Implement Service Ping

Service Ping consists of two kinds of data:

  • Counters: Track how often a certain event happened over time, such as how many CI/CD pipelines have run. They are monotonic and always trend up.
  • Observations: Facts collected from one or more GitLab instances and can carry arbitrary data. There are no general guidelines for how to collect those, due to the individual nature of that data.

To implement a new metric in Service Ping, follow these steps:

  1. Implement the required counter
  2. Name and place the metric
  3. Test counters manually using your Rails console
  4. Generate the SQL query
  5. Optimize queries with #database-lab
  6. Add the metric definition to the Metrics Dictionary
  7. Add the metric to the Versions Application
  8. Create a merge request
  9. Verify your metric
  10. Set up and test Service Ping locally

Instrumentation classes

NOTE: Implementing metrics directly in usage_data.rb is deprecated. When you add or change a Service Ping Metric, you must migrate metrics to instrumentation classes. For information about the progress on migrating Service Ping metrics, see this epic.

For example, we have the following instrumentation class: lib/gitlab/usage/metrics/instrumentations/count_boards_metric.rb.

You should add it to usage_data.rb as follows:

boards: add_metric('CountBoardsMetric', time_frame: 'all'),

Types of counters

There are several types of counters for metrics:

NOTE: Only use the provided counter methods. Each counter method contains a built-in fail-safe mechanism that isolates each counter to avoid breaking the entire Service Ping process.

Batch counters

For large tables, PostgreSQL can take a long time to count rows due to MVCC (Multi-version Concurrency Control). Batch counting is a counting method where a single large query is broken into multiple smaller queries. For example, instead of a single query querying 1,000,000 records, with batch counting, you can execute 100 queries of 10,000 records each. Batch counting is useful for avoiding database timeouts as each batch query is significantly shorter than one single long running query.

For GitLab.com, there are extremely large tables with 15 second query timeouts, so we use batch counting to avoid encountering timeouts. Here are the sizes of some GitLab.com tables:

Table Row counts in millions
merge_request_diff_commits 2280
ci_build_trace_sections 1764
merge_request_diff_files 1082
events 514

Batch counting requires indexes on columns to calculate max, min, and range queries. In some cases, you must add a specialized index on the columns involved in a counter.

Ordinary batch counters

Create a new database metrics instrumentation class with count operation for a given ActiveRecord_Relation

Method:

add_metric('CountIssuesMetric', time_frame: 'all')

Examples:

Examples using usage_data.rb have been deprecated. We recommend to use instrumentation classes.

Distinct batch counters

Create a new database metrics instrumentation class with distinct_count operation for a given ActiveRecord_Relation.

Method:

add_metric('CountUsersAssociatingMilestonesToReleasesMetric', time_frame: 'all')

WARNING: Counting over non-unique columns can lead to performance issues. For more information, see the iterating tables in batches guide.

Examples:

Examples using usage_data.rb have been deprecated. We recommend to use instrumentation classes.

Sum batch operation

There is no support for sum for database metrics.

Sum the values of a given ActiveRecord_Relation on given column and handles errors. Handles the ActiveRecord::StatementInvalid error

Method:

sum(relation, column, batch_size: nil, start: nil, finish: nil)

Arguments:

  • relation: the ActiveRecord_Relation to perform the operation
  • column: the column to sum on
  • batch_size: if none set it uses default value 1000 from Gitlab::Database::BatchCounter
  • start: custom start of the batch counting to avoid complex min calculations
  • end: custom end of the batch counting to avoid complex min calculations

Examples:

sum(JiraImportState.finished, :imported_issues_count)

Grouping and batch operations

The count, distinct_count, and sum batch counters can accept an ActiveRecord::Relation object, which groups by a specified column. With a grouped relation, the methods do batch counting, handle errors, and returns a hash table of key-value pairs.

Examples:

count(Namespace.group(:type))
# returns => {nil=>179, "Group"=>54}

distinct_count(Project.group(:visibility_level), :creator_id)
# returns => {0=>1, 10=>1, 20=>11}

sum(Issue.group(:state_id), :weight))
# returns => {1=>3542, 2=>6820}

Add operation

Sum the values given as parameters. Handles the StandardError. Returns -1 if any of the arguments are -1.

Method:

add(*args)

Examples:

project_imports = distinct_count(::Project.where.not(import_type: nil), :creator_id)
bulk_imports = distinct_count(::BulkImport, :user_id)

 add(project_imports, bulk_imports)

Estimated batch counters

Introduced in GitLab 13.7.

Estimated batch counter functionality handles ActiveRecord::StatementInvalid errors when used through the provided estimate_batch_distinct_count method. Errors return a value of -1.

WARNING: This functionality estimates a distinct count of a specific ActiveRecord_Relation in a given column, which uses the HyperLogLog algorithm. As the HyperLogLog algorithm is probabilistic, the results always include error. The highest encountered error rate is 4.9%.

When correctly used, the estimate_batch_distinct_count method enables efficient counting over columns that contain non-unique values, which can not be assured by other counters.

estimate_batch_distinct_count method

Method:

estimate_batch_distinct_count(relation, column = nil, batch_size: nil, start: nil, finish: nil)

The method includes the following arguments:

  • relation: The ActiveRecord_Relation to perform the count.
  • column: The column to perform the distinct count. The default is the primary key.
  • batch_size: From Gitlab::Database::PostgresHll::BatchDistinctCounter::DEFAULT_BATCH_SIZE. Default value: 10,000.
  • start: The custom start of the batch count, to avoid complex minimum calculations.
  • finish: The custom end of the batch count to avoid complex maximum calculations.

The method includes the following prerequisites:

  • The supplied relation must include the primary key defined as the numeric column. For example: id bigint NOT NULL.

  • The estimate_batch_distinct_count can handle a joined relation. To use its ability to count non-unique columns, the joined relation must not have a one-to-many relationship, such as has_many :boards.

  • Both start and finish arguments should always represent primary key relationship values, even if the estimated count refers to another column, for example:

      estimate_batch_distinct_count(::Note, :author_id, start: ::Note.minimum(:id), finish: ::Note.maximum(:id))

Examples:

  1. Simple execution of estimated batch counter, with only relation provided, returned value represents estimated number of unique values in id column (which is the primary key) of Project relation:

      estimate_batch_distinct_count(::Project)
  2. Execution of estimated batch counter, where provided relation has applied additional filter (.where(time_period)), number of unique values estimated in custom column (:author_id), and parameters: start and finish together apply boundaries that defines range of provided relation to analyze:

      estimate_batch_distinct_count(::Note.with_suggestions.where(time_period), :author_id, start: ::Note.minimum(:id), finish: ::Note.maximum(:id))
  3. Execution of estimated batch counter with joined relation (joins(:cluster)), for a custom column ('clusters.user_id'):

      estimate_batch_distinct_count(::Clusters::Applications::CertManager.where(time_period).available.joins(:cluster), 'clusters.user_id')

When instrumenting metric with usage of estimated batch counter please add _estimated suffix to its name, for example:

  "counts": {
    "ci_builds_estimated": estimate_batch_distinct_count(Ci::Build),
    ...

Redis counters

Handles ::Redis::CommandError and Gitlab::UsageDataCounters::BaseCounter::UnknownEvent. Returns -1 when a block is sent or hash with all values and -1 when a counter(Gitlab::UsageDataCounters) is sent. The different behavior is due to 2 different implementations of the Redis counter.

Method:

redis_usage_data(counter, &block)

Arguments:

  • counter: a counter from Gitlab::UsageDataCounters, that has fallback_totals method implemented
  • or a block: which is evaluated

Ordinary Redis counters

Examples of implementation:

UsageData API

You can use the UsageData API to track events. To track events, the usage_data_api feature flag must be enabled (set to default_enabled: true). Enabled by default in GitLab 13.7 and later.

UsageData API tracking
  1. Track events using the UsageData API.

    Increment event count using an ordinary Redis counter, for a given event name.

    API requests are protected by checking for a valid CSRF token.

    To increment the values, the related feature usage_data_<event_name> must be enabled.

    POST /usage_data/increment_counter
    Attribute Type Required Description
    event string yes The event name to track.

    Response:

    • 200 if the event was tracked.
    • 400 Bad request if the event parameter is missing.
    • 401 Unauthorized if the user is not authenticated.
    • 403 Forbidden if an invalid CSRF token is provided.
  2. Track events using the JavaScript/Vue API helper which calls the UsageData API.

    To track events, usage_data_api and usage_data_#{event_name} must be enabled.

    import api from '~/api';
    
    api.trackRedisCounterEvent('my_already_defined_event_name'),

Redis HLL counters

WARNING: HyperLogLog (HLL) is a probabilistic algorithm and its results always includes some small error. According to Redis documentation, data from used HLL implementation is "approximated with a standard error of 0.81%".

NOTE: A user's consent for usage_stats (User.single_user&.requires_usage_stats_consent?) is not checked during the data tracking stage due to performance reasons. Keys corresponding to those counters are present in Redis even if usage_stats_consent is still required. However, no metric is collected from Redis and reported back to GitLab as long as usage_stats_consent is required.

With Gitlab::UsageDataCounters::HLLRedisCounter we have available data structures used to count unique values.

Implemented using Redis methods PFADD and PFCOUNT.

Add new events
  1. Define events in known_events.

    Example event:

    - name: users_creating_epics
      category: epics_usage
      redis_slot: users
      aggregation: weekly
      feature_flag: track_epics_activity

    Keys:

    • name: unique event name.

      Name format for Redis HLL events <name>_<redis_slot>.

      See Metric name for a complete guide on metric naming suggestion.

      Consider including in the event's name the Redis slot to be able to count totals for a specific category.

      Example names: users_creating_epics, users_triggering_security_scans.

    • category: event category. Used for getting total counts for events in a category, for easier access to a group of events.

    • redis_slot: optional Redis slot. Default value: event name. Only event data that is stored in the same slot can be aggregated. Ensure keys are in the same slot. For example: users_creating_epics with redis_slot: 'users' builds Redis key {users}_creating_epics-2020-34. If redis_slot is not defined the Redis key will be {users_creating_epics}-2020-34. Recommended slots to use are: users, projects. This is the value we count.

    • expiry: expiry time in days. Default: 29 days for daily aggregation and 6 weeks for weekly aggregation.

    • aggregation: may be set to a :daily or :weekly key. Defines how counting data is stored in Redis. Aggregation on a daily basis does not pull more fine grained data.

    • feature_flag: optional default_enabled: :yaml. If no feature flag is set then the tracking is enabled. One feature flag can be used for multiple events. For details, see our GitLab internal Feature flags documentation. The feature flags are owned by the group adding the event tracking.

  2. Use one of the following methods to track the event:

    • In the controller using the RedisTracking module and the following format:

      track_redis_hll_event(*controller_actions, name:, if: nil, &block)

      Arguments:

      • controller_actions: the controller actions to track.
      • name: the event name.
      • if: optional custom conditions. Uses the same format as Rails callbacks.
      • &block: optional block that computes and returns the custom_id that we want to track. This overrides the visitor_id.

      Example:

      # controller
      class ProjectsController < Projects::ApplicationController
        include RedisTracking
      
        skip_before_action :authenticate_user!, only: :show
        track_redis_hll_event :index, :show, name: 'users_visiting_projects'
      
        def index
          render html: 'index'
        end
      
       def new
         render html: 'new'
       end
      
       def show
         render html: 'show'
       end
      end
    • In the API using the increment_unique_values(event_name, values) helper method.

      Arguments:

      • event_name: the event name.
      • values: the values counted. Can be one value or an array of values.

      Example:

      get ':id/registry/repositories' do
        repositories = ContainerRepositoriesFinder.new(
          user: current_user, subject: user_group
        ).execute
      
        increment_unique_values('users_listing_repositories', current_user.id)
      
        present paginate(repositories), with: Entities::ContainerRegistry::Repository, tags: params[:tags], tags_count: params[:tags_count]
      end
    • Using track_usage_event(event_name, values) in services and GraphQL.

      Increment unique values count using Redis HLL, for a given event name.

      Examples:

        track_usage_event(:incident_management_incident_created, current_user.id)
    • Using the UsageData API.

      Increment unique users count using Redis HLL, for a given event name.

      API requests are protected by checking for a valid CSRF token.

      POST /usage_data/increment_unique_users
      Attribute Type Required Description
      event string yes The event name to track

      Response:

      • 200 if the event was tracked, or if tracking failed for any reason.
      • 400 Bad request if an event parameter is missing.
      • 401 Unauthorized if the user is not authenticated.
      • 403 Forbidden if an invalid CSRF token is provided.
    • Using the JavaScript/Vue API helper, which calls the UsageData API.

      Example for an existing event already defined in known events:

      import api from '~/api';
      
      api.trackRedisHllUserEvent('my_already_defined_event_name'),
  3. Get event data using Gitlab::UsageDataCounters::HLLRedisCounter.unique_events(event_names:, start_date:, end_date:, context: '').

    Arguments:

    • event_names: the list of event names.
    • start_date: start date of the period for which we want to get event data.
    • end_date: end date of the period for which we want to get event data.
    • context: context of the event. Allowed values are default, free, bronze, silver, gold, starter, premium, ultimate.
  4. Testing tracking and getting unique events

Trigger events in rails console by using track_event method

Gitlab::UsageDataCounters::HLLRedisCounter.track_event('users_viewing_compliance_audit_events', values: 1)
Gitlab::UsageDataCounters::HLLRedisCounter.track_event('users_viewing_compliance_audit_events', values: [2, 3])

Next, get the unique events for the current week.

# Get unique events for metric for current_week
Gitlab::UsageDataCounters::HLLRedisCounter.unique_events(event_names: 'users_viewing_compliance_audit_events',
start_date: Date.current.beginning_of_week, end_date: Date.current.next_week)
Recommendations

We have the following recommendations for adding new events:

  • Event aggregation: weekly.
  • Key expiry time:
    • Daily: 29 days.
    • Weekly: 42 days.
  • When adding new metrics, use a feature flag to control the impact.
  • For feature flags triggered by another service, set default_enabled: false,
    • Events can be triggered using the UsageData API, which helps when there are > 10 events per change
Enable or disable Redis HLL tracking

Events are tracked behind optional feature flags due to concerns for Redis performance and scalability.

For a full list of events and corresponding feature flags see, known_events files.

To enable or disable tracking for specific event in https://gitlab.com or https://about.staging.gitlab.com, run commands such as the following to enable or disable the corresponding feature.

/chatops run feature set <feature_name> true
/chatops run feature set <feature_name> false

We can also disable tracking completely by using the global flag:

/chatops run feature set redis_hll_tracking true
/chatops run feature set redis_hll_tracking false
Known events are added automatically in Service Data payload

Service Ping adds all events known_events/*.yml to Service Data generation under the redis_hll_counters key. This column is stored in version-app as a JSON. For each event we add metrics for the weekly and monthly time frames, and totals for each where applicable:

  • #{event_name}_weekly: Data for 7 days for daily aggregation events and data for the last complete week for weekly aggregation events.
  • #{event_name}_monthly: Data for 28 days for daily aggregation events and data for the last 4 complete weeks for weekly aggregation events.

Redis HLL implementation calculates total metrics when both of these conditions are met:

  • The category is manually included in CATEGORIES_FOR_TOTALS.
  • There is more than one metric for the same category, aggregation, and Redis slot.

We add total unique counts for the weekly and monthly time frames where applicable:

  • #{category}_total_unique_counts_weekly: Total unique counts for events in the same category for the last 7 days or the last complete week, if events are in the same Redis slot and we have more than one metric.
  • #{category}_total_unique_counts_monthly: Total unique counts for events in same category for the last 28 days or the last 4 complete weeks, if events are in the same Redis slot and we have more than one metric.

Example of redis_hll_counters data:

{:redis_hll_counters=>
  {"compliance"=>
    {"users_viewing_compliance_dashboard_weekly"=>0,
     "users_viewing_compliance_dashboard_monthly"=>0,
     "users_viewing_compliance_audit_events_weekly"=>0,
     "users_viewing_audit_events_monthly"=>0,
     "compliance_total_unique_counts_weekly"=>0,
     "compliance_total_unique_counts_monthly"=>0},
 "analytics"=>
    {"users_viewing_analytics_group_devops_adoption_weekly"=>0,
     "users_viewing_analytics_group_devops_adoption_monthly"=>0,
     "analytics_total_unique_counts_weekly"=>0,
     "analytics_total_unique_counts_monthly"=>0},
   "ide_edit"=>
    {"users_editing_by_web_ide_weekly"=>0,
     "users_editing_by_web_ide_monthly"=>0,
     "users_editing_by_sfe_weekly"=>0,
     "users_editing_by_sfe_monthly"=>0,
     "ide_edit_total_unique_counts_weekly"=>0,
     "ide_edit_total_unique_counts_monthly"=>0}
 }

Example:

# Redis Counters
redis_usage_data(Gitlab::UsageDataCounters::WikiPageCounter)
redis_usage_data { ::Gitlab::UsageCounters::PodLogs.usage_totals[:total] }

# Define events in common.yml https://gitlab.com/gitlab-org/gitlab/-/blob/master/lib/gitlab/usage_data_counters/known_events/common.yml

# Tracking events
Gitlab::UsageDataCounters::HLLRedisCounter.track_event('users_expanding_vulnerabilities', values: visitor_id)

# Get unique events for metric
redis_usage_data { Gitlab::UsageDataCounters::HLLRedisCounter.unique_events(event_names: 'users_expanding_vulnerabilities', start_date: 28.days.ago, end_date: Date.current) }

Alternative counters

Handles StandardError and fallbacks into -1 this way not all measures fail if we encounter one exception. Mainly used for settings and configurations.

Method:

alt_usage_data(value = nil, fallback: -1, &block)

Arguments:

  • value: a simple static value in which case the value is simply returned.
  • or a block: which is evaluated
  • fallback: -1: the common value used for any metrics that are failing.

Example:

alt_usage_data { Gitlab::VERSION }
alt_usage_data { Gitlab::CurrentSettings.uuid }
alt_usage_data(999)

Add counters to build new metrics

When adding the results of two counters, use the add Service Data method that handles fallback values and exceptions. It also generates a valid SQL export.

Example:

add(User.active, User.bot)

Prometheus queries

In those cases where operational metrics should be part of Service Ping, a database or Redis query is unlikely to provide useful data. Instead, Prometheus might be more appropriate, because most GitLab architectural components publish metrics to it that can be queried back, aggregated, and included as Service Data.

NOTE: Prometheus as a data source for Service Ping is only available for single-node Omnibus installations that are running the bundled Prometheus instance.

To query Prometheus for metrics, a helper method is available to yield a fully configured PrometheusClient, given it is available as per the note above:

with_prometheus_client do |client|
  response = client.query('<your query>')
  ...
end

Refer to the PrometheusClient definition for how to use its API to query for data.

Fallback values for Service Ping

We return fallback values in these cases:

Case Value
Deprecated Metric -1000
Timeouts, general failures -1
Standard errors in counters -2

Test counters manually using your Rails console

# count
Gitlab::UsageData.count(User.active)
Gitlab::UsageData.count(::Clusters::Cluster.aws_installed.enabled, :cluster_id)

# count distinct
Gitlab::UsageData.distinct_count(::Project, :creator_id)
Gitlab::UsageData.distinct_count(::Note.with_suggestions.where(time_period), :author_id, start: ::User.minimum(:id), finish: ::User.maximum(:id))

Generate the SQL query

Your Rails console returns the generated SQL queries. For example:

pry(main)> Gitlab::UsageData.count(User.active)
   (2.6ms)  SELECT "features"."key" FROM "features"
   (15.3ms)  SELECT MIN("users"."id") FROM "users" WHERE ("users"."state" IN ('active')) AND ("users"."user_type" IS NULL OR "users"."user_type" IN (6, 4))
   (2.4ms)  SELECT MAX("users"."id") FROM "users" WHERE ("users"."state" IN ('active')) AND ("users"."user_type" IS NULL OR "users"."user_type" IN (6, 4))
   (1.9ms)  SELECT COUNT("users"."id") FROM "users" WHERE ("users"."state" IN ('active')) AND ("users"."user_type" IS NULL OR "users"."user_type" IN (6, 4)) AND "users"."id" BETWEEN 1 AND 100000

Optimize queries with #database-lab

#database-lab is a Slack channel that uses a production-sized environment to test your queries. Paste the SQL query into #database-lab to see how the query performs at scale.

  • GitLab.com's production database has a 15 second timeout.
  • Any single query must stay below the 1 second execution time with cold caches.
  • Add a specialized index on columns involved to reduce the execution time.

To understand the query's execution, we add the following information to a merge request description:

  • For counters that have a time_period test, we add information for both:
    • time_period = {} for all time periods.
    • time_period = { created_at: 28.days.ago..Time.current } for the last 28 days.
  • Execution plan and query time before and after optimization.
  • Query generated for the index and time.
  • Migration output for up and down execution.

We also use #database-lab and explain.depesz.com. For more details, see the database review guide.

Optimization recommendations and examples

Add the metric definition

See the Metrics Dictionary guide for more information.

Add the metric to the Versions Application

Check if the new metric must be added to the Versions Application. See the usage_data schema and Service Data parameters accepted. Any metrics added under the counts key are saved in the stats column.

Create a merge request

Create a merge request for the new Service Ping metric, and do the following:

  • Add the feature label to the merge request. A metric is a user-facing change and is part of expanding the Service Ping feature.
  • Add a changelog entry that complies with the changelog entries guide.
  • Ask for a Product Intelligence review. On GitLab.com, we have DangerBot set up to monitor Product Intelligence related files and recommend a Product Intelligence review.

Verify your metric

On GitLab.com, the Product Intelligence team regularly monitors Service Ping. They may alert you that your metrics need further optimization to run quicker and with greater success.

The Service Ping JSON payload for GitLab.com is shared in the #g_product_intelligence Slack channel every week.

You may also use the Service Ping QA dashboard to check how well your metric performs. The dashboard allows filtering by GitLab version, by "Self-managed" and "SaaS", and shows you how many failures have occurred for each metric. Whenever you notice a high failure rate, you can re-optimize your metric.

Use Metrics Dictionary copy query to clipboard feature to get a query ready to run in Sisense for a specific metric.

Set up and test Service Ping locally

To set up Service Ping locally, you must:

  1. Set up local repositories.
  2. Test local setup.
  3. Optional. Test Prometheus-based Service Ping.

Set up local repositories

  1. Clone and start GitLab.
  2. Clone and start Versions Application. Make sure you run docker-compose up to start a PostgreSQL and Redis instance.
  3. Point GitLab to the Versions Application endpoint instead of the default endpoint:
    1. Open service_ping/submit_service.rb locally and modify STAGING_BASE_URL.
    2. Set it to the local Versions Application URL: http://localhost:3000/usage_data.

Test local setup

  1. Using the gitlab Rails console, manually trigger Service Ping:

    ServicePing::SubmitService.new.execute
  2. Use the versions Rails console to check the Service Ping was successfully received, parsed, and stored in the Versions database:

    UsageData.last

Test Prometheus-based Service Ping

If the data submitted includes metrics queried from Prometheus you want to inspect and verify, you must:

  • Ensure that a Prometheus server is running locally.
  • Ensure the respective GitLab components are exporting metrics to the Prometheus server.

If you do not need to test data coming from Prometheus, no further action is necessary. Service Ping should degrade gracefully in the absence of a running Prometheus server.

Three kinds of components may export data to Prometheus, and are included in Service Ping:

  • node_exporter: Exports node metrics from the host machine.
  • gitlab-exporter: Exports process metrics from various GitLab components.
  • Other various GitLab services, such as Sidekiq and the Rails server, which export their own metrics.

Test with an Omnibus container

This is the recommended approach to test Prometheus-based Service Ping.

To verify your change, build a new Omnibus image from your code branch using CI/CD, download the image, and run a local container instance:

  1. From your merge request, select the qa stage, then trigger the package-and-qa job. This job triggers an Omnibus build in a downstream pipeline of the omnibus-gitlab-mirror project.
  2. In the downstream pipeline, wait for the gitlab-docker job to finish.
  3. Open the job logs and locate the full container name including the version. It takes the following form: registry.gitlab.com/gitlab-org/build/omnibus-gitlab-mirror/gitlab-ee:<VERSION>.
  4. On your local machine, make sure you are signed in to the GitLab Docker registry. You can find the instructions for this in Authenticate to the GitLab Container Registry.
  5. Once signed in, download the new image by using docker pull registry.gitlab.com/gitlab-org/build/omnibus-gitlab-mirror/gitlab-ee:<VERSION>
  6. For more information about working with and running Omnibus GitLab containers in Docker, refer to GitLab Docker images in the Omnibus documentation.

Test with GitLab development toolkits

This is the less recommended approach, because it comes with a number of difficulties when emulating a real GitLab deployment.

The GDK is not set up to run a Prometheus server or node_exporter alongside other GitLab components. If you would like to do so, Monitoring the GDK with Prometheus is a good start.

The GCK has limited support for testing Prometheus based Service Ping. By default, it comes with a fully configured Prometheus service that is set up to scrape a number of components. However, it has the following limitations:

  • It does not run a gitlab-exporter instance, so several process_* metrics from services such as Gitaly may be missing.
  • While it runs a node_exporter, docker-compose services emulate hosts, meaning that it normally reports itself as not associated with any of the other running services. That is not how node metrics are reported in a production setup, where node_exporter always runs as a process alongside other GitLab components on any given node. For Service Ping, none of the node data would therefore appear to be associated to any of the services running, because they all appear to be running on different hosts. To alleviate this problem, the node_exporter in GCK was arbitrarily "assigned" to the web service, meaning only for this service node_* metrics appears in Service Ping.

Aggregated metrics

WARNING: This feature is intended solely for internal GitLab use.

To add data for aggregated metrics to the Service Ping payload, add a corresponding definition to:

Each aggregate definition includes following parts:

  • name: Unique name under which the aggregate metric is added to the Service Ping payload.
  • operator: Operator that defines how the aggregated metric data is counted. Available operators are:
    • OR: Removes duplicates and counts all entries that triggered any of listed events.
    • AND: Removes duplicates and counts all elements that were observed triggering all of following events.
  • time_frame: One or more valid time frames. Use these to limit the data included in aggregated metric to events within a specific date-range. Valid time frames are:
    • 7d: Last seven days of data.
    • 28d: Last twenty eight days of data.
    • all: All historical data, only available for database sourced aggregated metrics.
  • source: Data source used to collect all events data included in aggregated metric. Valid data sources are:
  • events: list of events names to aggregate into metric. All events in this list must relay on the same data source. Additional data source requirements are described in the Database sourced aggregated metrics and Redis sourced aggregated metrics sections.
  • feature_flag: Name of development feature flag that is checked before metrics aggregation is performed. Corresponding feature flag should have default_enabled attribute set to false. The feature_flag attribute is optional and can be omitted. When feature_flag is missing, no feature flag is checked.

Example aggregated metric entries:

- name: example_metrics_union
  operator: OR
  events:
    - 'users_expanding_secure_security_report'
    - 'users_expanding_testing_code_quality_report'
    - 'users_expanding_testing_accessibility_report'
  source: redis
  time_frame:
    - 7d
    - 28d
- name: example_metrics_intersection
  operator: AND
  source: database
  time_frame:
    - 28d
    - all
  events:
    - 'dependency_scanning_pipeline_all_time'
    - 'container_scanning_pipeline_all_time'
  feature_flag: example_aggregated_metric

Aggregated metrics collected in 7d and 28d time frames are added into Service Ping payload under the aggregated_metrics sub-key in the counts_weekly and counts_monthly top level keys.

{
  :counts_monthly => {
    :deployments => 1003,
    :successful_deployments => 78,
    :failed_deployments => 275,
    :packages => 155,
    :personal_snippets => 2106,
    :project_snippets => 407,
    :aggregated_metrics => {
      :example_metrics_union => 7,
      :example_metrics_intersection => 2
    },
    :snippets => 2513
  }
}

Aggregated metrics for all time frame are present in the count top level key, with the aggregate_ prefix added to their name.

For example:

example_metrics_intersection

Becomes:

counts.aggregate_example_metrics_intersection

{
  :counts => {
    :deployments => 11003,
    :successful_deployments => 178,
    :failed_deployments => 1275,
    :aggregate_example_metrics_intersection => 12
  }
}

Redis sourced aggregated metrics

Introduced in GitLab 13.6.

To declare the aggregate of events collected with Redis HLL Counters, you must fulfill the following requirements:

  1. All events listed at events attribute must come from known_events/*.yml files.
  2. All events listed at events attribute must have the same redis_slot attribute.
  3. All events listed at events attribute must have the same aggregation attribute.
  4. time_frame does not include all value, which is unavailable for Redis sourced aggregated metrics.

Database sourced aggregated metrics

To declare an aggregate of metrics based on events collected from database, follow these steps:

  1. Persist the metrics for aggregation.
  2. Add new aggregated metric definition.

Persist metrics for aggregation

Only metrics calculated with Estimated Batch Counters can be persisted for database sourced aggregated metrics. To persist a metric, inject a Ruby block into the estimate_batch_distinct_count method. This block should invoke the Gitlab::Usage::Metrics::Aggregates::Sources::PostgresHll.save_aggregated_metrics method, which stores estimate_batch_distinct_count results for future use in aggregated metrics.

The Gitlab::Usage::Metrics::Aggregates::Sources::PostgresHll.save_aggregated_metrics method accepts the following arguments:

  • metric_name: The name of metric to use for aggregations. Should be the same as the key under which the metric is added into Service Ping.
  • recorded_at_timestamp: The timestamp representing the moment when a given Service Ping payload was collected. You should use the convenience method recorded_at to fill recorded_at_timestamp argument, like this: recorded_at_timestamp: recorded_at
  • time_period: The time period used to build the relation argument passed into estimate_batch_distinct_count. To collect the metric with all available historical data, set a nil value as time period: time_period: nil.
  • data: HyperLogLog buckets structure representing unique entries in relation. The estimate_batch_distinct_count method always passes the correct argument into the block, so data argument must always have a value equal to block argument, like this: data: result

Example metrics persistence:

class UsageData
  def count_secure_pipelines(time_period)
    ...
    relation = ::Security::Scan.by_scan_types(scan_type).where(time_period)

    pipelines_with_secure_jobs['dependency_scanning_pipeline'] = estimate_batch_distinct_count(relation, :pipeline_id, batch_size: 1000, start: start_id, finish: finish_id) do |result|
      ::Gitlab::Usage::Metrics::Aggregates::Sources::PostgresHll
        .save_aggregated_metrics(metric_name: 'dependency_scanning_pipeline', recorded_at_timestamp: recorded_at, time_period: time_period, data: result)
    end
  end
end

Add new aggregated metric definition

After all metrics are persisted, you can add an aggregated metric definition at aggregated_metrics/.

To declare the aggregate of metrics collected with Estimated Batch Counters, you must fulfill the following requirements:

  • Metrics names listed in the events: attribute, have to use the same names you passed in the metric_name argument while persisting metrics in previous step.
  • Every metric listed in the events: attribute, has to be persisted for every selected time_frame: value.

Example definition:

- name: example_metrics_intersection_database_sourced
  operator: AND
  source: database
  events:
    - 'dependency_scanning_pipeline'
    - 'container_scanning_pipeline'
  time_frame:
    - 28d
    - all