Summary

The backoff library's default logger produces a misleading message when user_defined_backoff_handler retries:

Backing off _send(...) for 0.0s (UserDefinedBackoffException: HTTP Status Code: 429...)

The 0.0s is the backoff library's interval parameter (set to 0), not the actual sleep time. The real sleep happens in the on_backoff callback (sleep_on_ratelimit) and is typically 60+ seconds. This generates confusion and incorrect bug reports from users.

Fix: Pass logger=None to backoff.on_exception() to suppress the library's default log handler, and improve the custom sleep_on_ratelimit callback to emit a single clear message:

UserDefinedBackoffException: Rate limit exceeded (HTTP 429). Retrying in 61 seconds.

No changes to retry logic or sleep behavior — only logging is affected.

Resolves https://github.com/airbytehq/airbyte-internal-issues/issues/16135
Related to https://github.com/airbytehq/oncall/issues/11837

Updates since last revision

• Prepended UserDefinedBackoffException: to both log messages in sleep_on_ratelimit per reviewer feedback, so the error source is immediately visible in logs.
• Fixed if exc.response: → exc.response is not None: requests.Response.__bool__ returns False for status >= 400, so the old check (pre-existing bug) would skip the HTTP status code branch on 429 responses.
• Log message now reports the actual sleep duration (retry_after + 1) instead of just retry_after, so the logged time matches the real sleep.
• Added kwargs.pop("logger", None) before passing logger=None to backoff.on_exception() to prevent TypeError if a caller ever passes logger in kwargs.

Review & Testing Checklist for Human

• exc.response is not None behavior change: The old code used if exc.response: which was falsey for HTTP error responses (including 429). The fix means the status code will now appear in logs where it previously did not. Confirm this is the desired behavior.
• Logged sleep time includes +1 buffer: Users will now see e.g. "Retrying in 61 seconds" (for a 60s backoff) instead of the old "Sleeping for 60 seconds". Confirm this is acceptable — it's more accurate but may look slightly odd.
• Response content no longer logged during retries: The old code logged raw Response Content: {exc.response.content!r} on every retry attempt. The new code omits this in favor of a cleaner user-facing message. Verify this trade-off is acceptable.
• Test plan: Trigger a 429 from any connector using a custom backoff strategy (e.g., source-amazon-seller-partner) and confirm: (1) no more "Backing off _send(...) for 0.0s" messages appear, (2) the new "UserDefinedBackoffException: Rate limit exceeded (HTTP 429). Retrying in {N} seconds." message appears instead, (3) retry/sleep behavior is unchanged.

Notes

• No existing unit tests cover the logging output of user_defined_backoff_handler. The change is logging-only and isolated to one function.
• The other backoff handlers (default_backoff_handler, http_client_default_backoff_handler, rate_limit_default_backoff_handler) are not affected — they use backoff.expo with real intervals so their default log messages are accurate.

Link to Devin session: https://app.devin.ai/sessions/2ced27cd22bd409f845816cd10a0c5f4