Value stream analytics development guide

Value stream analytics calculates the time between two arbitrary events recorded on domain objects and provides aggregated statistics about the duration.

For information on how to configure value stream analytics in GitLab, see our analytics documentation.


During development, events occur that move issues and merge requests through different stages of progress until they are considered finished. These stages can be expressed with the Stage model.

Example stage:

  • Name: Development
  • Start event: Issue created
  • End event: Issue first mentioned in commit
  • Parent: Group: gitlab-org


Events are the smallest building blocks of the value stream analytics feature. A stage consists of two events:

  • Start
  • End

These events play a key role in the duration calculation.

Formula: duration = end_event_time - start_event_time

To make the duration calculation flexible, each Event is implemented as a separate class. They're responsible for defining a timestamp expression that is used in the calculation query.

Implementing an Event class

There are a few methods that are required to be implemented, the StageEvent base class describes them in great detail. The most important ones are:

  • object_type
  • timestamp_projection

The object_type method defines which domain object is queried for the calculation. Currently two models are allowed:

  • Issue
  • MergeRequest

For the duration calculation the timestamp_projection method is used.

def timestamp_projection
  # your timestamp expression comes here

# event will use the issue creation time in the duration calculation
def timestamp_projection

More complex expressions are also possible (for example, using COALESCE). Review the existing event classes for examples.

In some cases, defining the timestamp_projection method is not enough. The calculation query should know which table contains the timestamp expression. Each Event class is responsible for making modifications to the calculation query to make the timestamp_projection work. This usually means joining an additional table.

Example for joining the issue_metrics table and using the first_mentioned_in_commit_at column as the timestamp expression:

def object_type

def timestamp_projection

def apply_query_customization(query)
  # in this case the query attribute will be based on the Issue model: `Issue.where(...)`

Validating start and end events

Some start/end event pairs are not "compatible" with each other. For example:

  • "Issue created" to "Merge Request created": The event classes are defined on different domain models, the object_type method is different.
  • "Issue closed" to "Issue created": Issue must be created first before it can be closed.
  • "Issue closed" to "Issue closed": Duration is always 0.

The StageEvents module describes the allowed start_event and end_event pairings (PAIRING_RULES constant). If a new event is added, it needs to be registered in this module. To add a new event:

  1. Add an entry in ENUM_MAPPING with a unique number, which is used in the Stage model as enum.
  2. Define which events are compatible with the event in the PAIRING_RULES hash.

Supported start/end event pairings:

graph LR;
  IssueCreated --> IssueClosed;
  IssueCreated --> IssueFirstAddedToBoard;
  IssueCreated --> IssueFirstAssociatedWithMilestone;
  IssueCreated --> IssueFirstMentionedInCommit;
  IssueCreated --> IssueLastEdited;
  IssueCreated --> IssueLabelAdded;
  IssueCreated --> IssueLabelRemoved;
  MergeRequestCreated --> MergeRequestMerged;
  MergeRequestCreated --> MergeRequestClosed;
  MergeRequestCreated --> MergeRequestFirstDeployedToProduction;
  MergeRequestCreated --> MergeRequestLastBuildStarted;
  MergeRequestCreated --> MergeRequestLastBuildFinished;
  MergeRequestCreated --> MergeRequestLastEdited;
  MergeRequestCreated --> MergeRequestLabelAdded;
  MergeRequestCreated --> MergeRequestLabelRemoved;
  MergeRequestLastBuildStarted --> MergeRequestLastBuildFinished;
  MergeRequestLastBuildStarted --> MergeRequestClosed;
  MergeRequestLastBuildStarted --> MergeRequestFirstDeployedToProduction;
  MergeRequestLastBuildStarted --> MergeRequestLastEdited;
  MergeRequestLastBuildStarted --> MergeRequestMerged;
  MergeRequestLastBuildStarted --> MergeRequestLabelAdded;
  MergeRequestLastBuildStarted --> MergeRequestLabelRemoved;
  MergeRequestMerged --> MergeRequestFirstDeployedToProduction;
  MergeRequestMerged --> MergeRequestClosed;
  MergeRequestMerged --> MergeRequestFirstDeployedToProduction;
  MergeRequestMerged --> MergeRequestLastEdited;
  MergeRequestMerged --> MergeRequestLabelAdded;
  MergeRequestMerged --> MergeRequestLabelRemoved;
  IssueLabelAdded --> IssueLabelAdded;
  IssueLabelAdded --> IssueLabelRemoved;
  IssueLabelAdded --> IssueClosed;
  IssueLabelRemoved --> IssueClosed;
  IssueFirstAddedToBoard --> IssueClosed;
  IssueFirstAddedToBoard --> IssueFirstAssociatedWithMilestone;
  IssueFirstAddedToBoard --> IssueFirstMentionedInCommit;
  IssueFirstAddedToBoard --> IssueLastEdited;
  IssueFirstAddedToBoard --> IssueLabelAdded;
  IssueFirstAddedToBoard --> IssueLabelRemoved;
  IssueFirstAssociatedWithMilestone --> IssueClosed;
  IssueFirstAssociatedWithMilestone --> IssueFirstAddedToBoard;
  IssueFirstAssociatedWithMilestone --> IssueFirstMentionedInCommit;
  IssueFirstAssociatedWithMilestone --> IssueLastEdited;
  IssueFirstAssociatedWithMilestone --> IssueLabelAdded;
  IssueFirstAssociatedWithMilestone --> IssueLabelRemoved;
  IssueFirstMentionedInCommit --> IssueClosed;
  IssueFirstMentionedInCommit --> IssueFirstAssociatedWithMilestone;
  IssueFirstMentionedInCommit --> IssueFirstAddedToBoard;
  IssueFirstMentionedInCommit --> IssueLastEdited;
  IssueFirstMentionedInCommit --> IssueLabelAdded;
  IssueFirstMentionedInCommit --> IssueLabelRemoved;
  IssueClosed --> IssueLastEdited;
  IssueClosed --> IssueLabelAdded;
  IssueClosed --> IssueLabelRemoved;
  MergeRequestClosed --> MergeRequestFirstDeployedToProduction;
  MergeRequestClosed --> MergeRequestLastEdited;
  MergeRequestClosed --> MergeRequestLabelAdded;
  MergeRequestClosed --> MergeRequestLabelRemoved;
  MergeRequestFirstDeployedToProduction --> MergeRequestLastEdited;
  MergeRequestFirstDeployedToProduction --> MergeRequestLabelAdded;
  MergeRequestFirstDeployedToProduction --> MergeRequestLabelRemoved;
  MergeRequestLastBuildFinished --> MergeRequestClosed;
  MergeRequestLastBuildFinished --> MergeRequestFirstDeployedToProduction;
  MergeRequestLastBuildFinished --> MergeRequestLastEdited;
  MergeRequestLastBuildFinished --> MergeRequestMerged;
  MergeRequestLastBuildFinished --> MergeRequestLabelAdded;
  MergeRequestLastBuildFinished --> MergeRequestLabelRemoved;
  MergeRequestLabelAdded --> MergeRequestLabelAdded;
  MergeRequestLabelAdded --> MergeRequestLabelRemoved;
  MergeRequestLabelRemoved --> MergeRequestLabelAdded;
  MergeRequestLabelRemoved --> MergeRequestLabelRemoved;


Teams and organizations might define their own way of building software, thus stages can be completely different. For each stage, a parent object needs to be defined.

Currently supported parents:

  • Project
  • Group

How parent relationship it work

  1. User navigates to the value stream analytics page.
  2. User selects a group.
  3. Backend loads the defined stages for the selected group.
  4. Additions and modifications to the stages are persisted within the selected group only.

Default stages

The original implementation of value stream analytics defined 7 stages. These stages are always available for each parent, however altering these stages is not possible.

To make things efficient and reduce the number of records created, the default stages are expressed as in-memory objects (not persisted). When the user creates a custom stage for the first time, all the stages are persisted. This behavior is implemented in the value stream analytics service objects.

The reason for this was that we'd like to add the abilities to hide and order stages later on.

Data Collector

DataCollector is the central point where the data is queried from the database. The class always operates on a single stage and consists of the following components:

  • BaseQueryBuilder:
    • Responsible for composing the initial query.
    • Deals with Stage specific configuration: events and their query customizations.
    • Parameters coming from the UI: date ranges.
  • Median: Calculates the median duration for a stage using the query from BaseQueryBuilder.
  • RecordsFetcher: Loads relevant records for a stage using the query from BaseQueryBuilder and specific Finder classes to apply visibility rules.
  • DataForDurationChart: Loads calculated durations with the finish time (end event timestamp) for the scatterplot chart.

For a new calculation or a query, implement it as a new method call in the DataCollector class.

Database query

Structure of the database query:

SELECT (customized by: Median or RecordsFetcher or DataForDurationChart)
FROM OBJECT_TYPE (Issue or MergeRequest)
INNER JOIN (several JOIN statements, depending on the events)
  (Filter by the PARENT model, example: filter Issues from Project A)
  (Date range filter based on the OBJECT_TYPE.created_at)
  (Check if the START_EVENT is earlier than END_EVENT, preventing negative duration)

Structure of the SELECT statement for Median:


Structure of the SELECT statement for DataForDurationChart:


High-level overview

  • Rails Controller (Analytics::CycleAnalytics module): Value stream analytics exposes its data via JSON endpoints, implemented within the analytics workspace. Configuring the stages are also implements JSON endpoints (CRUD).
  • Services (Analytics::CycleAnalytics module): All Stage related actions are delegated to respective service objects.
  • Models (Analytics::CycleAnalytics module): Models are used to persist the Stage objects ProjectStage and GroupStage.
  • Feature classes (Gitlab::Analytics::CycleAnalytics module):
    • Responsible for composing queries and define feature specific business logic.
    • DataCollector, Event, StageEvents, etc.


Since we have a lots of events and possible pairings, testing each pairing is not possible. The rule is to have at least one test case using an Event class.

Writing a test case for a stage using a new Event can be challenging since data must be created for both events. To make this a bit simpler, each test case must be implemented in the data_collector_spec.rb where the stage is tested through the DataCollector. Each test case is turned into multiple tests, covering the following cases:

  • Different parents: Group or Project
  • Different calculations: Median, RecordsFetcher or DataForDurationChart