Support custom redaction and message length cap in LLMRetryPolicy - #70830
Merged
Conversation
Lee-W
force-pushed
the
llm-retry-redaction-hardening
branch
from
July 31, 2026 13:20
d685c71 to
268fe6f
Compare
Lee-W
marked this pull request as ready for review
August 5, 2026 04:42
Lee-W
force-pushed
the
llm-retry-redaction-hardening
branch
3 times, most recently
from
August 5, 2026 08:19
9f4ac87 to
5fd7b26
Compare
kaxil
approved these changes
Aug 5, 2026
kaxil
approved these changes
Aug 5, 2026
Lee-W
force-pushed
the
llm-retry-redaction-hardening
branch
3 times, most recently
from
August 6, 2026 09:23
7bc6b44 to
1c44094
Compare
Airflow's secrets masker only masks values already registered via mask_secret(), so free-text PII in an exception message — email addresses, customer names, account numbers — still reached the external LLM provider verbatim. Users who handle that kind of data need a way to apply their own redaction rules before the prompt leaves the worker. Nothing bounded the message size either, so a long traceback could inflate token cost without limit on every retry decision. The redact_exception flag this replaces was merged but has not shipped in any release, so it is removed outright rather than deprecated.
Previously, turning off redaction required passing the non-obvious lambda s: s as a custom redactor. Passing redactor=None now disables redaction explicitly, while the default behaviour (Airflow's secrets masker) is unchanged.
Every other optional LLMRetryPolicy parameter treats None as "use the
default", so redactor=None meaning "send the exception text unredacted"
was a footgun: assembling kwargs programmatically, as in
redactor=cfg.get("redactor"), silently shipped raw exception text to the
LLM provider whenever that key was absent. Disabling redaction is now
spelled redact_exception=False, and combining it with an explicit
redactor is rejected at construction time instead of one of the two
settings quietly winning.
The default masker also had no name a Dag author could import. The docs
pointed at airflow.sdk.log.redact, which is not part of the Task SDK's
documented public surface and has moved home four times -- which is why
common.compat.sdk carries a fallback chain for it. Exporting
default_redactor gives callers a stable path that survives the next move.
Two tests registered a secret without first resetting the masker, so they inherited patterns left by earlier tests and leaked their own secret into whatever ran next. That makes them order-dependent: passing or failing depended on what the rest of the module had already registered.
default_redactor read as an attribute rather than a callable, against the project's action-verb naming rule, and the docs example calls it directly. The new name also carries the limitation that matters most here: it masks only values already registered via mask_secret(), so a Dag author reading redactor=redact_registered_secrets learns the boundary from the name instead of having to reach the warning block to find it.
Lee-W
force-pushed
the
llm-retry-redaction-hardening
branch
from
August 7, 2026 07:23
1c44094 to
60b2ce0
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Airflow's secrets masker only masks values already registered via mask_secret(), so free-text PII in an exception message — email addresses, customer names, account numbers — still reached the external LLM provider verbatim. Users who handle that kind of data need a way to apply their own redaction rules before the prompt leaves the worker.
Nothing bounded the message size either, so a long traceback could inflate token cost without limit on every retry decision.
The redact_exception flag this replaces was merged but has not shipped in any release, so it is removed outright rather than deprecated.
Was generative AI tooling used to co-author this PR?
Generated-by: [Claude] following the guidelines
{pr_number}.significant.rst, in airflow-core/newsfragments. You can add this file in a follow-up commit after the PR is created so you know the PR number.