Cocoon Scheduler

This Dart project contains logic for constructing infrastructure configs to validate commits in the repositories owned by Flutter.

ci.yaml

This is the config file in a repository used to tell Cocoon what tasks are used to validate commits. It includes both the tasks used in presubmit and postsubmit.

In addition, it supports tasks from different infrastructures as long as cocoon supports that scheduler. Only luci and cocoon are supported, but contributions are welcome.

Example config:

# /.ci.yaml

# Enabled branches is a list of regexes, with the assumption that these are full line matches.
# Internally, Cocoon prefixes these with $ and suffixes with ^ to enable matches.
enabled_branches:
  - main
  - flutter-\\d+\\.\\d+-candidate\\.\\d+

# Platform properties defines common properties shared among targets from the same platform.
platform_properties:
  linux:
    properties:
      # os will be inherited by all Linux targets, but it can be overrided at the target level
      os: Linux

targets:
# A Target is an individual unit of work that is scheduled by Flutter infra
# Target's are composed of the following properties:
# name: A human readable string to uniquely identify this target.
#       The first word indicates the platform this test will be run on. This should match
#       to an existing platform under platform_properties.
# recipes: LUCI recipes the target follows to run tests
#          https://flutter.googlesource.com/recipes/+/refs/heads/main/recipes/
# bringup: Whether this target is under active development and should not block the tree.
#          If true, will not run in presubmit and will not block postsubmit.
# presubmit: Whether to run this target on presubmit (defaults to true).
# postsubmit: Whether to run this target on postsubmit (defaults to true).
# run_if: List of path regexes that can trigger this target on presubmit.
#         If none are passed, will always run in presubmit.
# enabled_branches: List of strings of branches this target can run on.
#                   This overrides the global enabled_branches.
# properties: A map of string, string. Values are parsed to their closest data model.
# postsubmit_properties: Properties that are only run on postsubmit.
#
# Minimal example:
# Linux analyze will run on all presubmit and in postsubmit.
 - name: Linux analyze
#
# Bringup example:
# Linux licenses will run on postsubmit, but it also passes the properties
# `analyze=true` to the builder. Since `bringup=true`, presubmit is not run,
# and postsubmit runs will not block the tree.
 - name: Linux licenses
   bringup: true
   properties:
     - analyze: license

#
# Tags example:
# This test will be categorized as host only framework test.
# Postsubmit runs will be passed "upload_metrics: true".
 - name: Linux analyze
   properties:
     tags: >-
       ["framework", "hostonly"]
   postsubmit_properties:
     - upload_metrics: "true"

#
# Devicelab example:
# For tests that are located https://github.com/flutter/flutter/tree/master/dev/devicelab/bin/tasks:
# 1) target name follows format of `<platform> <taskname>`
# 2) properties
#    2.1) update `tags` based on hosts, devices, and tests type. These tags will be used for statistic analysis.
#    2.2) a `taskname` property is required, which should match the task name
#
# Here is the target config for a task named: `analyzer_benchmark.dart`.
 - name: Linux_android analyzer_benchmark
   recipe: devicelab/devicelab_drone
   presubmit: false
   properties:
     tags: >
       ["devicelab", "android", "linux"]
     task_name: analyzer_benchmark

Adding new targets

All new targets should be added as bringup: true to ensure they do not block the tree.

Targets first need to be mirrored to flutter/infra before they will be run. This propagation takes about 30 minutes, and will only run as non-blocking in postsubmit.

The target will show runs in https://ci.chromium.org/p/flutter (under the repo). See https://github.com/flutter/flutter/wiki/Adding-a-new-Test-Shard for up to date information on the steps to promote your target to blocking.

For flutter/flutter, there's a GitHub bot that will promote a test that has been passing for the past 50 runs.

Test Ownership

This only applies to flutter/flutter*

To prevent tests from rotting, all targets are required to have a clear owner. Add an owner in TESTOWNERS

Properties

Targets support specifying properties that can be passed throughout infrastructure. The following are a list of keys that are reserved for special use.

Properties is a Map<String, String> and any special values must be JSON encoded (i.e. no trailing commas). Additionally, these strings must be compatible with YAML multiline strings

add_recipes_cq: String boolean whether to add this target to flutter/recipes CQ. This ensures changes to flutter/recipes pass on this target before landing.

dependencies: JSON list of objects with “dependency” and optionally “version”. The list of supported deps is in flutter_deps recipe_module. Dependencies generate a corresponding swarming cache that can be used in the recipe code. The path of the cache will be the name of the dependency.

Versions can be located in CIPD

Example

dependencies: >-
  [
    {"dependency": "android_sdk"},
    {"dependency": "chrome_and_driver", "version": "latest"},
    {"dependency": "clang"},
    {"dependency": "goldctl"}
  ]

tags: JSON list of strings. These are currently only used in flutter/flutter to help with TESTOWNERSHIP and test flakiness.

Example

tags: >
  ["devicelab","hostonly"]

Upgrading dependencies

  1. Find the cipd ref to upgrade to
  2. In ci.yaml, find a target that would be impacted by this change
    • Override the version specified in dependencies
      - name: Linux Host Engine
        recipe: engine
        properties:
          build_host: "true"
          dependencies: >-
          [
              {"dependency": "open_jdk", "version": "11"}
          ]
        timeout: 60
      
    - Send PR, wait for the checks to go green (the change takes effect on presubmit)
    
  3. If the check is red, add patches to get it green
  4. Once the PR has landed, infrastructure may take 1 or 2 commits to apply the latest properties

External Tests

Cocoon supports tests that are not owned by Flutter infrastructure. By default, these should not block the tree but act as FYI to the gardeners.

  1. Contact flutter-infra@ with your request (go/flutter-infra-office-hours)
  2. Add your system to SchedulerSystem (https://github.com/flutter/cocoon/blob/master/app_dart/lib/src/model/proto/internal/scheduler.proto)
  3. Add your service account to https://github.com/flutter/cocoon/blob/master/app_dart/lib/src/request_handling/swarming_authentication.dart
  4. Add a custom frontend icon - https://github.com/flutter/cocoon/blob/master/dashboard/lib/widgets/task_icon.dart
  5. Add a custom log link - https://github.com/flutter/cocoon/blob/master/dashboard/lib/logic/qualified_task.dart
  6. Wait for the next prod roll (every weekday)
  7. Add a target to .ci.yaml
    # .ci.yaml
    # Name is an arbitrary string that will show on the build dashboard
    - name: my_external_test_a
      # External tests should not block the tree
      bringup: true
      presubmit: false
      # Scheduler must match what was added to scheduler.proto (any unique name works)
      scheduler: my_external_location
    
  8. Send updates to https://flutter-dashboard.appspot.com/api/update-task-status - https://github.com/flutter/cocoon/blob/master/app_dart/lib/src/request_handlers/update_task_status.dart

Scheduling Targets

For targets using the Cocoon scheduler, they can run on:

By default, all targets should use the Cocoon scheduler.

Presubmit Features

  1. GitHub checks enable targets to run immediately, and are available on the pull request page.
  2. Changes to the ci.yaml will be applied during those presubmit runs.
  3. New targets are required to be brought up with bringup: true

Postsubmit Features

  1. Targets are immediately triggered on GitHub webhooks for merged pull requests
  2. Updates are made immediate via LUCI PubSub notifications
  3. Prioritizes recently failed targets (to unblock the tree quicker)
  4. Backfills targets at a low swarming priority when nothing is actively running
  5. Batches targets that have a high queue time, and backfills in off peak hours
  6. Flakiness monitoring