ref(typing)!: enable strict mypy repo-wide by joshuarli · Pull Request #129444 · getsentry/sentry

joshuarli · 2026-11-01T07:12:38Z

Draft until #129431 lands. This branch is stacked on it, so the diff below includes those commits until it merges and this branch is rebased.

Every first-party Python module now passes mypy --strict, so the project config enables strict = true directly and the ratchet that tracked the gap is gone.

Result

Measurement	Before	After
`mypy --strict` errors	1,411 in 712 files	0
`no-any-return`	1,065	0
`no-untyped-call`	271	0
`typeddict-item`	29	0
`untyped-decorator`	26	0
`extra_checks` / `misc` (subclassing `Any`)	20	0

[tool.mypy] now sets strict = true and drops the individual flags it implies. check_strict_typing.py, strict_baseline.json, their pre-commit hook, and the gate unit tests are deleted; there is nothing left to baseline.

How the remaining errors were removed

No error was suppressed and no module was opted out. Each fix reads the value where it actually enters first-party code:

HTTP clients: self.client.* responses are declared as Response / HttpResponseBase, and shared integration clients return BaseApiResponse instead of object.
Decoded JSON: json.loads / orjson.loads results are checked for object or array shape before use, or cast to a named TypedDict with a # justification: comment.
Caches: values read from Redis/Django cache are validated or cast to the type the corresponding writer stores.
Framework seams: DRF validators, Django model metadata, and serializer spreads are read as the type their callers receive.
Provider SDKs: added local stubs for sentry_relay, selenium, slack_sdk, google.cloud, google.api_core, emmett55, and django_zero_downtime_migrations, plus typed adapters for datadog, google, and drf-spectacular.
Typed helpers: get_cached returns BaseApiResponse; callers that need structured payloads use get_mapping / get_sequence / get_cached_mapping / get_cached_sequence.

Casts at trust boundaries carry a # justification: comment describing the guarantee. mypy --strict is the only gate: it now fails on any new Any return, untyped call, or untyped decorator.

Verification

SENTRY_LOG_LEVEL=WARNING .venv/bin/python -m mypy (config strict = true) → Success: no issues found
prek run -q → pass on the full change set, including the mypy, check-mypy-bypass, weaklist, and typing-campaign hooks
All 32 backend test shards pass on cdb948aa6c9.
Targeted pytest for every path the strict pass touched: 685 + 310 + 114 + 185 + 123 + 56 + 40 tests pass serially, including the boundary-reading fixes for session counts, the legacy indexer cache, SMS enrollment, workflow configs, Jira issue types, backup import, and the Seer and integration clients.

Remaining failures are not caused by this branch:

pre-commit lint fails with Argument list too long because the stacked diff feeds 2,660 files to prek run --files; it passes on refactor(typing): enforce strong mypy checks across Sentry's entire Python codebase #129431 and locally in batches, and should clear after rebase.
migration drift and the OpenAPI schema diff also fail on refactor(typing): enforce strong mypy checks across Sentry's entire Python codebase #129431.
getsentry dispatch needs a same-repo token, and Vercel needs deploy authorization; neither runs for fork PRs.

Notes

snuba_sdk.* keeps follow_imports = "skip"; following it surfaces 963 errors and it is a deliberately tracked exception.
No new # type: ignore was added; the handful that remain are pre-existing and each carries an error code.

Shipping

Stacked on #129431. Merge that PR first.

After refactor(typing): enforce strong mypy checks across Sentry's entire Python codebase #129431 lands, rebase away the phase-one commits and keep only the 12 phase-two commits:
```
git fetch origin
git rebase --onto origin/master e7cba434a78 josh/typing-strict
git push --force-with-lease fork josh/typing-strict
```

Re-run the checks on the rebased head:

SENTRY_LOG_LEVEL=WARNING .venv/bin/python -m mypy
.venv/bin/prek run -q

pre-commit lint should now pass: its Argument list too long failure comes from feeding the stacked 2,660-file diff to prek run --files.
Undraft, get review approval, merge. The PR body and the commit both carry the strict-mode change; land it after phase one so master is never left with strict = true and unresolved errors.

Rebuild the accepted typing campaign on current master as an annotation-only change. Excludes the campaign evaluator, its policy modules, the campaign log, and owner-owned baseline edits. Behavior changes are carved into separate PRs; this commit preserves runtime behavior, including ReleaseFile.save positional compatibility and the release serializer's existing stats and newGroups output. Fix two refactor regressions: starred group search views keep their createdBy/starred/stars fields, and Seer commit lookups stay keyed by primary key instead of sha. Allow the weaklist guard to accept narrowing a wildcard into named modules, which is how the flags endpoints wildcard graduates.

OpenAPI schema generation raised UnableToProceedError for object-valued mappings and abstract Mapping/Sequence annotations introduced by the typing pass. Resolve object as a generic object and treat Mapping and Sequence like dict and list.

BasePaginator.build_queryset assigned the reversed ordering to queryset.query.order_b, a typo for order_by, when the queryset was already ordered by the descending key and a previous-page cursor needed ascending order. The assignment was silently dropped, so the previous page was filtered as if ascending but still sorted descending, returning the wrong rows. The ordering is now written to order_by, matching the sibling branch.

Exercise the branch that reverses an explicitly descending queryset for a previous-page cursor. Without the order_by fix, the previous page returns the current row instead of the preceding page.

Postponed annotations hide NotRequired from TypedDict runtime introspection, so ActorSerializerResponse.email was emitted as required and rejected the team-owner examples. Remove the postponed annotations from that module. Resolve object-valued mappings as unconstrained values instead of objects; the examples contain arrays and strings in those maps.

The release archive manifest and artifact index are user-uploaded JSON. Validate that the decoded value is an object before treating it as one, so malformed input fails at the boundary instead of raising AttributeError deeper in the archive reader. Review split: getsentry#129432.

Bring the reviewed behavior changes into the canonical campaign branch so the typing diff reflects the final contracts: release serializer defaults, Teams missing-id failure, blank origin preservation, Jira query preconditions, and the paginator previous-page fix.

The manifest root is validated, but `files` still needs an object shape before it is counted and indexed. Reject non-object values at the archive boundary and cover both malformed shapes. Review split: getsentry#129432.

Reject non-object thresholds, unknown or non-integer maximum keys, and non-string units at write time so the serializer casts are backed by a validated invariant. Review split: getsentry#129434.

Manifest headers come from uploaded archives. Treat any non-string debug-id value as absent instead of passing it into symbolic, and drop the casts that assumed the header value was always a string. Review split: getsentry#129435.

Gate new Any usage in graduated modules and require a `# justification:` comment for casts that read JSON, external responses, metadata, or config. The hook checks staged changes locally; `--base` and `--status` support branch-wide checks and campaign metrics. Run the hook against the stage and add tests for graduated/weaklisted modules, trust-boundary casts, and local casts.

Compare Any tokens against the file's previous revision instead of flagging every modified line that still contains an existing Any. Limit the whole-branch mode to Python files under src/.

Use a validating sourcemap-header helper instead of casting arbitrary header values, and document the remaining provider-response casts with explicit justifications so the campaign gate enforces the boundary.

Annotate the remaining functions in these modules and remove their weaklist entries. Full mypy passes with the stricter config, so these modules are now enforced rather than merely annotated.

getsentry deployments extend the admin broadcast response with plans, roles, trial status, region, platform, product, early-adopter, and organization filters. Model them as optional fields on the shared response so getsentry code stays type-compatible without a companion change.

Audit the full weaklist against a no-weaklist mypy run. These 399 modules and 12 wildcard families already pass the stricter config, so remove their exemptions and keep the enforcement they already satisfy.

`--goal` reports the remaining weaklist count and exits nonzero until it reaches zero, making the campaign target explicit and measurable.

Annotate the transfer payload and artifact download return types and remove the two weaklist exemptions. Full mypy and the endpoint tests pass.

Track the two burn-down shapes separately so wildcard narrowing and exact-module graduation are both visible.

Type the broadcast serializer/validator selection and API key serializer and endpoint parameters, then remove both weaklist exemptions. Full mypy and the endpoint tests pass.

Type the alert payload, notification responses, and metric-alert status lookup, then remove the weaklist exemption. Full mypy and the opsgenie tests pass.

Add precise annotations so each module passes strict mypy without its weaklist override. The uptime consumer tuple annotation follows the now-typed redis client.

Annotate 25 lane 1 modules so they type-check without the disallow_untyped_defs override. Leave EmptyListField weaklisted: its empty-string passthrough widens the return beyond ListField's contract, which would need Any or an override ignore to express.

Integrate the three lane batches and remove their weaklist entries. Fix the dashboard permission user-id widening and the Seer detected-issue mapping variance exposed by the stricter signatures.

Return an empty list for blank input so the field satisfies the DRF list contract; allowed-domain validation already treats empty input as invalid, so endpoint behavior is unchanged.

Annotate revision, version, and docker helpers and remove the package weaklist exemption. Full mypy passes.

Annotate the legacy culprit generators and remove the weaklist exemption. Full mypy and the culprit tests pass.

Replace the new Any usages introduced by the lane batches with precise contracts, and allow justified Any at dynamic boundaries in the gate.

Annotate the lane 6 modules so their disallow_untyped_defs weaklist entries can be removed. Widen two helper parameter annotations the new typed call sites exercise.

Remove the remaining `disallow_untyped_defs` weaklist entries owned by this lane, narrowing decoded JSON and framework values at their boundaries instead of leaving unchecked `Any`. Also repair test harness regressions from earlier typing lanes that prevented collection and teardown: lazy model imports in pytest plugins/conftest, self forwarding in the commit-hook override, and the runtime-safe generic base for the GitLab search endpoint. Widen `shortId` to reflect that a group may not have a qualified short id.

All source and test modules now pass strict mypy, so the campaign's disallow_untyped_defs weaklist and the remaining disable_error_code allowlist are removed. The stronglist check accepts absent lists, and the campaign gate compares a whole-branch diff against the base revision's weaklist instead of treating every changed module as newly graduated. Test helpers now type the elevated-mode request, the stored performance event, and the nullable metric and uptime inputs without changing runtime behavior.

Integration form defaults and choices are arbitrary JSON, so they keep the Any contract the schema generator can render. Client key choices are emitted as lists, which matches the documented array schema and avoids the tuple type drf-spectacular cannot resolve.

The backend typing job regenerates the module ignore block and fails on any resulting diff. With no ignored modules the generator now removes the block instead of appending an empty override.

Several annotated contracts were narrower than the values production actually carries, and a few checks changed semantics while becoming type-checkable. Restore the runtime behavior: - nullable file and blob sizes keep their documented None state, and Content-Length still reflects it; group.times_seen again treats only None as unset so reprocessing can seed a zero count - Snuba filter values and result keys keep their real types, with tag keys narrowed at the tagstore boundary - monitor environment routing consumes the URL environment parameter again; GitLab allows a missing installation id; GitHub Enterprise check_file returns the API response it always did - missing samples return None again, dropped orderbys stay None, and runs without a state id are serialized instead of asserted away

github-actions · 2026-11-01T07:13:08Z

🚨 Warning: This pull request contains Frontend and Backend changes!

It's discouraged to make changes to Sentry's Frontend and Backend in a single pull request. The Frontend and Backend are not atomically deployed. If the changes are interdependent of each other, they must be separated into two pull requests and be made forward or backwards compatible, such that the Backend or Frontend can be safely deployed independently.

Have questions? Please ask in the #discuss-dev-infra channel.

            # grab the filter out of the brackets of the string that looks
            # like so: members[value eq "123124"]
-            regex_search = re.search(r"\[(.*?)\]", operation["path"])
+            regex_search = re.search(r"\[(.*?)\]", cast(str, operation["path"]))


Full strict mode still reports 1,348 errors in 694 files, mostly Any values returned across dynamic boundaries. Record those per file and error code, and fail changed files that add a strict error or leave a stale baseline entry. This makes the remaining gap explicit and monotonic instead of invisible. Also enable disallow_incomplete_defs, remove three stale third-party import suppressions, and type the cache-key callables and role decorator that were losing their signatures at call sites.

Mypy strict can only check values that cross a typed boundary. Relay, selenium, slack_sdk, google.cloud, drf-spectacular and datadog all return Any, so callers either leaked Any or suppressed the diagnostic. Add local stubs for the surface we use and typed adapters where the library cannot be annotated, so the strict pass can see these values.

The client returned Any from every request, so integration call sites could not be checked and silently accepted any JSON shape. Give the request methods their real response type and add mapping/sequence accessors that validate the payload shape at the API boundary. Clients that parse a documented response narrow it locally instead of indexing Any.

Settings, caches, decoded JSON, and dashboard payloads all hand back Any. Read the expected type at the boundary instead of passing Any through, so the strict pass can follow these values into their callers. Where the runtime value is genuinely dynamic, cast at the boundary with the reason recorded next to it.

Settings, options, JSON fields, and untyped SDKs all hand back Any. Read the expected type where the value crosses into first-party code instead of letting Any flow into typed returns. Where the runtime value is genuinely dynamic, the cast says which contract it satisfies and why.

JSON fields, Redis replies, third-party SDKs, and DRF validate hooks all return Any. Check the shape where the value enters first-party code and keep the narrowed type, so callers no longer have to trust an implicit contract.

Provider SDKs, caches, and profile samples hand back decoded values with no usable annotation. Check their shape where first-party code consumes them, and keep the cast reason next to the few genuinely dynamic reads.

DRF validators, Django model metadata, request-response clients, and provider SDKs return Any at the framework seam. Check the value where it enters first-party code, so the declared return type is the type callers actually receive.

Tests, fixtures, and the last production modules returned values that their declared types could not prove: DRF and Django client responses, cached payloads, decoded JSON, mock call arguments, and SDK objects. Read each of those at the boundary that produced it so the declaration reflects what callers actually receive. Add stubs for the google.api_core surface the code uses and keep the Bigtable table stub precise, with the node-store fake casting only its own duck-typed rows.

Every first-party module now type-checks under strict mode, so the project config can require it directly instead of relying on a per-file error baseline. The ratchet, its baseline, and its gate have nothing left to track. BREAKING CHANGE: mypy runs with strict = true for all first-party code.

Reading values at their trust boundaries must keep what the boundary actually returned. Several validations were stricter than the data: session counts arrive as JSON numbers, the legacy indexer cache stores raw integers, unenrolled SMS interfaces have a null phone number, and workflow configs are arbitrary JSON. Accept those shapes again, keep Jira's issue type metadata a mapping, and let the GitHub Enterprise installation id stay an integer formatted into URLs. Revert the seer chat suggestion helper to its nullable result and register the GitLab note response as the empty JSON object the client parses.

joshuarli · 2026-11-01T16:06:25Z

Superseded by #129447, which combines this PR and its stacked counterpart into a single branch.

joshuarli added 30 commits October 31, 2026 09:49

fix(apidocs): resolve precise container hints

0dd5ddc

OpenAPI schema generation raised UnableToProceedError for object-valued mappings and abstract Mapping/Sequence annotations introduced by the typing pass. Resolve object as a generic object and treat Mapping and Sequence like dict and list.

test(api): cover previous-page ordering for descending querysets

0e0dad4

Exercise the branch that reverses an explicitly descending queryset for a previous-page cursor. Without the order_by fix, the previous page returns the current row instead of the preceding page.

fix(releases): validate manifest files shape

9eb2181

The manifest root is validated, but `files` still needs an object shape before it is counted and indexed. Reject non-object values at the archive boundary and cover both malformed shapes. Review split: getsentry#129432.

fix(dashboards): validate threshold JSON shape

2bcd33b

Reject non-object thresholds, unknown or non-integer maximum keys, and non-string units at write time so the serializer casts are backed by a validated invariant. Review split: getsentry#129434.

fix(artifactbundle): reject non-string debug ids

a189274

Manifest headers come from uploaded archives. Treat any non-string debug-id value as absent instead of passing it into symbolic, and drop the casts that assumed the header value was always a string. Review split: getsentry#129435.

fix(typing): count only new Any in the campaign gate

2dd3ccb

Compare Any tokens against the file's previous revision instead of flagging every modified line that still contains an existing Any. Limit the whole-branch mode to Python files under src/.

ref(typing): justify and validate trust-boundary casts

1e9e791

Use a validating sourcemap-header helper instead of casting arbitrary header values, and document the remaining provider-response casts with explicit justifications so the campaign gate enforces the boundary.

ref(typing): graduate 16 touched weaklist modules

8d17ea4

Annotate the remaining functions in these modules and remove their weaklist entries. Full mypy passes with the stricter config, so these modules are now enforced rather than merely annotated.

ref(typing): remove 411 clean weaklist entries

40e297b

Audit the full weaklist against a no-weaklist mypy run. These 399 modules and 12 wildcard families already pass the stricter config, so remove their exemptions and keep the enforcement they already satisfy.

feat(typing): add zero-weaklist campaign goal gate

e97b207

`--goal` reports the remaining weaklist count and exits nonzero until it reaches zero, making the campaign target explicit and measurable.

ref(typing): graduate two endpoint weaklist modules

5399f8e

Annotate the transfer payload and artifact download return types and remove the two weaklist exemptions. Full mypy and the endpoint tests pass.

feat(typing): report exact and wildcard weaklist counts

b5089e6

Track the two burn-down shapes separately so wildcard narrowing and exact-module graduation are both visible.

ref(typing): graduate two admin endpoint weaklist modules

8b7268a

Type the broadcast serializer/validator selection and API key serializer and endpoint parameters, then remove both weaklist exemptions. Full mypy and the endpoint tests pass.

ref(typing): graduate opsgenie client weaklist module

d2bb3b2

Type the alert payload, notification responses, and metric-alert status lookup, then remove the weaklist exemption. Full mypy and the opsgenie tests pass.

refactor(typing): graduate lane 2 weaklist modules

9cbf4aa

ref(typing): graduate lane 3 weaklist modules

18ba4f3

Add precise annotations so each module passes strict mypy without its weaklist override. The uptime consumer tuple annotation follows the now-typed redis client.

ref(typing): remove 76 graduated lane weaklist entries

6cdbc7f

Integrate the three lane batches and remove their weaklist entries. Fix the dashboard permission user-id widening and the Seer detected-issue mapping variance exposed by the stricter signatures.

ref(typing): graduate the empty list field

3511032

Return an empty list for blank input so the field satisfies the DRF list contract; allowed-domain validation already treats empty input as invalid, so endpoint behavior is unchanged.

ref(typing): graduate sentry package metadata module

fb091fe

Annotate revision, version, and docker helpers and remove the package weaklist exemption. Full mypy passes.

ref(typing): graduate culprit weaklist module

99c19d3

Annotate the legacy culprit generators and remove the weaklist exemption. Full mypy and the culprit tests pass.

fix(typing): keep graduated modules Any-free

ebfd3f5

Replace the new Any usages introduced by the lane batches with precise contracts, and allow justified Any at dynamic boundaries in the gate.

refactor(typing): graduate lane 6 weaklist modules

3af6fc2

Annotate the lane 6 modules so their disallow_untyped_defs weaklist entries can be removed. Widen two helper parameter annotations the new typed call sites exercise.

joshuarli added 5 commits October 31, 2026 20:27

fix(typing): keep the mypy allowlist absent when empty

85f2923

The backend typing job regenerates the module ignore block and fails on any resulting diff. With no ignored modules the generator now removes the block instead of appending an empty override.

github-actions Bot added Scope: Frontend Automatically applied to PRs that change frontend components Scope: Backend Automatically applied to PRs that change backend components labels Nov 1, 2026

joshuarli mentioned this pull request Nov 1, 2026

refactor(typing): enforce strong mypy checks across Sentry's entire Python codebase #129431

Closed

joshuarli force-pushed the josh/typing-strict branch 3 times, most recently from 5b1be2c to f86d249 Compare November 1, 2026 07:22

github-advanced-security Bot found potential problems Nov 1, 2026

View reviewed changes

Comment thread src/sentry/core/endpoints/scim/teams.py

# grab the filter out of the brackets of the string that looks

# like so: members[value eq "123124"]

regex_search = re.search(r"\[(.*?)\]", operation["path"])

regex_search = re.search(r"\[(.*?)\]", cast(str, operation["path"]))

joshuarli force-pushed the josh/typing-strict branch from f86d249 to 3463e32 Compare November 1, 2026 07:34

joshuarli added 10 commits November 1, 2026 01:23

ref(typing): validate values at untyped API boundaries

3666b62

JSON fields, Redis replies, third-party SDKs, and DRF validate hooks all return Any. Check the shape where the value enters first-party code and keep the narrowed type, so callers no longer have to trust an implicit contract.

ref(typing): read provider and cache payload types

bdc4a0c

Provider SDKs, caches, and profile samples hand back decoded values with no usable annotation. Check their shape where first-party code consumes them, and keep the cast reason next to the few genuinely dynamic reads.

ref(typing): read cache and SDK values as their declared types

899e62d

joshuarli changed the title ~~ref(typing): ratchet strict mypy errors~~ ref(typing)!: enable strict mypy repo-wide Nov 1, 2026

joshuarli mentioned this pull request Nov 1, 2026

ref(typing)!: enforce strict mypy across Sentry's Python codebase #129447

Draft

joshuarli closed this Nov 1, 2026

Navigation Menu

Repository navigation

ref(typing)!: enable strict mypy repo-wide - #129444

ref(typing)!: enable strict mypy repo-wide#129444
joshuarli wants to merge 103 commits into
getsentry:masterfrom
joshuarli:josh/typing-strict

joshuarli commented Nov 1, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Nov 1, 2026

Uh oh!

joshuarli commented Nov 1, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

joshuarli commented Nov 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Result

How the remaining errors were removed

Verification

Notes

Shipping

Uh oh!

github-actions Bot commented Nov 1, 2026

Uh oh!

joshuarli commented Nov 1, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

joshuarli commented Nov 1, 2026 •

edited

Loading