airflow/dev/breeze/doc/ci/04_selective_checks.md

Table of Contents generated with DocToc

• Selective CI Checks
  • Why selective checks exist (the optimisation goal)
  • Mental model: events, modes, file groups, outputs
  • The decision pipeline
  • Individually simple rules
  • Worked examples: pull requests, pushes, and canary runs
  • Troubleshooting: my CI run does too much
  • Groups of files that selective check make decisions on
  • Selective check decision rules
  • Skipping prek hooks (Static checks)
  • Suspended providers
  • Selective check outputs
  • Self-hosted (RunsOn) vs. public runners
  • Changing behaviours of the CI runs by setting labels

Selective CI Checks

In order to optimise our CI jobs, we've implemented optimisations to only run selected checks for some kind of changes. The logic implemented reflects the internal architecture of Airflow 2.0 packages, and it helps to keep down both the usage of jobs in GitHub Actions and CI feedback time to contributors in case of simpler changes.

Note

All the logic described here lives in a single file: dev/breeze/src/airflow_breeze/utils/selective_checks.py. When you change that file, update this document in the same PR so the behaviour and the documentation stay in sync.

breeze verify maps the run_* flags to local commands in dev/breeze/src/airflow_breeze/utils/verification_plan.py. Adding a run_* flag to SelectiveChecks also means classifying it there; dev/breeze/tests/test_verification_plan.py fails until you do.

Why selective checks exist (the optimisation goal)

Airflow's full CI matrix is large: multiple Python versions, multiple databases (Postgres, MySQL, SQLite), Kubernetes versions, Helm tests, 100+ provider packages, UI tests, doc builds, image builds for two architectures, and dozens of static checks. Running everything for every change would burn an enormous amount of GitHub Actions minutes and make contributors wait an hour or more for feedback on a one-line doc fix.

Selective checks exist to answer one question, cheaply and conservatively:

Given exactly which files changed (plus the event, the branch and the labels), what is the smallest set of CI work that still gives us confidence the change is correct?

Three design assumptions drive everything:

1. The changed file list is a reliable proxy for risk. A change to airflow-core/.../scheduler can break anything, so it must run broadly. A change confined to providers/amazon can only plausibly break Amazon (and its direct dependents), so only those need to run. A change to a .md file can't break runtime behaviour at all.
2. Be conservative when in doubt — never trade correctness for speed. Every rule is biased toward running more. If a change touches something foundational (the environment, the CI scripts, the dependency set, the API contract, core test utilities), selective checks fall back to the full matrix. Skipping is only ever done when a rule can prove the skipped work is irrelevant.
3. Trust must be earned on main. Pull-request runs are optimised aggressively for fast feedback, but the merge to main (and the nightly canary run) always runs the full matrix on all versions. So even if a PR-time optimisation was too aggressive, the canary catches it before a release.

Mental model: events, modes, file groups, outputs

Selective checks take a handful of inputs and produce a set of GitHub Actions outputs that the workflows read to decide what to run. Everything in between is pure, deterministic computation — same inputs always give the same outputs, and you can reproduce it locally (see Troubleshooting).

flowchart LR
    subgraph IN[Inputs]
        F[changed files]
        EV[github event<br/>PR / PUSH / SCHEDULE / DISPATCH]
        BR[target branch<br/>main / v3-x-test]
        LB[PR labels]
        RP[repository<br/>apache/airflow or fork]
    end
    IN --> SC[SelectiveChecks]
    SC --> MODE[Run MODE<br/>full_tests_needed?<br/>all_versions?<br/>is_canary_run?]
    SC --> MATCH[FILE-GROUP matches<br/>regex per area]
    MODE --> OUT[GitHub Actions outputs]
    MATCH --> OUT
    OUT --> JOBS[which jobs run<br/>which versions<br/>which test types<br/>which prek hooks]

The two halves that matter are:

• The run MODE — a few booleans (full_tests_needed, all_versions, is_canary_run) computed mostly from the event, the branch and a handful of high-impact files/labels. The mode decides whether we take the "run everything" shortcut or the "be selective" path.
• The FILE-GROUP matches — every area of the codebase (UI, Helm, docs, Kubernetes, each provider, each mypy target, …) has a list of regexps in CI_FILE_GROUP_MATCHES. A group is "active" when at least one changed file matches it. In selective mode, active groups are what turn individual jobs on.

The key helper that ties the two together is _should_be_run(group):

_should_be_run(group):
    if full_tests_needed:  return True          # full mode → every area is "on"
    return (any changed file matches group)     # selective mode → only matched areas are "on"

Almost every run_* / *_build output is just _should_be_run(<some group>), so once you understand those two halves you understand the whole system.

The decision pipeline

The computation happens in three conceptual stages. Each individual rule is tiny; the apparent complexity is only the number of rules, not their difficulty.

Stage 1 — pick the run mode

First selective checks decide whether this run takes the full-matrix shortcut. The foundational helper is _should_run_all_tests_and_versions(), which is driven by the event and a few high-blast-radius signals:

flowchart TD
    A([_should_run_all_tests_and_versions]) --> T{fork push with<br/>truncated change set?}
    T -->|yes| YES0[True<br/>run everything]
    T -->|no| FP{fork push?}
    FP -->|yes - treated as a pull request| E
    FP -->|no| B{event is<br/>PUSH / SCHEDULE / DISPATCH?}
    B -->|yes| C{PUSH and only<br/>.txt/.md files changed?}
    C -->|yes| NO1[False<br/>skip full tests]
    C -->|no| D{PUSH with<br/>--selective-tests-on-push?}
    D -->|yes - release branch push in ci-amd.yml| NO2[treated like PULL_REQUEST<br/>selective]
    NO2 --> E
    D -->|no| YES1[True<br/>run everything]
    B -->|no - PULL_REQUEST| E{commit_ref missing?}
    E -->|yes| YES2[True]
    E -->|no| Fp{pyproject.toml changed?}
    Fp -->|yes| YES3[True]
    Fp -->|no| G{generated provider<br/>dependencies changed?}
    G -->|yes| YES4[True]
    G -->|no| NO3[False]

full_tests_needed then layers a short-circuit ladder on top: it is True if _should_run_all_tests_and_versions() is true or any one of a small list of high-impact file groups changed or the full tests needed label is set. The first matching rule wins — order does not matter because any single hit forces the full matrix:

flowchart TD
    A([full_tests_needed]) --> M{_should_run_all_tests_and_versions?}
    M -->|yes| T[TRUE - full matrix]
    M -->|no| E1{environment files?<br/>test workflows, dev/breeze/src,<br/>Dockerfile, scripts/ci/docker-compose+kubernetes, ...<br/>excl. prek + non-test workflows}
    E1 -->|yes| T
    E1 -->|no| E2{API contract / codegen?<br/>generated OpenAPI spec or generator}
    E2 -->|yes| T
    E2 -->|no| E3{git or standard<br/>provider files?}
    E3 -->|yes| T
    E3 -->|no| E4{test helper loaded by every test run?<br/>tests_common pytest plugin + its imports}
    E4 -->|yes| T
    E4 -->|no| E5{'full tests needed' label?}
    E5 -->|yes| T
    E5 -->|no| Fa[FALSE - be selective]

Two related booleans share this machinery:

• all_versions — whether to expand the Python/Kubernetes/DB version axis to all supported versions instead of just the defaults. Labels (default versions only, latest versions only, all versions) override; otherwise it follows _should_run_all_tests_and_versions().
• is_canary_run — True for SCHEDULE/PUSH/DISPATCH events on apache/airflow (not forks) when more than just .txt/.md changed, or whenever the canary label is set. A PUSH with --selective-tests-on-push (env SELECTIVE_TESTS_ON_PUSH) is not a canary: ci-amd.yml passes that flag, so a push to a release branch runs the selective tests of the changes since the last successful run of the branch (--push-base-commit), exactly as a pull request would. Workflows that run on push for other reasons (constraints, image cache) do not pass it, and keep getting the full set of Python versions they need. Canary runs always run unit tests even when full_tests_needed is False.

Stage 2 — decide which jobs run

In full mode every area is "on". In selective mode, each job is gated on its file group via _should_be_run. The image builds are derived (they turn on only because something that needs them turned on):

flowchart TD
    CB{target branch chart/... or airflow-ctl/...?} -->|yes| COMP[only that component's tests ON,<br/>whatever changed; see below]
    CB -->|no| FTN
    FTN{full_tests_needed?} -->|yes| ALL[every run_* area = ON<br/>all versions, all test types]
    FTN -->|no| SEL[selective: gate each area on its file group]
    SEL --> RUT{run_unit_tests?}
    RUT -->|canary OR any source file matched| UNIT[unit tests ON<br/>test types chosen by matched files]
    RUT -->|only new-UI files, or no source files| NOUNIT[unit tests OFF]
    SEL --> AREAS[run_ui_tests, run_kubernetes_tests,<br/>run_helm_tests*, docs_build, run_api_tests,<br/>run_python_scans, run_task_sdk_tests, ...]
    UNIT --> IMG
    AREAS --> IMG
    IMG[ci_image_build = unit OR docs OR k8s OR ui<br/>OR pyproject/provider-yaml changed OR prod_image_build]
    AREAS --> PROD[prod_image_build = k8s OR helm OR<br/>integration/e2e tests]
    PROD --> IMG

* Helm tests only run on the main branch and on chart/ branches.

The release branches of a single component (chart/vX-Yx-test|stable, airflow-ctl/vX-Y-test|stable) only ever get backports of that component, so their runs are scoped to it. The target branch comes from the GitHub context (base_ref for a pull request, ref_name for a push or a dispatched run) and must match one of those names: a feature branch named chart/... is not a component branch, and neither is any push in a fork, which is selected against main. On such a branch every output decorated with _scoped_to_component_branch is off, except those COMPONENT_BRANCH_OUTPUTS lists for the component: on chart/, Helm and Kubernetes tests always run (kustomize overlays as usual) with the CI and PROD images they need; on airflow-ctl/, only the airflow-ctl tests run, in a local virtualenv, so no image is built there: static checks run on the host (basic-checks-only), and the airflow-ctl integration tests, the docs and the dependency upgrade check, which need an image, do not run. The PROD image is built for the chart's tests only, so run-generic-prod-image-tests (Docker Compose, image-building examples, basic and canary end-to-end tests) is off, and so is run-generic-basic-tests (the standalone Airflow and shared distributions tests). Provider tests are skipped, and on chart/ docs, when they changed, build only the chart's package. Static checks, scripts and Breeze integration tests, and the CodeQL scans run as usual. A test in test_selective_checks.py fails when a new run_* output is neither scoped nor listed as a general check, so a new job does not start running on these branches by accident.

Stage 3 — choose which test types

When unit tests run, selective checks narrow which test types execute, separately for core and providers:

• Core test types (_get_core_test_types_to_run): Always is always included. Each specific type (API, CLI, Serialization, …) is added if its files matched. Then the escape hatch: if any changed source file is left over after removing provider files, test files, UI files, etc. — i.e. something "core/other" changed — all core test types are added, because a core change can affect anything. In full mode, all core types run unconditionally.
• Provider test types (_get_providers_test_types_to_run): empty on non-main branches. In full mode (or when dependencies were upgraded) → Providers (all). Otherwise selective checks compute the affected providers from the changed files and add their direct upstream and downstream dependents (not the whole transitive closure). Changes to common provider code under devel-common/ escalate to all providers, except a changed tests_common helper, which instead selects the test files that import it (directly or through other helpers) as if those files had changed. Suspended providers are excluded (and a PR that touches one fails unless it carries the allow suspended provider changes label). Providers whose DB tests leave process-global state behind (PROVIDERS_WITH_PROCESS_GLOBAL_TEST_SIDE_EFFECTS, currently cncf.kubernetes) are added to the Providers[...] test type when a provider whose own files changed sorts after them: on canary the DB tests of Providers[-amazon,celery,google,standard] run in one pytest process with provider test folders in sorted order, so the changed provider's tests run after the leaked state is in place. Running them together in the PR surfaces a test that depends on clean state in the PR instead of after merge; a new leak added on the cncf.kubernetes side is not caught this way. Dependents pulled in only for coverage do not trigger this. The individually-listed test types run each provider on its own and are not widened, and the job description names the selected providers.

The same matched-file approach drives the prek hook skip list (skip_prek_hooks): each mypy / compile / lint hook is skipped when nothing in its area changed. See Skipping prek hooks.

The run's platform (--platform)

Some integrations and providers only work on one CPU architecture, so the selection above is filtered by the platform the run's tests will execute on:

• Integrations in DISABLE_TESTABLE_INTEGRATIONS_FROM_ARM are dropped on ARM.
• Providers that declare excluded-platforms in their provider.yaml (e.g. ibm.mq and ibm.db2 exclude linux/arm64) are removed from the providers test-type matrix on that platform. Only linux/* values can match a run; a provider may also list a non-CI platform such as darwin/arm64, which never affects the matrix and exists solely to add an install-time platform_machine marker.

Individually simple rules

The list of rules is long, but each rule is a one-liner you can reason about in isolation. A few representative examples (file → effect):

A change to …	… turns on	because
README.md only	(almost) nothing	.md/.txt are text-non-doc; on a push this even skips full tests
airflow-core/docs/...rst	docs_build (+ CI image)	matches DOC_FILES; docs need the CI image to render
providers/amazon/.../s3.py	unit tests with Providers[amazon, …dependents] + Always	a provider file → only that provider and its direct dependents
airflow-core/src/airflow/jobs/scheduler_job_runner.py	unit tests with all core test types	a core/other file → escape hatch runs all core types
pyproject.toml	full matrix, all versions	dependency surface changed → _should_run_all_tests_and_versions
.github/workflows/ci-amd.yml or dev/breeze/src/...	full matrix	environment files → can change the whole CI environment
.github/workflows/codeql-analysis.yml (non-test workflow)	basic checks only	non-test workflow → cannot affect tests (env-files carve-out)
scripts/ci/prek/check_*.py (static-check hook)	CI image + static checks, no full matrix	prek hooks are static checks → Prek files carve-out
the generated OpenAPI spec	full matrix	the API contract ripples to UI codegen + every client
core_api/openapi/_private_ui.yaml (UI-only spec)	UI compile/lint prek hooks, no full matrix	matches UI OpenAPI files; the UI codegen input must be type-checked, but the public contract is unchanged
chart/templates/...yaml (on main)	run_helm_tests (+ PROD image)	matches HELM_FILES; Helm tests only on main
task-sdk/.../task_runner.py or airflow-core/tests/integration/otel/...	the otel core integration	matches OTEL_FILES; the otel integration tests assert the span hierarchy task_runner emits
airflow-core/src/airflow/ui/...tsx only	run_ui_tests, no unit tests	"only new-UI files" short-circuit skips Python unit tests

The "complexity" you feel reading the code is just many such rules stacked up — each one on its own is simple, conservative, and independently testable (see dev/breeze/tests/test_selective_checks.py).

Worked examples: pull requests, pushes, and canary runs

The same engine produces very different results depending on the event, the branch and the files. These four scenarios cover the common cases:

flowchart TD
    subgraph PR[Pull request - small provider change]
        PR1[event=PULL_REQUEST<br/>files=providers/amazon/.../s3.py] --> PR2[full_tests_needed=False<br/>all_versions=False<br/>is_canary_run=False]
        PR2 --> PR3[unit tests: Providers amazon + Always<br/>default versions only<br/>amazon prek/mypy only]
    end
    subgraph RELFULL[Release branch full run, dispatched]
        RF1[event=WORKFLOW_DISPATCH<br/>branch=v3-x-test] --> RF2[full_tests_needed=True<br/>all_versions=True<br/>is_canary_run=True]
        RF2 --> RF3[full matrix, all versions,<br/>all canary-only jobs]
    end
    subgraph RELPUSH[Push to release branch v3-x-test]
        RP1[event=PUSH<br/>--selective-tests-on-push] --> RP2[full_tests_needed=False<br/>is_canary_run=False]
        RP2 --> RP3[same selection as a PR<br/>with the files changed since<br/>the last successful run<br/>default versions]
    end
    subgraph CANARY[Scheduled nightly canary]
        CN1[event=SCHEDULE on apache/airflow] --> CN2[full_tests_needed=True<br/>all_versions=True<br/>is_canary_run=True]
        CN2 --> CN3[everything, all versions]
    end

1. Pull request, small provider change (providers/amazon/.../s3.py). Not an environment/contract file, commit_ref present, no special label → full_tests_needed=False, all_versions=False. Result: unit tests run only Providers[amazon] (plus amazon's direct dependents) and Always, on default versions only; only the amazon-related prek/mypy hooks run. Fast PR feedback.
2. Pull request, core change (scheduler_job_runner.py). Still full_tests_needed=False, but the core/other escape hatch adds all core test types. Providers are not pulled in (no provider files changed). Default versions.
3. Pull request that changes pyproject.toml (or a test workflow like .github/workflows/ci-amd.yml, or the OpenAPI spec). full_tests_needed=True (and for pyproject.toml also all_versions=True). The PR runs the full matrix — same as a canary — because the change can affect everything. (A non-test workflow such as codeql-analysis.yml, or a scripts/ci/prek/... static-check hook, does not force the full matrix — see the Environment files carve-outs above.)
4. Full run of a release branch (WORKFLOW_DISPATCH on v3-1-test, started by release-branch-ci.yml a few times a day or by a release manager). full_tests_needed=True, all_versions=True, is_canary_run=True. The full matrix plus all canary-only jobs run, and the branch's constraints and image cache are refreshed. main gets the same from the scheduled canary; neither builds the full matrix on push.
5. Push to a release branch (v3-1-test). ci-amd.yml passes --selective-tests-on-push and --push-base-commit (the commit of the last successful run of the workflow on the branch), so the push is treated like a pull request whose files are those changed since that commit: full_tests_needed and is_canary_run follow the PR rules, and the run selects only the tests those changes need, on default versions. If the diff cannot be computed, all tests run.
6. Scheduled nightly canary (SCHEDULE on apache/airflow). Full matrix on all versions, every canary-only job. The most thorough run; it is what gives us confidence that the PR-time shortcuts never let a regression through.
7. Push in a contributor fork (AIP-120, see Fork CI for external contributors). The changed files come from the cross-repository compare API against apache/airflow's branch line of the pushed code (breeze's default branch at that commit: main, or the release branch it was cut from), not from git diff-tree, with a renamed file counted under both paths, and the run is selected exactly like a pull request with those files. is_committer_build is always False because forks have no self-hosted runners. When the file list reaches the compare API's 300-file cap, or the compare fails, everything runs, including the UI E2E tests, which otherwise only follow matching files or the full tests needed label. Checks that compare a file with its previous version (pyproject.toml and the provider dependency bump check) read the previous version at the compare's merge base in apache/airflow, not at the pushed commit's parent, which after a merge of upstream main is the contributor's own previous tip. A file missing at the merge base counts as new; any other failure to read it fails the selective checks instead of skipping the check. The labels come from the open apache/airflow pull request of the pushed branch, so full tests needed, allow provider dependency bump and the other labels work as on a pull request; a label added later takes effect on the next fork run.

Troubleshooting: my CI run does too much

If a CI run is slower or broader than you expect (full matrix on a small change, all providers running, all versions), the cause is almost always a single rule that fired. To find it:

1. Read the selective-checks job output. Selective checks print a [warning] … line for every decision, e.g. "Running full set of tests because env files changed" or "Running everything with all versions: changed pyproject.toml". That line names the exact rule (and usually the file group) that escalated the run. This is the fastest way to diagnose.

2. Reproduce locally with Breeze, pointing at the squashed commit of your change:

   breeze ci selective-check --commit-ref <commit_sha>
   

   It prints the same outputs and the same [warning] reasons CI uses, so you can iterate without pushing.

3. Check the usual escalation triggers (any one of these forces the full matrix):

   • an environment file changed — .github/workflows/*, dev/breeze/src/*, Dockerfile*, scripts/ci/*, scripts/docker/*, (often this is the surprise: editing CI/breeze itself runs everything);
   • pyproject.toml or generated provider dependencies changed (also forces all_versions);
   • the generated OpenAPI spec or the client generator changed (the API contract);
   • a tests_common helper loaded by every test run (the pytest plugin, anything it imports, conftest or package __init__ modules), a helper whose importers cannot be narrowed (it was deleted or renamed, git grep failed, or an importer lies outside the known test trees), or git/standard provider files changed;
   • the full tests needed or all versions label is set on the PR.

4. All providers running? That means selective checks decided all providers are affected — usually because common provider code under devel-common/ changed, or because dependencies were upgraded, or full_tests_needed is on. The reason is printed in the provider-selection [warning] lines.

5. Want to confirm an optimisation is safe? Remember the canary on main always runs everything — the worst case of an over-aggressive skip is caught at merge time, not in production.

The authoritative, exhaustive rule list (kept in sync with the code) is in Selective check decision rules below; the sections above are the "why" and the shape, this is the "what" in full detail.

Groups of files that selective check make decisions on

We have the following Groups of files for CI that determine which tests are run:

• Environment files - if any of those changes, that forces 'full tests needed' mode, because changes there might simply change the whole environment of what is going on in CI (Container image, dependencies). Two deliberate carve-outs do not force full tests: static-check hooks (scripts/ci/prek/...) drive prek static checks rather than the test matrix, so they only trigger the CI image build (via the Prek files group) — mypy-scripts and the image-based static checks still run, but the full test matrix does not; and non-test workflows (.github/workflows/ for security scans, doc publishing, notifications, backporting, calendar/stale bots, …) cannot change test outcomes. Workflows that run or configure the test suite (ci-amd, ci-arm, run-unit-tests, k8s-tests, helm-tests, image builds, …) still force the full matrix.
• Prek files - scripts/ci/prek/... static-check hooks. They do not force full tests, but they trigger the CI image build so mypy-scripts and the image-based static checks run for a prek-only change.
• Python production files and Javascript production files - this area is useful in CodeQL Security scanning
  • if any of the python or javascript files for airflow "production" changed, this means that the security scans should run
• Always test files - Files that belong to "Always" run tests.
• API tests files and Codegen test files - those are OpenAPI definition files that impact Open API specification and determine that we should run dedicated API tests.
• UI OpenAPI files - the OpenAPI spec yamls under core_api/openapi/ and under the simple auth manager's openapi/ that are the inputs of the UI client codegen. Membership in this group does not force full tests (a generated spec still does, via Codegen test files); it only keeps the UI compile/lint prek hooks from being skipped (see Skipping prek hooks).
• Helm files - change in those files impacts helm "rendering" tests - chart folder (which contains the chart sources and tests under chart/tests/).
• Build files - change in the files indicates that we should run upgrade to newer dependencies - build dependencies in pyproject.toml and generated dependencies files in generated folder. The dependencies are automatically generated from the provider.yaml files in provider by the hatch_build.py build hook. The provider.yaml is a single source of truth for each provider and hatch_build.py for all regular dependencies.
• DOC files - change in those files indicate that we should run documentation builds (both airflow sources and airflow documentation)
• UI files - those are files for the new full React UI (useful to determine if UI tests should run)
• System test files - those are the files that are part of system tests (system tests are not automatically run in our CI, but Airflow stakeholders are running the tests and expose dashboards for them at System Test Dashboards
• Kubernetes files - determine if any of Kubernetes related tests should be run
• All Python files - if none of the Python file changed, that indicates that we should not run unit tests
• All source files - if none of the sources change, that indicates that we should probably not build an image and run any image-based static checks
• All Airflow Python files - files that are checked by mypy-airflow-core static checks
• All Providers Python files - files that are checked by mypy-providers static checks
• All Dev Python files - files that are checked by mypy-dev static checks
• All Scripts Python files - files that are checked by mypy-scripts static checks
• Task SDK files - files that are checked by mypy-task-sdk static checks
• All Airflow CTL Python files - files that are checked by mypy-airflow-ctl static checks
• All Devel Common Python files - files that are checked by mypy-devel-common static checks
• All Helm Tests Python files / All Docker Tests Python files / All Kubernetes Tests Python files / All Airflow E2E Tests Python files / Airflow CTL Integration Test files / Task SDK Integration Test files - files that are checked by the respective mypy-helm-tests / mypy-docker-tests / mypy-kubernetes-tests / mypy-airflow-e2e-tests / mypy-airflow-ctl-tests / mypy-task-sdk-integration-tests hooks
• shared/<dist>/**/*.py - each shared/<dist> workspace member has its own mypy-shared-<dist> hook (selective-checks enumerates them at runtime)
• All Provider Yaml files - all provider yaml files

We have a number of TEST_TYPES that can be selectively disabled/enabled based on the content of the incoming PR. Usually they are limited to a sub-folder of the "tests" folder but there are some exceptions. You can read more about those in testing.rst <contributing-docs/09_testing.rst>. Those types are determined by selective checks and are used to run DB and Non-DB tests.

The DB tests inside each TEST_TYPE are run sequentially (because they use DB as state) while TEST_TYPES are run in parallel - each within separate docker-compose project. The Non-DB tests are all executed together using pytest-xdist (pytest-xdist distributes the tests among parallel workers).

Selective check decision rules

• Full tests case is enabled when the event is PUSH, SCHEDULE or WORKFLOW_DISPATCH, or we miss commit info, or any of the important environment files (pyproject.toml, Dockerfile, scripts, etc.) changed, or the API contract changed (the generated OpenAPI spec or the client generator — plain API source/test edits that leave the committed spec untouched do not force full tests), or a tests_common helper loaded by every test run or whose importers cannot be narrowed (helper deleted or renamed, git grep failed, or an importer outside the known test trees), or git / standard provider files changed, or when the full tests needed label is set. Any other changed tests_common helper only selects the tests that import it. That enables all matrix combinations of variables (representative) and all possible test type. No further checks are performed. See also [1] note below. Two exceptions narrow this: a PUSH that changed only .txt/.md files skips full tests, and a PUSH with --selective-tests-on-push (which ci-amd.yml passes for pushes to release branches) is treated like a pull request — selective tests on default versions, not a canary run. A PUSH in a contributor fork is not a PUSH for this rule either: it is selected like a pull request against apache/airflow, and runs everything only when its change set reaches the compare API's 300-file cap (a list of 300 may be missing files) or the compare fails. The incremental base of a release-branch push (the last successful run of the branch) is only looked up in apache/airflow. The high-level flow and worked examples for these cases are in The decision pipeline and Worked examples above.
• Python, Kubernetes, Backend, Kind, Helm versions are limited to "defaults" only unless Full tests mode is enabled.
• Python scans, Javascript scans, API tests/codegen, UI, WWW, Kubernetes tests and DOC builds are enabled if any of the relevant files have been changed.
• Helm tests are run only if relevant files have been changed and if current branch is main.
• On a component's release branch (chart/..., airflow-ctl/...) only that component's tests run, always, with the images they need (none for airflow-ctl); other tests do not run, provider tests are skipped, and docs build only the chart's package. Static checks still run. See Stage 2.
• If no Source files are changed - no tests are run and no further rules below are checked.
• CI Image building is enabled if either test are run, docs are build.
• PROD Image building is enabled when kubernetes tests are run.
• In case of Providers test in regular PRs, additional check is done in order to determine which providers are affected and the actual selection is made based on that:
  • if directly provider code is changed (either in the provider, test or system tests) then this provider is selected.
  • if there are any providers that depend on the affected providers, they are also included in the list of affected providers (but not recursively - only direct dependencies are added)
  • if there are any changes to "common" provider code not belonging to any provider (usually system tests or tests), then tests for all Providers are run
  • if a provider whose own files changed sorts after a provider with process-global DB-test side effects (PROVIDERS_WITH_PROCESS_GLOBAL_TEST_SIDE_EFFECTS, currently cncf.kubernetes), the side-effect provider is added to the same Providers[...] test type, because the two share one pytest process on canary and the changed tests must pass after it
• TS SDK unit tests (the vitest suite run by pnpm test in ts-sdk/, exposed as the run-ts-sdk-tests output) run when TS SDK files change: anything under ts-sdk/ except .md files, ts-sdk/api-docs/, and ts-sdk/docs/ (the docs toolchain's own package).
• Lang SDK serialization conformance (the tests-lang-sdk-conformance job, which serializes the Dags of scripts/ci/lang_sdk_serialization/test_dags.yaml with an SDK and with Airflow and compares the two) runs for each SDK listed in the lang-sdk-conformance-sdks output: go when Go SDK conformance files change, java when Java SDK conformance files change and ts when TS SDK conformance files change. Each group is the SDK's own files (as in Go SDK files, Java SDK files and TS SDK files) plus what the check compares the SDK against: Airflow's serializer (serialized_objects.py), airflow-core/src/airflow/serialization/schema.json and the shared harness under scripts/ci/lang_sdk_serialization/. None of those three forces full_tests_needed, so a change to any of them runs the check for every SDK. The list is empty on a component branch.
• Java SDK E2E tests (the java_sdk mode of the deployed-stack tests, exposed as the run-java-sdk-e2e-tests output) run when the Java SDK sources (java-sdk/, excluding .md), the Java test-fixture and native-Dag bundles (airflow-e2e-tests/java-test-bundle/, airflow-e2e-tests/java-native-bundle/), the Java e2e suite or its Docker files (airflow-e2e-tests/tests/airflow_e2e_tests/java_sdk_tests/, airflow-e2e-tests/docker/java.yml, airflow-e2e-tests/docker/Dockerfile.java), the Java coordinator (task-sdk/src/airflow/sdk/coordinators/java/, _subprocess.py), or the native Lang-SDK Dag parsing sources change. Those are the Dag processing files importer_routing.py and lang_sdk_processor.py, models/dagcode.py, serialization/serialized_objects.py, the Task SDK importers (task-sdk/src/airflow/sdk/importers/), the coordinator registry (task-sdk/src/airflow/sdk/execution_time/coordinator.py), and coordinators/_dag_importer.py; they trigger the TypeScript SDK E2E tests too. Shared Dag processing files such as manager.py are left to canary runs. Like the other deployed e2e suites, enabling them forces PROD Image building.
• YAML Dag E2E tests (the yaml_dag mode of the deployed-stack tests, exposed as the run-yaml-dag-e2e-tests output) run when the YAML Dag fixture or its suite change (airflow-e2e-tests/yaml-dags/, airflow-e2e-tests/tests/airflow_e2e_tests/yaml_dag_tests/, airflow-e2e-tests/docker/yaml-dag.yml), when the Go example bundle's handler for that Dag changes (go-sdk/example/bundle/yamldag/, go-sdk/example/bundle/main.go — the mode runs the Dag's one code task on that bundle, so the rest of the Go SDK is left to the Go SDK E2E tests), when the executable coordinator changes (task-sdk/src/airflow/sdk/coordinators/executable/, _subprocess.py), or when the native Lang-SDK Dag parsing sources listed above change — the YAML importer itself lives in task-sdk/src/airflow/sdk/importers/. Like the other deployed e2e suites, enabling them forces PROD Image building.
• OpenLineage E2E tests (the openlineage mode of the deployed-stack tests under airflow-e2e-tests/tests/airflow_e2e_tests/openlineage_tests, exposed as the run-openlineage-e2e-tests output) run when the openlineage or common providers or the openlineage e2e suite change — and always on canary runs (where full tests needed also covers core/task-sdk changes). Like the other deployed e2e suites, enabling them forces PROD Image building.
• OpenLineage E2E compat tests (the same suite rerun against older released Airflow versions with current provider code, exposed as the run-openlineage-e2e-compat-tests output) are costly, so they do NOT run on every OpenLineage PR. They run on canary runs, when the full tests needed label is explicitly set, or when a file that drives the compat setup changes. These files are the compat workflow (.github/workflows/openlineage-e2e-compat-tests.yml), the compat Dockerfile (airflow-e2e-tests/docker/openlineage-compat.Dockerfile), the compat database override (airflow-e2e-tests/docker/openlineage-compat-db.yml) and the shared e2e harness (airflow-e2e-tests/tests/airflow_e2e_tests/conftest.py / constants.py). Like run-ui-e2e-tests, run-openlineage-e2e-compat-tests ignores full_tests_needed. Many changes that do not touch the compat setup also turn on full_tests_needed, for example changes to environment files or the API spec. So a PR that changes only breeze's e2e command does not run the compat tests, and the next canary run tests that change.
• The specific unit test type is enabled only if changed files match the expected patterns for each type (API, CLI, WWW, Providers etc.). The Always test type is added always if any unit tests are run. Providers tests are removed if current branch is different than main
• If there are no files left in sources after matching the test types and Kubernetes files, then apparently some Core/Other files have been changed. This automatically adds all test types to execute. This is done because changes in core might impact all the other test types.
• if CI Image building is disabled, only basic prek hooks are enabled - no 'image-depending` prek hooks are enabled.
• If there are some build dependencies changed (hatch_build.py and updated system dependencies in the pyproject.toml - then upgrade to newer dependencies is enabled.
• If docs are build, the docs-list-as-string will determine which docs packages to build. This is based on several criteria: if any of the airflow core, charts, docker-stack, Apache Airflow Mypy, providers files or docs have changed, then corresponding packages are built (including cross-dependent providers). If any of the core files changed, also providers docs are built because all providers depend on airflow docs. If any of the docs build python files changed or when build is "canary" type in main - all docs packages are built.

Skipping prek hooks (Static checks)

Our CI always run prek checks with --all-files flag. This is in order to avoid cases where different check results are run when only subset of files is used. This has an effect that the prek tests take a long time to run when all of them are run. Selective checks allow to save a lot of time for those tests in regular PRs of contributors by smart detection of which prek hooks should be skipped when some files are not changed. Those are the rules implemented:

• The identity check is always skipped (saves space to display all changed files in CI)
• The provider specific checks are skipped when builds are running in v2_* branches (we do not build providers from those branches. Those are the checks skipped in this case:
  • check-airflow-provider-compatibility
  • check-extra-packages-references
  • check-provider-yaml-valid
  • lint-helm-chart
  • mypy-providers
• If "full tests" mode is detected, no more prek hooks are skipped - we run all of them
• The following checks are skipped if those files are not changed:
  • if no All Providers Python files changed - mypy-providers check is skipped
  • if no All Airflow Python files changed - mypy-airflow-core check is skipped
  • if no All Dev Python files changed - mypy-dev check is skipped
  • if no All Scripts Python files changed - mypy-scripts check is skipped
  • if no Task SDK files changed - mypy-task-sdk check is skipped
  • if no All Airflow CTL Python files changed - mypy-airflow-ctl check is skipped
  • if no All Devel Common Python files changed - mypy-devel-common check is skipped
  • if no files under the matching folder changed, the corresponding per-folder mypy hook is skipped (mypy-airflow-ctl-tests, mypy-helm-tests, mypy-airflow-e2e-tests, mypy-task-sdk-integration-tests, mypy-docker-tests, mypy-kubernetes-tests)
  • for each shared/<dist> workspace member, mypy-shared-<dist> is skipped when no file under shared/<dist>/ changed (enumerated at runtime)
  • if neither UI files nor UI OpenAPI files changed - the ts-compile-lint-ui and ts-compile-lint-simple-auth-manager-ui checks are skipped. The UI OpenAPI files group covers the union of those hooks' own openapi files: triggers, so a spec-only change (e.g. _private_ui.yaml) still runs them - otherwise a stale committed client can mask type errors (see #68919)
  • if no All Python files changed - flynt check is skipped
  • if no Helm files changed - lint-helm-chart check is skipped
  • if no Java SDK files (java-sdk/, excluding .md) changed - ktlint, spotless and regenerate-java-sdk-verification-metadata checks are skipped. All three run the java-sdk Gradle wrapper, and regenerate-java-sdk-verification-metadata also resolves the whole Java SDK dependency graph from Maven Central
  • if no Java SDK conformance files changed - sync-java-sdk-dag-schema is skipped (it runs the :sdk:syncDagSchema Gradle task). The group is wider than Java SDK files: it also covers Airflow's serializer, airflow-core/src/airflow/serialization/schema.json and the shared harness under scripts/ci/lang_sdk_serialization/ (see Lang SDK serialization conformance above). Because airflow-core's schema.json is in the group, sync-java-sdk-dag-schema still runs on a PR that changes only that file. Such a change is what leaves the Java SDK's copy in java-sdk/sdk/schema/dag-schema.json out of date
  • whichever of the four Java SDK hooks above runs first in a job pays for Gradle's cold start, which includes downloading the Gradle distribution. A PR skips the cold start only when it changes none of the Java SDK conformance files
  • neither Java SDK group includes .pre-commit-config.yaml or CI setup files. A PR that changes only the definition of one of the five hooks therefore does not run the hooks in its own CI. The same holds for a PR whose only change is CI setup that does not force full_tests_needed, for example a composite action under .github/actions/. The scheduled canary runs skip none of the five hooks, so they still check such a change
  • if no TS SDK files (ts-sdk/) changed - check-ts-sdk-supervisor-schema check is skipped (it regenerates and diffs the generated ts-sdk file from the vendored schema). A change to the supervisor wire schema on the Python side is not deferred to a follow-up PR: sync-ts-sdk-schemas is triggered by that source, refreshes the vendored copy, and fails on the PR that caused it, so the drift surfaces there rather than shipping unnoticed
  • check-provider-yaml-valid is skipped unless at least one of these changed: All Providers Python files, All Providers Distribution Config files (which includes provider.yaml, pyproject.toml, and providers/.pre-commit-config.yaml), or Prek files (scripts/ci/prek/). The last condition ensures the check runs when the check script itself is modified.

Suspended providers

The selective checks will fail in PR if it contains changes to a suspended provider unless you set the label allow suspended provider changes in the PR. This is to prevent accidental changes to suspended providers.

Selective check outputs

The selective check outputs available are described below. In case of list-as-string values, empty string means everything, where lack of the output means nothing and list elements are separated by spaces. This is to accommodate for the way how outputs of this kind can be easily used by GitHub Actions to pass the list of parameters to a command to execute

Output	Meaning of the output	Example value	List
all-python-versions	List of all python versions there are available in the form of JSON array	['3.11', '3.12']
all-python-versions-list-as-string	List of all python versions there are available in the form of space separated string	3.11 3.12	*
all-versions	If set to true, then all python, k8s, DB versions are used for tests.	false
basic-checks-only	Whether to run all static checks ("false") or only basic set of static checks ("true")	false
ci-image-build	Whether CI image build is needed	true
core-test-types-list-as-strings-in-json	Which test types should be run for unit tests for core (2 slices; 1 on RunsOn)	API Always Providers	*
debug-resources	Whether resources usage should be printed during parallel job execution ("true"/ "false")	false
default-branch	Which branch is default for the build ("main" for main branch, "v2-4-test" for 2.4 line etc.)	main
default-constraints-branch	Which branch is default for the build ("constraints-main" for main branch, "constraints-2-4" etc.)	constraints-main
default-helm-version	Which Helm version to use as default	v3.9.4
default-kind-version	Which Kind version to use as default	v0.16.0
default-kubernetes-version	Which Kubernetes version to use as default	v1.25.2
default-mysql-version	Which MySQL version to use as default	5.7
default-postgres-version	Which Postgres version to use as default	10
default-python-version	Which Python version to use as default	3.11
disable-airflow-repo-cache	Disables cache of the repo main cache in CI - airflow will be installed without main installation cache	true
docker-cache	Which cache should be used for images ("registry", "local" , "disabled")	registry
docs-build	Whether to build documentation ("true"/"false")	true
docs-list-as-string	What filter to apply to docs building - based on which documentation packages should be built	apache-airflow helm-chart google	*
excluded-providers-as-string	List of providers that should be excluded from the build as space-separated string	amazon google	*
force-pip	Whether pip should be forced in the image build instead of uv ("true"/"false")	false
full-tests-needed	Whether this build runs complete set of tests or only subset (for faster PR builds) [1]	false
generated-dependencies-changed	Whether generated dependencies have changed ("true"/"false")	false
has-migrations	Whether the PR has migrations ("true"/"false")	false
hatch-build-changed	When hatch build.py changed in the PR.	false
helm-test-packages	List of helm packages to test as JSON array (["All"] on RunsOn: one job runs them all)	["airflow_aux", "airflow_core"]	*
helm-version	Which Helm version to use for tests	v3.15.3
in-memory-runner-type	Like runner-type, but the RunsOn template puts docker in RAM (extras=tmpfs) unless disabled	["ubuntu-22.04"]
include-success-outputs	Whether to include outputs of successful parallel tests ("true"/"false")	false
individual-providers-test-types-list-as-strings-in-json	Which test types should be run for unit tests for providers (individually listed; 5 slices, 3 on RunsOn)	Providers[\amazon] Providers[google]	*
is-committer-build	Whether the build is triggered by a committer	false
is-legacy-ui-api-labeled	Whether the PR is labeled as legacy UI/API	false
java-sdk-version	JDK version used to build the lang-SDK Java artifacts natively in CI	17
kind-version	Which Kind version to use for tests	v0.24.0
kubernetes-combos-list-as-string	All combinations of Python version and Kubernetes version to use for tests as space-separated string	3.11-v1.25.2 3.12-v1.28.13	*
kubernetes-versions	All Kubernetes versions to use for tests as JSON array	['v1.25.2']
kubernetes-versions-list-as-string	All Kubernetes versions to use for tests as space-separated string	v1.25.2	*
lang-sdk-conformance-sdks	Language SDKs whose serialization conformance check runs, as JSON array	["go", "java", "ts"]
latest-versions-only	If set, the number of Python, Kubernetes, DB versions will be limited to the latest ones.	false
mypy-checks	List of folders to be considered for mypy checks	["airflow_aux", "airflow_core"]
mysql-exclude	Which versions of MySQL to exclude for tests as JSON array	[]
mysql-versions	Which versions of MySQL to use for tests as JSON array	['8.0']
postgres-exclude	Which versions of Postgres to exclude for tests as JSON array	[]
postgres-versions	Which versions of Postgres to use for tests as JSON array	['12']
prod-image-build	Whether PROD image build is needed	true
providers-compatibility-tests-matrix	Matrix of providers compatibility tests: (python_version, airflow_version, removed_providers). Excludes unbuilt python versions and the entry whose airflow_version equals the current branch version (unreleased on a vX-Y-test branch until GA)	[{}]
providers-test-types-list-as-strings-in-json	Which test types should be run for unit tests for providers	Providers Providers[-google]	*
providers-split-test-types-list-as-strings-in-json	Provider test types for the jobs running them in parallel in the CI image (one per provider on RunsOn)	Providers Providers[-google]	*
providers-xdist-test-types-list-as-strings-in-json	Which test types the Non-DB (xdist) provider tests run, with no provider split out	Providers	*
pyproject-toml-changed	When pyproject.toml changed in the PR.	false
python-versions	List of python versions to use for that build	['3.11']
python-versions-list-as-string	Which versions of MySQL to use for tests as space-separated string	3.11	*
run-agent-framework-tests	Whether the common.ai agent framework adapter tests should be run ("true"/"false")	true
run-amazon-tests	Whether Amazon tests should be run ("true"/"false")	true
run-api-codegen	Whether "api-codegen" are needed to run ("true"/"false")	true
run-api-tests	Whether "api-tests" are needed to run ("true"/"false")	true
run-generic-prod-image-tests	Whether the PROD image's own tests run (Docker Compose, examples, basic e2e): off on component branches	true
run-generic-basic-tests	Whether the standalone Airflow and shared distributions tests run: off on component branches	true
run-helm-tests	Whether Helm tests are needed to run ("true"/"false")	true
run-kubernetes-tests	Whether Kubernetes tests should be run ("true"/"false")	true
run-mypy	Whether mypy check is supposed to run in this build	true
run-system-tests	Whether system tests should be run ("true"/"false")	true
run-task-sdk-tests	Whether Task SDK tests should be run ("true"/"false")	true
run-ts-sdk-docs	Whether the TypeScript SDK API reference should be built — on ts-sdk/api-docs/, ts-sdk/docs/, or ts-sdk/src/ changes, including Markdown ("true"/"false")	true
run-ts-sdk-e2e-tests	Whether TypeScript SDK e2e tests should be run — on runtime-affecting ts-sdk/, TS e2e test, Node coordinator, or native Lang-SDK Dag parsing changes ("true"/"false")	true
run-ts-sdk-tests	Whether TypeScript SDK unit tests should be run — on ts-sdk/ changes except Markdown, api-docs/, and docs/ ("true"/"false")	true
run-ui-tests	Whether UI tests should be run ("true"/"false")	true
run-unit-tests	Whether unit tests should be run ("true"/"false")	true
run-www-tests	Whether Legacy WWW tests should be run ("true"/"false")	true
run-yaml-dag-e2e-tests	Whether YAML Dag e2e tests should be run — on YAML Dag fixture/suite, its Go handler, executable coordinator, or native Lang-SDK Dag parsing changes ("true"/"false")	true
amd-runners	List of labels assigned for runners for that build for public AMD runners	["ubuntu-22.04"]
arm-runners	List of labels assigned for runners for that build for public ARM runners	["ubuntu-22.04-arm"]
runner-type	Runner labels for the CI jobs on this platform: public, or a RunsOn template jobs fill with cpu/ram	["ubuntu-22.04"]
selected-providers-list-as-string	List of providers affected when they are selectively affected.	airbyte http	*
skip-prek-hooks	Which prek hooks should be skipped during the static-checks run	flynt,identity
skip-providers-tests	When provider tests should be skipped (on non-main branch or when no provider changes detected)	true
sqlite-exclude	Which versions of Sqlite to exclude for tests as JSON array	[]
testable-core-integrations	List of core integrations that are testable in the build as JSON array	['celery', 'kerberos']
testable-providers-integrations	List of core integrations that are testable in the build as JSON array	['mongo', 'kafka']
upgrade-to-newer-dependencies	Whether the image build should attempt to upgrade all dependencies (true/false or commit hash); false when no CI image is built	false

Self-hosted (RunsOn) vs. public runners

The Tests (AMD) / Tests (ARM) jobs run either on GitHub-hosted ("public") runners or on self-hosted RunsOn EC2 runners in the Airflow AWS account. Selective checks route runs of code from apache/airflow itself there; this is a routing preference, not what keeps fork code off RunsOn (see below). The runner-type output is decided as follows, first match wins:

1. Runs in any repository other than apache/airflow (forks running their own CI) use public runners.
2. Any run uses public runners while the repository variable USE_PUBLIC_RUNNERS is true: the switch for moving main off RunsOn without a PR, e.g. when the stack is unavailable.
3. PRs with the use public runners label use public runners.
4. Any push, schedule or workflow_dispatch in apache/airflow (canary runs and the release branch pushes, all triggered by people with write access) uses RunsOn.
5. Other events than pull_request use public runners.
6. PRs with the use self-hosted runners label use RunsOn, if a committer's (COMMITTERS in global_constants.py) latest review of the PR's head commit is an approval and no committer's request for changes stands. The label outlives the commit it was added for, so on its own it would vouch for every later push: a new push sends the PR back to public runners until the head commit is approved again, and so does a dismissed approval if no other committer's approval of the head commit remains. A committer's request for changes keeps the PR on public runners across pushes, as it blocks a merge on GitHub, until that committer reviews again or the review is dismissed. The label must also still be on the PR: a re-run reuses the labels of its original event, so they are checked against the PR's current ones. If the PR or its reviews cannot be read, the label is ignored as well.
7. PRs with the non committer build label use public runners.
8. PRs opened by a bot (Dependabot, the ci-upgrade-* PRs and the like), or from a dependabot/ branch, use public runners: they carry code nobody has reviewed yet. The exception is a backport PR: one authored by github-actions[bot] (matched on the exact login) from a backport-* branch (the name cherry_picker gives it) of apache/airflow. A committer's merge started it, or a user with write access dispatched it for a commit they chose. It continues with rule 9; the same PR from a fork does not.
9. PRs from a branch of apache/airflow use RunsOn; PRs from forks use public runners.

Labels are read from the event that started the run, and the CI workflows do not run on labeled. After adding or removing a label, push a commit or close and reopen the PR; re-running an earlier run reuses its original labels, and "Re-run failed jobs" also reuses that run's runner decision. A push voids the approval use self-hosted runners needs, so for that label approve the head commit and then close and reopen the PR.

The RunsOn value is a label template such as ["runs-on=<run_id>/family=.../image=ubuntu22-full-x64/volume=150gb/spot=co/cpu={0}/ram={1}"]. Every job picks its size class with format(fromJSON(inputs.runners)[0], <cpu>, <ram>) ('4+8', '16+32', '8+16', '32+64' or '32+64', '128+256'), or, for a single-job reusable workflow, its caller passes format(<runners>, <cpu>, <ram>). Each class is a range up to twice its size, so Spot can come from the larger instance types too; xdist follows the host's CPUs, so a larger runner runs more test workers. Jobs using --run-in-parallel (the DB, lowest-dependencies and special tests) run one serial pytest per test type instead, so they never use more CPUs than they have test types, and jobs running a single pytest (integration and system tests) use one. The size classes follow that: the xdist jobs and every job running provider tests in parallel take '32+64', '128+256', the others '8+16', '32+64' or smaller: a job running one process (an end-to-end test against one Airflow stack, a Gradle test run) takes '8+16', '32+64', and a job with a minute or two of work takes '4+8', '16+32'. On RunsOn the Helm unit tests run all their packages in one job per Kubernetes version, since each package takes about a minute under xdist against two to three minutes of preparing the image. On RunsOn providers-split-test-types-list-as-strings-in-json, which the provider jobs running in the CI image read, lists each provider as its own test type, in one job, instead of the long-running providers plus one group of all the others, so the job is not left waiting on that group. The provider compatibility tests keep the grouped providers-test-types-list-as-strings-in-json, because each of their test types installs an Airflow version first. The Non-DB xdist jobs read providers-xdist-test-types-list-as-strings-in-json instead, which never splits the providers: xdist runs them all in one pytest anyway. The matrix slices follow it too. On public runners the core test types are split into NUMBER_OF_CORE_SLICES (2) jobs, to keep the Non-DB xdist collection inside the runner's memory, and the individually listed providers into NUMBER_OF_LOW_DEP_SLICES (5). On RunsOn the core test types stay in one job (NUMBER_OF_CORE_SLICES_ON_RUNS_ON), because a core DB job lasts as long as its slowest test type whichever slice it is in, and the providers are split into jobs of about 40 test types (LOW_DEP_TEST_TYPES_PER_JOB_ON_RUNS_ON), which the heaviest-first ordering keeps the 32 CPUs busy with: 3 jobs for all providers, one for a change touching a few. On RunsOn the longest-running providers (PROVIDER_TEST_SPLITS: amazon, google and standard) are also split into several test types such as Providers[amazon:hooks], with Providers[amazon:other] running the rest of the provider's tests, so none of them leaves the job waiting on one long serial pytest. spot=co takes Spot from the pools with the most capacity regardless of price, which is cheaper than the on-demand launch and whole-job rerun a capacity miss or interruption costs; on a public label the same call returns the label unchanged. Steps that only make sense on RunsOn are gated on runner.environment == 'self-hosted':

• The EC2 cost summary of each job, with charts of the job's CPU and memory use (metrics: cpu,memory of runs-on/action), the measurements to size a job's runner class by.
• The runner settings for parallel tests (scripts/ci/configure_runner_for_parallel_tests.sh). Every test type of a parallel run is a docker compose project with a network and a database of its own. Docker's default address pools fit about 30 networks, fewer than a RunsOn runner runs test types at once, so the script splits the same range into /24 subnets; and it raises fs.aio-max-nr, which a few dozen MySQL servers exhaust.
• The image stashes (.github/actions/stash-restore / stash-save). On RunsOn the image tarballs go to the stack's S3 cache bucket, named by their sha256, and only the digest is stashed as a GitHub artifact. Any RunsOn job can write to the bucket, so a tarball is used only if it matches the digest stashed by the same run, branch and repository (or, for a PR, its base branch). The smaller caches (mount cache, prek, docs inventory) use the ASF stash action on both kinds of runners, so PRs on public runners keep the ones canary runs save; the price is that every RunsOn canary uploads the mount cache from AWS to GitHub.

The in-memory-runner-type output is the same decision with a second RunsOn template, used by the jobs whose databases wait on disk syncs: the Postgres, MySQL and Sqlite test jobs, the lowest-dependencies core tests and the SQLAlchemy, Boto and Pendulum special tests. Its label adds extras=tmpfs, which puts /var/lib/docker, /tmp and /home/runner on a RAM disk, and also accepts the memory-optimized r-family. The jobs ask for 8 to 16 CPUs and 32 to 128 GB (cpu=8+16/ram=32+128), as the RAM disk is capped at the runner's memory before RunsOn v3.5.0. On public runners the output equals runner-type, and so it does on RunsOn while the repository variable USE_IN_MEMORY_DOCKER is false: the switch for moving docker back to disk without a PR.

The decision above is not a security boundary. Under the pull_request event the code that evaluates it is the PR's own: a fork PR runs its own workflow files and breeze sources, and can name RunsOn labels in runs-on: directly. RunsOn launches a runner for any job of a repository its GitHub App is installed on, so nothing in this repository can prevent that. Keeping untrusted code away from anything that matters has to be enforced outside the repository, by:

• The repository's Actions setting "Approval for running fork pull request workflows from contributors", which must be "Require approval for all external contributors". That still exempts members of the apache GitHub organization (ASF committers of any project who linked their GitHub account) and repository collaborators, but every exempt person is identifiable.
• The RunsOn stack, which must be dedicated to this repository and give the runners an instance role that reaches nothing beyond the stack's own cache bucket (whose contents are never trusted, see the stash actions), no EBS permissions, no SSH, network isolation between runners and a maximum run time. RunsOn has no cap on instance sizes or counts and its budget only alerts, so the AWS account's EC2 vCPU quotas are the cost ceiling.

Pushes in a contributor fork are never committer builds, whoever pushes, because forks have no self-hosted runners.

Changing behaviours of the CI runs by setting labels

PRs from apache/airflow branches run on self-hosted RunsOn runners and PRs from forks run on public runners (see above). Set use public runners to move a PR to public runners; a committer can set use self-hosted runners and approve the PR's head commit to move a fork PR to RunsOn, provided no committer's request for changes stands, then close and reopen the PR (a push would need a fresh approval; a committer should do the reopen, since an external author's reopen still needs a maintainer to approve the workflow run).

If you are testing CI workflow changes and want to test it for more complete matrix combinations generated by the jobs - you can set all versions label in the PR. This will run the PRs with the same combinations of versions as the canary main build. Using all versions is automatically set when build dependencies change in pyproject.toml.

If you are running an apache PR, you can also set canary label for such PR and in this case, all the canary properties of build will be used: full tests needed mode, all versions as well as all canary-specific jobs will run there. You can modify this behaviour of the canary run by applying use public runners, and default versions only labels to the PR as well which will still run a canary equivalent build but with public runners an default Python/K8S versions only - respectively.

If you are testing CI workflow changes and change pyproject.toml and you want to limit the number of matrix combinations generated by the jobs - you can set default versions only label in the PR. This will limit the number of versions used in the matrix to the default ones (default Python version and default Kubernetes version).

If you are testing CI workflow changes and want to limit the number of matrix combinations generated by the jobs - you can also set latest versions only label in the PR. This will limit the number of versions used in the matrix to the latest ones (latest Python version and latest Kubernetes version).

You can also disable cache if you want to make sure your tests will run with image that does not have left-over package installed from the past cached image - by setting disable image cache label in the PR.

By default, all outputs of successful parallel tests are not shown. You can enable them by setting include success outputs label in the PR. This makes the logs of mostly successful tests a lot longer and more difficult to sift through, but it might be useful in case you want to compare successful and unsuccessful runs of the tests.

This table summarizes the labels you can use on PRs to control the selective checks and the CI runs:

Label	Affected outputs	Meaning
all versions	all-versions, -versions-	Run tests for all python and k8s versions.
allow suspended provider changes	allow-suspended-provider-changes	Allow changes to suspended providers.
area:e2e-tests	prod-image-build	If set, the Airflow E2E tests are run regardless of changed files (does not force the full test matrix).
area:kubernetes-tests	run-kubernetes-tests	If set, the Kubernetes tests job is run regardless of changed files (does not force the full test matrix).
canary	is-canary-run	If set, the PR run from apache/airflow repo behaves as canary run.
debug ci resources	debug-ci-resources	If set, then debugging resources is enabled during parallel tests and you can see them.
default versions only	all-versions, -versions-	If set, the number of Python and Kubernetes, DB versions are limited to the default ones.
disable image cache	docker-cache	If set, the image cache is disables when building the image.
force pip	force-pip	If set, the image build uses pip instead of uv.
full tests needed	full-tests-needed	If set, complete set of tests are run
include success outputs	include-success-outputs	If set, outputs of successful parallel tests are shown not only failed outputs.
latest versions only	-versions-, -versions-	If set, the number of Python, Kubernetes, DB versions will be limited to the latest ones.
non committer build	is-committer-build	If set, the scripts used for images are used from target branch for committers.
upgrade to newer dependencies	upgrade-to-newer-dependencies	If set to true (default false) then dependencies in the CI image build are upgraded.

---

Read next about Workflows