Aaron He created FLINK-40559:
--------------------------------

             Summary: Application mode silently ignores the --claimMode / 
--restoreMode command-line option
                 Key: FLINK-40559
                 URL: https://issues.apache.org/jira/browse/FLINK-40559
             Project: Flink
          Issue Type: Bug
          Components: Client / Job Submission, Deployment / Kubernetes
    Affects Versions: 2.2.1, 2.3.0, 2.1.0, 1.20.2, 1.18.1, 2.0.0
            Reporter: Aaron He


Flink silently discards an explicitly passed {{--claimMode}} (or its deprecated 
alias {{{}-restoreMode{}}}) command-line argument in two independent places. 
The user receives no warning or error; the recovery claim mode falls back to 
{{{}NO_CLAIM{}}}.

Verified on {{master}} at 
[b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0|https://github.com/apache/flink/commit/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0].
h2. References
 * [Application-mode parser 
options|https://github.com/apache/flink/blob/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0/flink-container/src/main/java/org/apache/flink/container/entrypoint/StandaloneApplicationClusterConfigurationParserFactory.java#L75-L86]
 * [CommandLineParser using stopAtNonOption = 
true|https://github.com/apache/flink/blob/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0/flink-runtime/src/main/java/org/apache/flink/runtime/entrypoint/parser/CommandLineParser.java#L40-L50]
 * [createSavepointRestoreSettings() savepoint-path 
gate|https://github.com/apache/flink/blob/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0/flink-clients/src/main/java/org/apache/flink/client/cli/CliFrontendParser.java#L654-L686]
 * [SavepointRestoreSettings.toConfiguration() / 
fromConfiguration()|https://github.com/apache/flink/blob/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0/flink-runtime/src/main/java/org/apache/flink/runtime/jobgraph/SavepointRestoreSettings.java#L195-L229]
 * [Application-mode direct configuration 
propagation|https://github.com/apache/flink/blob/b53ea1b1c4d5ab0ec882559a4a050d82ae5740e0/flink-container/src/main/java/org/apache/flink/container/entrypoint/StandaloneApplicationClusterEntryPoint.java#L115-L121]
 * FLINK-39673 and its merged fix, 
[apache/flink#28295|https://github.com/apache/flink/pull/28295]. The PR title 
says {{{}FLINK-36973{}}}, but its body correctly links {{{}FLINK-39673{}}}.
 * [Flink code contribution 
process|https://flink.apache.org/how-to-contribute/contribute-code/]
 * [Flink pull-request quality 
guide|https://flink.apache.org/how-to-contribute/code-style-and-quality-pull-requests/]

h2. Defect A — the option is never registered in application mode

{{StandaloneApplicationClusterConfigurationParserFactory.getOptions()}} 
registers {{SAVEPOINT_PATH_OPTION}} and 
{{{}SAVEPOINT_ALLOW_NON_RESTORED_OPTION{}}}, but not {{SAVEPOINT_CLAIM_MODE}} 
or {{{}SAVEPOINT_RESTORE_MODE{}}}:
{code:java}
options.addOption(CliFrontendParser.SAVEPOINT_PATH_OPTION);
options.addOption(CliFrontendParser.SAVEPOINT_ALLOW_NON_RESTORED_OPTION);
// no SAVEPOINT_CLAIM_MODE / SAVEPOINT_RESTORE_MODE
{code}
{{CommandLineParser}} calls {{parser.parse(options, args, true)}} with 
{{{}stopAtNonOption = true{}}}, so an unrecognized {{--claimMode}} does not 
raise an error. It terminates option parsing and the remainder is swallowed 
into {{{}getArgs(){}}}. The claim mode is therefore unreachable in application 
mode regardless of how it is passed.
h2. Defect B — the parse is gated on --fromSavepoint

{{CliFrontendParser.createSavepointRestoreSettings()}} reads the claim mode 
only inside the savepoint-path branch:
{code:java}
if (commandLine.hasOption(SAVEPOINT_PATH_OPTION.getOpt())) {
    ...
    if (commandLine.hasOption(SAVEPOINT_CLAIM_MODE)) { ... }
    return SavepointRestoreSettings.forPath(
            savepointPath, allowNonRestoredState, recoveryClaimMode);
} else {
    return SavepointRestoreSettings.none(); // claim mode discarded
}
{code}
When {{--claimMode}} is passed without {{{}--fromSavepoint{}}}, the option is 
parsed and then dropped. This is the normal shape for an HA-recovery restart, 
where the JobManager resolves its checkpoint from HA storage rather than from a 
command-line savepoint path.

The two defects are independent: fixing A alone still leaves B.
h2. How to reproduce

Add these tests to 
{{{}StandaloneApplicationClusterConfigurationParserFactoryTest{}}}:
{code:java}
@Test
void testClaimModeWithSavepoint() throws FlinkParseException {
    final String restorePath = "s3://test/savepoint";
    final String[] args = {
        "-c", confDirPath, "-j", JOB_CLASS_NAME,
        "--fromSavepoint", restorePath,
        "--claimMode", RecoveryClaimMode.CLAIM.name()
    };

    final SavepointRestoreSettings settings =
            commandLineParser.parse(args).getSavepointRestoreSettings();

    assertThat(settings.getRestorePath()).isEqualTo(restorePath);
    
assertThat(settings.getRecoveryClaimMode()).isEqualTo(RecoveryClaimMode.CLAIM);
}

@Test
void testClaimModeWithoutSavepoint() throws FlinkParseException {
    final String[] args = {
        "-c", confDirPath, "-j", JOB_CLASS_NAME,
        "--claimMode", RecoveryClaimMode.CLAIM.name()
    };

    final SavepointRestoreSettings settings =
            commandLineParser.parse(args).getSavepointRestoreSettings();

    assertThat(settings.restoreSavepoint()).isFalse();
    
assertThat(settings.getRecoveryClaimMode()).isEqualTo(RecoveryClaimMode.CLAIM);
}
{code}
Run:
{code:bash}
./mvnw -pl flink-container 
-Dtest='StandaloneApplicationClusterConfigurationParserFactoryTest#testClaimModeWithSavepoint+testClaimModeWithoutSavepoint'
 test
{code}
*Observed on unmodified {{{}master{}}}:* both tests fail because Defect A masks 
Defect B.
{noformat}
[ERROR] Tests run: 2, Failures: 2, Errors: 0, Skipped: 0
[ERROR] ...testClaimModeWithSavepoint
expected: CLAIM
 but was: NO_CLAIM
[ERROR] ...testClaimModeWithoutSavepoint
expected: CLAIM
 but was: NO_CLAIM
{noformat}
To isolate Defect B, temporarily add {{SAVEPOINT_CLAIM_MODE}} and 
{{SAVEPOINT_RESTORE_MODE}} to {{getOptions()}} and rerun:
{noformat}
[ERROR] Tests run: 2, Failures: 1, Errors: 0, Skipped: 0
[ERROR] ...testClaimModeWithoutSavepoint
expected: CLAIM
 but was: NO_CLAIM
{noformat}
{{testClaimModeWithSavepoint}} now passes while 
{{testClaimModeWithoutSavepoint}} still fails, confirming that the two defects 
are independent.
h2. Relationship to FLINK-39673

FLINK-39673 / [PR #28295|https://github.com/apache/flink/pull/28295] made 
{{SavepointRestoreSettings.toConfiguration()}} skip options that were not 
explicitly set, and {{none()}} is now {{{}(null, null, null){}}}. As a result, 
an unset claim mode is no longer written to the {{{}Configuration{}}}, so a 
value supplied through {{flink-conf.yaml}} / {{flinkConfiguration}} survives 
correctly.

This ticket is therefore not about configuration being overwritten. The 
remaining defect is narrower: an explicitly passed command-line {{--claimMode}} 
is silently ignored, so the CLI option is non-functional in application mode.
h2. Proposed fix

Two small, independently reviewable commits:
 # *{{{}flink-container{}}}:* Register {{SAVEPOINT_CLAIM_MODE}} and 
{{SAVEPOINT_RESTORE_MODE}} in 
{{{}StandaloneApplicationClusterConfigurationParserFactory.getOptions(){}}}.
 # *{{flink-clients}} / {{{}flink-runtime{}}}:* Restructure 
{{createSavepointRestoreSettings()}} to branch on option presence rather than 
on the presence of a savepoint path, and add a factory on 
{{SavepointRestoreSettings}} that carries a claim mode with no restore path.

The second change must preserve the post-FLINK-39673 "explicitly set" 
semantics. {{allowNonRestoredState}} must remain {{null}} when the flag is 
absent, so the new code path does not begin writing 
{{execution.state-recovery.ignore-unclaimed-state}} into the {{Configuration}} 
where it previously wrote nothing.

Behavior is unchanged for jobs that do not pass {{{}--claimMode{}}}.
h2. Related but out of scope

{{SavepointRestoreSettings.fromConfiguration()}} applies the same 
savepoint-path gate, so the {{flink run}} client path through 
{{ExecutionConfigAccessor}} also drops a pathless claim mode on the 
{{Configuration}} round-trip. Application mode does not go through that path: 
{{StandaloneApplicationClusterEntryPoint}} calls {{toConfiguration()}} 
directly. It is therefore not required to fix the defect described here. I am 
happy to include it or split it out, whichever the reviewer prefers.

I have both fixes and test coverage prepared and am happy to take this ticket.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to