, for example:

Gitlab::Metrics::Sli::Apdex.initialize_sli(:received_email, [
  {
    feature_category: :team_planning,
    email_type: :create_issue
  },
  {
    feature_category: :service_desk,
    email_type: :service_desk
  },
  {
    feature_category: :code_review_workflow,
    email_type: :create_merge_request
  }
])

Metrics must be initialized before they get scraped for the first time. This currently happens during the on_master_start life-cycle event. Since this delays application readiness until metrics initialization returns, make sure the overhead this adds is understood and acceptable.

Tracking operations for an SLI

Tracking an operation in the newly defined SLI can be done like this:

Gitlab::Metrics::Sli::Apdex[:received_email].increment(
  labels: {
    feature_category: :service_desk,
    email_type: :service_desk
  },
  success: issue_created?
)

Calling #increment on this SLI will increment the total Prometheus counter

gitlab_sli:received_email_apdex:total{ feature_category='service_desk', email_type='service_desk' }

If the success: argument passed is truthy, then the success counter will also be incremented:

gitlab_sli:received_email_apdex:success_total{ feature_category='service_desk', email_type='service_desk' }

For error rate SLIs, the equivalent argument is called error::

Gitlab::Metrics::Sli::ErrorRate[:merge].increment(
  labels: {
    merge_type: :fast_forward
  },
  error: !merge_success?
)

Using the SLI in service monitoring and alerts

When the application is emitting metrics for a new SLI, they need to be consumed from the metrics catalog to result in alerts, and included in the error budget for stage groups and GitLab.com’s overall availability.

Start by adding the new SLI to the Application-SLI library. After that, add the following information:

  • name: the name of the SLI as defined in code. For example received_email.
  • significantLabels: an array of Prometheus labels that belong to the metrics. For example: ["email_type"]. If the significant labels for the SLI include feature_category, the metrics will also feed into the error budgets for stage groups.
  • featureCategory: if the SLI applies to a single feature category, you can specify it statically through this field to feed the SLI into the error budgets for stage groups.
  • description: a Markdown string explaining the SLI. It will be shown on dashboards and alerts.
  • kind: the kind of indicator. For example sliDefinition.apdexKind.

When done, run make generate to generate recording rules for the new SLI. This command creates recordings for all services emitting these metrics aggregated over significantLabels.

Open up a merge request with these changes and request review from a Scalability team member.

When these changes are merged, and the aggregations in Thanos recorded, query Thanos to see the success ratio of the new aggregated metrics. For example:

sum by (environment, stage, type)(application_sli_aggregation:rails_request:apdex:success:rate_1h)
/
sum by (environment, stage, type)(application_sli_aggregation:rails_request:apdex:weight:score_1h)

This shows the success ratio, which can guide you to set an appropriate SLO when adding this SLI to a service.

Then, add the SLI to the appropriate service catalog file. For example, the web service:

rails_requests:
  sliLibrary.get('rails_request_apdex')
    .generateServiceLevelIndicator({ job: 'gitlab-rails' })

To pass extra selectors and override properties of the SLI, see the service monitoring documentation.

SLIs with statically defined feature categories can already receive alerts about the SLI in specified Slack channels. For more information, read the alert routing documentation. In this project we are extending this so alerts for SLIs with a feature_category label in the source metrics can also be routed.

For any question, don’t hesitate to create an issue in the Scalability issue tracker or come find us in #g_scalability on Slack.