Aaron He created FLINK-40562:
--------------------------------
Summary: Preserve pathless claim mode across configuration round
trips
Key: FLINK-40562
URL: https://issues.apache.org/jira/browse/FLINK-40562
Project: Flink
Issue Type: Improvement
Components: Client / Job Submission
Reporter: Aaron He
h2. Problem
{{SavepointRestoreSettings.fromConfiguration()}} returns {{none()}} whenever
{{StateRecoveryOptions.SAVEPOINT_PATH}} is absent, even when
{{StateRecoveryOptions.RESTORE_MODE}} is explicitly configured.
The reconstructed settings therefore report the default {{NO_CLAIM}} instead of
an explicitly configured {{CLAIM}}. Serializing those settings into a fresh
{{Configuration}} also loses the explicit mode.
This is the configuration-reconstruction follow-up described in the "Related
but out of scope" section of
[FLINK-40559|https://issues.apache.org/jira/browse/FLINK-40559]. It is separate
from registering and parsing the command-line options.
Session-mode submission is the motivating path, but the affected method is
shared code, not a session-only implementation. Other callers that reconstruct
settings from a configuration without a savepoint path can encounter the same
loss. This report does not establish the end-to-end impact for every deployment
mode.
h2. Minimal example
The behavior follows directly from the current implementation and is explicitly
asserted by the existing {{testFromConfigurationWithNoPath}}. Source inspected
on {{master}} at
[3593f080f09fb9f974dca62279a98cc16505acb9|https://github.com/apache/flink/commit/3593f080f09fb9f974dca62279a98cc16505acb9].
This report does not claim an end-to-end cluster reproduction.
{noformat}
Configuration input = new Configuration();
input.set(StateRecoveryOptions.RESTORE_MODE, RecoveryClaimMode.CLAIM);
// No StateRecoveryOptions.SAVEPOINT_PATH.
SavepointRestoreSettings settings =
SavepointRestoreSettings.fromConfiguration(input);
settings.restoreSavepoint(); // false
settings.getRecoveryClaimMode(); // NO_CLAIM, although input contains CLAIM
Configuration output = new Configuration();
SavepointRestoreSettings.toConfiguration(settings, output);
output.containsKey(StateRecoveryOptions.RESTORE_MODE.key()); // false
{noformat}
The input configuration is not modified. The loss occurs in the reconstructed
settings and in a subsequent serialization to a fresh configuration.
h2. Submission-path relevance
After the parser changes proposed in
[apache/flink#29109|https://github.com/apache/flink/pull/29109] for FLINK-40559:
# {{ProgramOptions}} can hold a claim mode without a savepoint path.
# {{ProgramOptions.applyToConfiguration()}} serializes that explicit mode.
# {{ExecutionConfigAccessor.getSavepointRestoreSettings()}} calls
{{SavepointRestoreSettings.fromConfiguration()}} and loses the mode because the
path is absent.
# {{PipelineExecutorUtils}} uses the reconstructed settings when preparing the
{{StreamGraph}}.
An illustrative CLI input for this path is:
{noformat}
bin/flink run --claimMode CLAIM job.jar
{noformat}
On master before FLINK-40559, the parser can already drop the pathless option;
that is a separate, earlier loss. The direct configuration example above
isolates this issue without depending on the parser fix.
h2. Proposed behavior and scope
Preserve an explicitly configured recovery claim mode when reconstructing
settings without a savepoint path, while keeping {{restoreSavepoint()}} false
and {{getRestorePath()}} null. The pathless factory proposed in FLINK-40559
could be reused once available.
Preserve the explicit/unset distinction introduced by FLINK-39673:
* Explicit {{CLAIM}} and explicit {{NO_CLAIM}} must both survive reconstruction
and serialization into a fresh configuration.
* An absent claim mode must not become an explicitly written default.
* An absent {{SAVEPOINT_IGNORE_UNCLAIMED_STATE}} must not become an explicitly
written false value.
* Existing with-savepoint behavior must remain unchanged.
The existing {{testFromConfigurationWithNoPath}} deliberately expects
{{none()}} despite a configured {{CLAIM}}. This proposal changes that contract,
not just an untested branch. Maintainer agreement is needed on supporting
pathless claim mode in general configuration consumers and session submission.
Retaining the value alone does not establish that every execution path uses it
during HA recovery.
The intended scope is configuration reconstruction and propagation, not
changing checkpoint ownership rules or making a new job restore state without a
restore source.
h2. Suggested coverage
* Runtime configuration round trips without a path, covering explicit
{{CLAIM}}, explicit {{NO_CLAIM}}, and absent mode; inspect key presence as well
as effective values.
* Preserve the existing explicit/unset semantics for the
allow-non-restored-state option.
* Client coverage through {{ProgramOptions}} and {{ExecutionConfigAccessor}},
rather than only invoking the CLI parser directly, after FLINK-40559.
* Retain coverage for settings with a savepoint path.
h2. Related issues and duplicate search
Searched Apache Flink Jira across all statuses on 2026-09-05 using claimMode,
restoreMode, "claim mode", "restore mode", SavepointRestoreSettings,
ExecutionConfigAccessor, the claim-mode configuration key, and
pathless/fromConfiguration terms. No separate exact duplicate was found; this
is not a guarantee that no differently worded report exists.
* [FLINK-40559|https://issues.apache.org/jira/browse/FLINK-40559] (Open):
parent context; explicitly excludes this reconstruction issue.
* [FLINK-26316|https://issues.apache.org/jira/browse/FLINK-26316] (Open):
dynamic savepoint configuration overwritten by {{none()}} during
{{ProgramOptions.applyToConfiguration()}}, rather than a configured mode
discarded during reconstruction.
* [FLINK-34015|https://issues.apache.org/jira/browse/FLINK-34015] (Open):
dynamic properties overwritten by CLI defaults; its discussion describes a
restore path supplied via dynamic configuration.
* [FLINK-39673|https://issues.apache.org/jira/browse/FLINK-39673] (Closed,
Fixed): prevents unspecified defaults from overriding user configuration during
serialization. Its explicit/unset semantics must be preserved.
* [FLINK-28651|https://issues.apache.org/jira/browse/FLINK-28651] (Closed,
Fixed): REST {{JarRunHandler}} choosing a default instead of the configured
restore mode, a different consumer and failure point.
h2. Source references
* [SavepointRestoreSettings.fromConfiguration() and
toConfiguration()|https://github.com/apache/flink/blob/3593f080f09fb9f974dca62279a98cc16505acb9/flink-runtime/src/main/java/org/apache/flink/runtime/jobgraph/SavepointRestoreSettings.java]
* [Existing
testFromConfigurationWithNoPath|https://github.com/apache/flink/blob/3593f080f09fb9f974dca62279a98cc16505acb9/flink-runtime/src/test/java/org/apache/flink/runtime/jobgraph/SavepointRestoreSettingsTest.java]
*
[ProgramOptions.applyToConfiguration()|https://github.com/apache/flink/blob/3593f080f09fb9f974dca62279a98cc16505acb9/flink-clients/src/main/java/org/apache/flink/client/cli/ProgramOptions.java#L179-L189]
*
[ExecutionConfigAccessor.getSavepointRestoreSettings()|https://github.com/apache/flink/blob/3593f080f09fb9f974dca62279a98cc16505acb9/flink-clients/src/main/java/org/apache/flink/client/cli/ExecutionConfigAccessor.java#L93-L95]
* [PipelineExecutorUtils propagation to
StreamGraph|https://github.com/apache/flink/blob/3593f080f09fb9f974dca62279a98cc16505acb9/flink-clients/src/main/java/org/apache/flink/client/deployment/executors/PipelineExecutorUtils.java#L107-L110]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
