Skip to content

Fix API client errors losing detail behind a reverse proxy - #75493

Merged
henry3260 merged 13 commits into
apache:mainfrom
rjgoyln:ctl-json-media-type-errors
Oct 25, 2026
Merged

henry3260 merged 13 commits into
apache:mainfrom
rjgoyln:ctl-json-media-type-errors

Conversation

@rjgoyln

@rjgoyln rjgoyln commented Oct 23, 2026

Copy link
Copy Markdown
Contributor

Continues #70936, which was approved and then closed by the open PR limit.

Summary

Both API clients parse an error body only when content-type is exactly application/json. Behind a proxy that rewrites it to application/json; charset=utf-8, or one answering with application/problem+json:

  • airflowctl shows only the HTTP status line, not the server's error message.
  • The Task SDK raises a bare HTTPStatusError instead of ServerResponseError, so the status-specific handling in the supervisor and the client's operations (a 404 meaning a missing connection or variable, for example) stops matching.

Airflow's own API server sends the bare media type, so this needs a proxy in between.

Change

  • Match the media type case-insensitively, ignoring parameters and accepting +json suffixes, in both clients.
  • In airflowctl, treat an undecodable JSON body as a plain HTTP error, and log a non-object JSON body instead of crashing on it.

Since the approval on #70936 the branch is rebased onto the Task SDK's httpx2 migration (#73116), and the newsfragment covers only the Task SDK side, which is what ships in airflow-core.


Was generative AI tooling used to co-author this PR?
  • Yes — Claude Code (Opus 5.5)

Generated-by: Claude Code (Opus 5.5) following the guidelines

airflowctl compared the error response content-type against the exact
string "application/json". Any reverse proxy, ingress, or API gateway
that returns "application/json; charset=utf-8" — or emits its own
"application/problem+json" error — fell outside that comparison, so
ServerResponseError was never constructed and the friendly "not found"
messages in the dags and tasks commands were skipped. The bare
httpx.HTTPStatusError raised instead is not handled by
safe_call_command, so users saw a raw traceback. Airflow's own API
server does not send a charset, which is why this only surfaces once
something sits in front of it.

Media types are case-insensitive and may carry parameters (RFC 9110),
so the header has to be parsed rather than compared whole.

Matching more responses also routes bodies into the error path that are
valid JSON but not objects, which dict() rejected with a ValueError from
that same unprotected path, and bodies that do not decode at all.
The Task SDK client carries the identical exact-string content-type
comparison, so a task talking to the API server through a proxy that
declares a charset loses the server's error detail and, because the
raised exception is then a plain httpx.HTTPStatusError, no longer hits
the status-specific handling in the supervisor and the task runner.
The airflowctl and Task SDK copies are the same three lines with
different comments, and they already drifted in the commits that
created them. A pointer in each makes the next person notice there
are two.
Requested during review: the Task SDK half of the change ships in
airflow-core, and both effects are visible to users running behind a
proxy.
The Task SDK moved from httpx to httpx2 (apache#73116) after this branch was
written, so the rebased helper and its tests referred to a module that
is no longer imported there.
Since apache#73307 airflowctl catches the bare HTTPStatusError and prints one
line instead of a traceback, so the newsfragment overstated it; the
server's own message is still what gets lost.
The bare media type and the 404/500 axis repeat what
test_error_parsing_with_unread_body already covers, and dict() of an
object body worked before; none of them fail without the fix.
The previous comment described the dict() call the guard replaced,
which is no longer in the code.
airflow-core's release notes ship only the Task SDK side; airflowctl's
changelog is built from git log. The task runner never handled
ServerResponseError, so it is no longer named.
The longer versions restated what the diff and the PR already explain,
and the two _has_json_body copies now carry the same note.

@henry3260 henry3260 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the fix! LGTM

@henry3260
henry3260 enabled auto-merge (squash) October 25, 2026 17:05
@henry3260

Copy link
Copy Markdown
Contributor

It’s weird that it’s taking so long to get merged.

@henry3260
henry3260 disabled auto-merge October 25, 2026 18:43
@henry3260
henry3260 enabled auto-merge (squash) October 25, 2026 18:43
@rjgoyln

rjgoyln commented Oct 25, 2026

Copy link
Copy Markdown
Contributor Author

Thanks for the review and the branch update!

@henry3260
henry3260 merged commit 352ca60 into apache:main Oct 25, 2026
97 checks passed
@github-actions

Copy link
Copy Markdown
Contributor

Backport successfully created: airflow-ctl/v0-1-test

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status Branch Result
✅ airflow-ctl/v0-1-test PR Link

potiuk pushed a commit that referenced this pull request Oct 25, 2026
…reverse proxy (#75493) (#75622)

* Fix airflowctl tracebacks on JSON errors behind a reverse proxy

airflowctl compared the error response content-type against the exact
string "application/json". Any reverse proxy, ingress, or API gateway
that returns "application/json; charset=utf-8" — or emits its own
"application/problem+json" error — fell outside that comparison, so
ServerResponseError was never constructed and the friendly "not found"
messages in the dags and tasks commands were skipped. The bare
httpx.HTTPStatusError raised instead is not handled by
safe_call_command, so users saw a raw traceback. Airflow's own API
server does not send a charset, which is why this only surfaces once
something sits in front of it.

Media types are case-insensitive and may carry parameters (RFC 9110),
so the header has to be parsed rather than compared whole.

Matching more responses also routes bodies into the error path that are
valid JSON but not objects, which dict() rejected with a ValueError from
that same unprotected path, and bodies that do not decode at all.

* Move media type rationale from docstring into an inline comment

* Apply the same JSON media type handling to the Task SDK client

The Task SDK client carries the identical exact-string content-type
comparison, so a task talking to the API server through a proxy that
declares a charset loses the server's error detail and, because the
raised exception is then a plain httpx.HTTPStatusError, no longer hits
the status-specific handling in the supervisor and the task runner.

* Cross-reference the twin JSON media type checks

The airflowctl and Task SDK copies are the same three lines with
different comments, and they already drifted in the commits that
created them. A pointer in each makes the next person notice there
are two.

* Add newsfragment for the API client error handling fix

Requested during review: the Task SDK half of the change ships in
airflow-core, and both effects are visible to users running behind a
proxy.

* Use httpx2 in the Task SDK JSON media type check

The Task SDK moved from httpx to httpx2 (#73116) after this branch was
written, so the rebased helper and its tests referred to a module that
is no longer imported there.

* Describe the airflowctl symptom as it is on main now

Since #73307 airflowctl catches the bare HTTPStatusError and prints one
line instead of a traceback, so the newsfragment overstated it; the
server's own message is still what gets lost.

* Keep only the airflowctl test cases the fix changes

The bare media type and the 404/500 axis repeat what
test_error_parsing_with_unread_body already covers, and dict() of an
object body worked before; none of them fail without the fix.

* Explain the non-object log guard by what it protects

The previous comment described the dict() call the guard replaced,
which is no longer in the code.

* Limit the newsfragment to the Task SDK fix

airflow-core's release notes ship only the Task SDK side; airflowctl's
changelog is built from git log. The task runner never handled
ServerResponseError, so it is no longer named.

* Rename the newsfragment to this PR's number

* Shorten the media type comments

The longer versions restated what the diff and the PR already explain,
and the two _has_json_body copies now carry the same note.

---------
(cherry picked from commit 352ca60)

Co-authored-by: rjgoyln <151457491+rjgoyln@users.noreply.github.com>
Co-authored-by: Henry Chen <henrychen@apache.org>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants