[FLINK-35414] Rework last-state upgrade mode to support job cancellation as suspend mechanism

Author: gyfora

docs/content/docs/custom-resource/job-management.md

@@ -84,13 +84,13 @@ Supported values: `stateless`, `savepoint`, `last-state`
 
 The `upgradeMode` setting controls both the stop and restore mechanisms as detailed in the following table:
 
-|                        | Stateless               | Last State                                 | Savepoint                              |
-|------------------------|-------------------------|--------------------------------------------|----------------------------------------|
-| Config Requirement     | None                    | Checkpointing & HA Enabled                 | Checkpoint/Savepoint directory defined |
-| Job Status Requirement | None                    | HA metadata available                      | Job Running*                           |
-| Suspend Mechanism      | Cancel / Delete         | Delete Flink deployment (keep HA metadata) | Cancel with savepoint                  |
-| Restore Mechanism      | Deploy from empty state | Recover last state using HA metadata       | Restore From savepoint                 |
-| Production Use         | Not recommended         | Recommended                                | Recommended                            |
+|                        | Stateless       | Last State                         | Savepoint                              |
+|------------------------|-----------------|------------------------------------|----------------------------------------|
+| Config Requirement     | None            | Checkpointing Enabled              | Checkpoint/Savepoint directory defined |
+| Job Status Requirement | None            | Job or HA metadata accessible      | Job Running*                           |
+| Suspend Mechanism      | Cancel / Delete | Cancel / Delete (keep HA metadata) | Cancel with savepoint                  |
+| Restore Mechanism      | Empty state     | Use HA metadata or last cp/sp      | Restore From savepoint                 |
+| Production Use         | Not recommended | Recommended                        | Recommended                            |
 
 
 *\* When HA is enabled the `savepoint` upgrade mode may fall back to the `last-state` behaviour in cases where the job is in an unhealthy state.*
@@ -149,10 +149,6 @@ spec:
     state: running
 ```
 
-{{< hint warning >}}
-Last state upgrade mode is currently only supported for `FlinkDeployments`.
-{{< /hint >}}
-
 ### Application restarts without spec change
 
 There are cases when users would like to restart the Flink deployments to deal with some transient problem.

docs/content/docs/custom-resource/overview.md

@@ -215,10 +215,6 @@ COPY flink-hadoop-fs-1.19-SNAPSHOT.jar $FLINK_PLUGINS_DIR/hadoop-fs/
 
 Alternatively, if you use helm to install flink-kubernetes-operator, it allows you to specify a postStart hook to download the required plugins.
 
-### Limitations
-
-- Last-state upgradeMode is currently not supported for FlinkSessionJobs
-
 ## Further information
 
  - [Snapshots]({{< ref "docs/custom-resource/snapshots" >}})

docs/content/docs/custom-resource/reference.md

@@ -452,7 +452,7 @@ This serves as a full reference for FlinkDeployment and FlinkSessionJob custom r
 
 | Parameter | Type | Docs |
 | ----------| ---- | ---- |
-| lastSavepoint | org.apache.flink.kubernetes.operator.api.status.Savepoint | Last completed savepoint by the operator for manual and periodic snapshots. Only used if FlinkStateSnapshot resources are disabled. |
+| lastSavepoint | org.apache.flink.kubernetes.operator.api.status.Savepoint | Last completed savepoint by the operator. |
 | triggerId | java.lang.String | Trigger id of a pending savepoint operation. |
 | triggerTimestamp | java.lang.Long | Trigger timestamp of a pending savepoint operation. |
 | triggerType | org.apache.flink.kubernetes.operator.api.status.SnapshotTriggerType | Savepoint trigger mechanism. |

docs/layouts/shortcodes/generated/dynamic_section.html

@@ -122,6 +122,12 @@
             <td>Boolean</td>
             <td>Enables last-state fallback for savepoint upgrade mode. When the job is not running thus savepoint cannot be triggered but HA metadata is available for last state restore the operator can initiate the upgrade process when the flag is enabled.</td>
         </tr>
+        <tr>
+            <td><h5>kubernetes.operator.job.upgrade.last-state.job-cancel.enabled</h5></td>
+            <td style="word-wrap: break-word;">false</td>
+            <td>Boolean</td>
+            <td>Cancel jobs during last-state upgrade. This config is ignored for session jobs where cancel is the only mechanism to perform this type of upgrade.</td>
+        </tr>
         <tr>
             <td><h5>kubernetes.operator.job.upgrade.last-state.max.allowed.checkpoint.age</h5></td>
             <td style="word-wrap: break-word;">(none)</td>

docs/layouts/shortcodes/generated/kubernetes_operator_config_configuration.html

@@ -212,6 +212,12 @@
             <td>Boolean</td>
             <td>Enables last-state fallback for savepoint upgrade mode. When the job is not running thus savepoint cannot be triggered but HA metadata is available for last state restore the operator can initiate the upgrade process when the flag is enabled.</td>
         </tr>
+        <tr>
+            <td><h5>kubernetes.operator.job.upgrade.last-state.job-cancel.enabled</h5></td>
+            <td style="word-wrap: break-word;">false</td>
+            <td>Boolean</td>
+            <td>Cancel jobs during last-state upgrade. This config is ignored for session jobs where cancel is the only mechanism to perform this type of upgrade.</td>
+        </tr>
         <tr>
             <td><h5>kubernetes.operator.job.upgrade.last-state.max.allowed.checkpoint.age</h5></td>
             <td style="word-wrap: break-word;">(none)</td>

e2e-tests/test_sessionjob_operations.sh

@@ -54,6 +54,7 @@ if [ "$location" == "" ];then
   exit 1
 fi
 
+echo "Starting sessionjob savepoint upgrade test"
 # Testing savepoint mode upgrade
 # Update the FlinkSessionJob and trigger the savepoint upgrade
 kubectl patch sessionjob ${SESSION_JOB_NAME} --type merge --patch '{"spec":{"job": {"parallelism": 1 } } }'
@@ -67,6 +68,24 @@ assert_available_slots 1 $CLUSTER_ID
 
 echo "Successfully run the sessionjob savepoint upgrade test"
 
+flink_version=$(kubectl get $SESSION_CLUSTER_IDENTIFIER -o yaml | yq '.spec.flinkVersion')
+
+if [ "$flink_version" != "v1_16" ]; then
+  echo "Starting sessionjob last-state upgrade test"
+  # Testing last-state mode upgrade
+  # Update the FlinkSessionJob and trigger the last-state upgrade
+  kubectl patch sessionjob ${SESSION_JOB_NAME} --type merge --patch '{"spec":{"job": {"parallelism": 2, "upgradeMode": "last-state" } } }'
+
+  # Check the job was restarted with the new parallelism
+  wait_for_status $SESSION_JOB_IDENTIFIER '.status.jobStatus.state' CANCELLING ${TIMEOUT} || exit 1
+  wait_for_status $SESSION_JOB_IDENTIFIER '.status.jobStatus.state' RUNNING ${TIMEOUT} || exit 1
+  assert_available_slots 0 $CLUSTER_ID
+
+  echo "Successfully run the sessionjob last-state upgrade test"
+else
+  echo "Skipping last-state test for flink version 1.16"
+fi
+
 # Test Operator restart
 echo "Delete session job " + $SESSION_JOB_NAME
 kubectl delete flinksessionjob $SESSION_JOB_NAME

flink-kubernetes-operator-api/src/main/java/org/apache/flink/kubernetes/operator/api/docs/CrdReferenceDoclet.java

@@ -219,6 +219,9 @@ public Void scan(Element e, Integer depth) {
                     }
                     break;
                 case FIELD:
+                    if (e.getModifiers().contains(Modifier.STATIC)) {
+                        return null;
+                    }
                     out.println(
                             "| "
                                     + getNameOrJsonPropValue(e)

flink-kubernetes-operator-api/src/main/java/org/apache/flink/kubernetes/operator/api/status/CommonStatus.java

@@ -18,12 +18,10 @@
 package org.apache.flink.kubernetes.operator.api.status;
 
 import org.apache.flink.annotation.Experimental;
-import org.apache.flink.annotation.Internal;
 import org.apache.flink.kubernetes.operator.api.lifecycle.ResourceLifecycleState;
 import org.apache.flink.kubernetes.operator.api.spec.AbstractFlinkSpec;
 import org.apache.flink.kubernetes.operator.api.spec.JobState;
 
-import com.fasterxml.jackson.annotation.JsonIgnore;
 import io.fabric8.crd.generator.annotation.PrinterColumn;
 import lombok.AllArgsConstructor;
 import lombok.Data;
@@ -98,10 +96,4 @@ public ResourceLifecycleState getLifecycleState() {
 
         return ResourceLifecycleState.DEPLOYED;
     }
-
-    /**
-     * Internal flag to signal that due to some condition we need to schedule a new reconciliation
-     * loop immediately. For example autoscaler overrides have changed and we need to apply them.
-     */
-    @JsonIgnore @Internal private boolean immediateReconciliationNeeded = false;
 }

flink-kubernetes-operator-api/src/main/java/org/apache/flink/kubernetes/operator/api/status/ReconciliationStatus.java

@@ -20,7 +20,6 @@
 import org.apache.flink.annotation.Experimental;
 import org.apache.flink.kubernetes.operator.api.AbstractFlinkResource;
 import org.apache.flink.kubernetes.operator.api.spec.AbstractFlinkSpec;
-import org.apache.flink.kubernetes.operator.api.spec.JobState;
 import org.apache.flink.kubernetes.operator.api.utils.SpecUtils;
 import org.apache.flink.kubernetes.operator.api.utils.SpecWithMeta;
 
@@ -100,22 +99,4 @@ public boolean isLastReconciledSpecStable() {
     public boolean isBeforeFirstDeployment() {
         return lastReconciledSpec == null;
     }
-
-    /**
-     * This method is only here for backward compatibility reasons. The current version of the
-     * operator does not leave the resources in UPGRADING state during in-place scaling therefore
-     * this method will always return false.
-     *
-     * @return True if in-place scaling is in progress.
-     */
-    @JsonIgnore
-    @Deprecated
-    public boolean scalingInProgress() {
-        if (isBeforeFirstDeployment() || state != ReconciliationState.UPGRADING) {
-            return false;
-        }
-        var job = deserializeLastReconciledSpec().getJob();
-        // For regular full upgrades the jobstate is suspended in UPGRADING state
-        return job != null && job.getState() == JobState.RUNNING;
-    }
 }

flink-kubernetes-operator-api/src/main/java/org/apache/flink/kubernetes/operator/api/status/SavepointInfo.java

@@ -24,6 +24,8 @@
 import lombok.Builder;
 import lombok.Data;
 import lombok.NoArgsConstructor;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
 
 import java.util.ArrayList;
 import java.util.List;
@@ -40,6 +42,9 @@ public class SavepointInfo implements SnapshotInfo {
      * Last completed savepoint by the operator for manual and periodic snapshots. Only used if
      * FlinkStateSnapshot resources are disabled.
      */
+    private static final Logger LOG = LoggerFactory.getLogger(SavepointInfo.class);
+
+    /** Last completed savepoint by the operator. */
     private Savepoint lastSavepoint;
 
     /** Trigger id of a pending savepoint operation. */
@@ -82,7 +87,11 @@ public void resetTrigger() {
      * @param savepoint Savepoint to be added.
      */
     public void updateLastSavepoint(Savepoint savepoint) {
-        if (lastSavepoint == null || !lastSavepoint.getLocation().equals(savepoint.getLocation())) {
+        if (savepoint == null) {
+            lastSavepoint = null;
+        } else if (lastSavepoint == null
+                || !lastSavepoint.getLocation().equals(savepoint.getLocation())) {
+            LOG.debug("Updating last savepoint to {}", savepoint);
             lastSavepoint = savepoint;
             savepointHistory.add(savepoint);
             if (savepoint.getTriggerType() == SnapshotTriggerType.PERIODIC) {