[2.8] Narrow client failure reporting

[YuanTingHsieh]

Summary

Narrow the client job-failure reporting path added for 2.8 so generic launcher JobReturnCode.EXECUTION_ERROR is not promoted into an authoritative server-side job failure.

Why

client_api_qa can complete the FL workflow successfully and then hit local worker teardown noise that surfaces as a generic launcher execution error. Reporting that generic code to the server causes the server to fail an already-finished job.

This keeps the explicit failure cases used by the recent job-timeout status fix:

• ProcessExitCode.EXCEPTION
• ProcessExitCode.CONFIG_ERROR
• ProcessExitCode.UNSAFE_COMPONENT
• JobReturnCode.ABORTED

The K8s pending-timeout path still reports JobReturnCode.EXCEPTION, so it should continue to show FINISHED:EXECUTION_EXCEPTION rather than RUNNING.

[Copilot]

Pull request overview

This PR narrows the client → server “job failure” reporting pathway introduced in 2.8 so that a generic launcher-level JobReturnCode.EXECUTION_ERROR reported by a client does not get promoted into an authoritative server-side job failure (avoiding failing jobs that already completed successfully server-side).

Changes:

• Server: stop treating reported JobReturnCode.EXECUTION_ERROR as a run-failing signal in process_job_failure.
• Client: stop reporting JobReturnCode.EXECUTION_ERROR as a “reportable job failure” to the server.
• Tests: update and add unit tests to enforce the narrower reporting/handling contract.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
nvflare/private/fed/server/fed_server.py	Narrows process_job_failure to only fail runs for reported CONFIG_ERROR / EXCEPTION (no longer for EXECUTION_ERROR).
nvflare/private/fed/client/client_executor.py	Removes EXECUTION_ERROR from REPORTABLE_JOB_FAILURES, preventing it from being reported to the server.
tests/unit_test/private/fed/server/fed_server_test.py	Updates parametrized expectations and adds a test asserting EXECUTION_ERROR is ignored by the server.
tests/unit_test/private/fed/client/client_executor_test.py	Updates expected reportable failures and asserts EXECUTION_ERROR is non-reportable from the client.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[greptile-apps]

Greptile Summary

This PR narrows the set of client job-failure codes that are promoted to an authoritative server-side job failure, fixing a false-failure scenario where client_api_qa worker teardown noise (surfaced as JobReturnCode.EXECUTION_ERROR) was causing the server to fail an already-completed job.

• client_executor.py: Removes JobReturnCode.EXECUTION_ERROR from REPORTABLE_JOB_FAILURES, so clients no longer send a failure message to the server for generic launcher exit codes.
• fed_server.py: Removes JobReturnCode.EXECUTION_ERROR from the fail_run branch in process_job_failure, ensuring even legacy clients that still send this code will have it silently ignored rather than triggering a job failure.
• Tests: Updated parametrized and new explicit tests cleanly cover both the retained fail/abort paths and the newly ignored EXECUTION_ERROR path.

Confidence Score: 5/5

Safe to merge — the change removes a single over-broad failure code from a two-layer guard (client reporting + server handling), the retained codes match the PR description, and the new test explicitly exercises the ignored path.

The change is small and surgical: one dict entry removed on the client, one tuple element removed on the server, and tests updated consistently. The K8s pending-timeout path continues to report JobReturnCode.EXCEPTION so that path is unaffected. No regressions are apparent.

No files require special attention.

Important Files Changed

Filename	Overview
nvflare/private/fed/client/client_executor.py	Removes JobReturnCode.EXECUTION_ERROR from REPORTABLE_JOB_FAILURES; the retained set (EXCEPTION, UNSAFE_COMPONENT, CONFIG_ERROR, ABORTED) is consistent with the server-side handling.
nvflare/private/fed/server/fed_server.py	Removes JobReturnCode.EXECUTION_ERROR from the fail_run branch; the two-branch if/elif structure (fail vs. stop) is preserved and correctly aligned with the client-side reportable set.
tests/unit_test/private/fed/client/client_executor_test.py	Moves EXECUTION_ERROR into the non-failure parametrize set and updates the expected-failures dict to match production code.
tests/unit_test/private/fed/server/fed_server_test.py	Adds a new explicit test confirming EXECUTION_ERROR leaves fail_run and stop_run uncalled; existing parametrized test correctly trimmed to the two remaining fail-run codes.

Sequence Diagram

sequenceDiagram
    participant Client
    participant Server

    Note over Client: Job finishes successfully
    Note over Client: Worker teardown raises noise<br/>(EXECUTION_ERROR)

    alt Before this PR
        Client->>Server: process_job_failure(EXECUTION_ERROR)
        Server->>Server: fail_run(job_id, EXCEPTION)
        Note over Server: Job incorrectly marked FAILED
    else After this PR
        Note over Client: EXECUTION_ERROR not in REPORTABLE_JOB_FAILURES
        Note over Client: No message sent to server
        Note over Server: Job stays FINISHED (correct)
    end

    Note over Client: EXCEPTION / CONFIG_ERROR / UNSAFE_COMPONENT / ABORTED
    Client->>Server: process_job_failure(code)
    alt code in (EXCEPTION, CONFIG_ERROR)
        Server->>Server: fail_run(job_id, EXCEPTION)
    else code in (UNSAFE_COMPONENT, ABORTED)
        Server->>Server: stop_run(job_id)
    end

Loading

_{Reviews (2): Last reviewed commit: "Merge branch '2.8' into codex/narrow-cli..." | Re-trigger Greptile}