Merge branch 'master' into andyz/fixFailureOnSumittedNode

Netflix · Nov 14, 2024 · 9f50786 · 9f50786
2 parents aa15cdf + 4da2e05
commit 9f50786
Show file tree

Hide file tree

Showing 7 changed files with 51 additions and 7 deletions.
diff --git a/...mantis-control-plane-server/src/main/java/io/mantisrx/master/jobcluster/job/JobActor.java b/...mantis-control-plane-server/src/main/java/io/mantisrx/master/jobcluster/job/JobActor.java
@@ -1906,13 +1906,29 @@ public void checkHeartBeats(Instant currentTime) {
                 for (JobWorker worker : stage.getAllWorkers()) {
                     IMantisWorkerMetadata workerMeta = worker.getMetadata();
                     if (!workerMeta.getLastHeartbeatAt().isPresent()) {
-                        // the worker is still waiting for resource allocation and the scheduler should take care of
-                        // the retry logic.
                         Instant acceptedAt = Instant.ofEpochMilli(workerMeta.getAcceptedAt());
-                        LOGGER.warn("Job {}, Worker {} stuck in accepted state since {}",
-                            this.jobMgr.getJobId(),
-                            workerMeta.getWorkerId(),
-                            acceptedAt);
+                        if(!scheduler.schedulerHandlesAllocationRetries()) {
+                            // worker stuck in accepted and the scheduler will not retry allocation requests, so
+                            // we must resubmit
+                            if (Duration.between(acceptedAt, currentTime).getSeconds() > stuckInSubmitToleranceSecs) {
+                                LOGGER.info("Resubmitting Job {}, Worker {} that has been stuck in accepted state for {}", this.jobMgr.getJobId(),
+                                    workerMeta.getWorkerId(), Duration.between(acceptedAt, currentTime).getSeconds());
+                                workersToResubmit.add(worker);
+                                eventPublisher.publishStatusEvent(new LifecycleEventsProto.WorkerStatusEvent(
+                                    WARN,
+                                    "worker stuck in Accepted state, resubmitting worker",
+                                    workerMeta.getStageNum(),
+                                    workerMeta.getWorkerId(),
+                                    workerMeta.getState()));
+                            }
+                        } else {
+                            // the worker is still waiting for resource allocation and the scheduler should take care of
+                            // the retry logic.
+                            LOGGER.warn("Job {}, Worker {} stuck in accepted state since {}",
+                                this.jobMgr.getJobId(),
+                                workerMeta.getWorkerId(),
+                                acceptedAt);
+                        }
                     } else {
                         if (Duration.between(workerMeta.getLastHeartbeatAt().get(), currentTime).getSeconds()
                             > missedHeartBeatToleranceSecs) {

diff --git a/...trol-plane-server/src/main/java/io/mantisrx/server/master/config/MasterConfiguration.java b/...trol-plane-server/src/main/java/io/mantisrx/server/master/config/MasterConfiguration.java
@@ -363,6 +363,10 @@ default Duration getSchedulerIntervalBetweenRetries() {
     @Default("3")
     int getSlaMaxHeadroomForAccepted();
 
+    @Config("mantis.scheduler.handlesAllocationRetries")
+    @Default("true")
+    boolean getSchedulerHandlesAllocationRetries();
+
     default Duration getHeartbeatInterval() {
         return Duration.ofMillis(getHeartbeatIntervalInMs());
     }

diff --git a/...ntrol-plane-server/src/main/java/io/mantisrx/server/master/scheduler/MantisScheduler.java b/...ntrol-plane-server/src/main/java/io/mantisrx/server/master/scheduler/MantisScheduler.java
@@ -76,4 +76,11 @@ public interface MantisScheduler {
      */
     void initializeRunningWorker(final ScheduleRequest scheduleRequest, final String hostname, final String hostID);
 
+    /**
+     * This should return true if the underlying scheduler handles retrying worker allocations.
+     *
+     * @return If there are not enough resources to schedule the worker and the scheduler automatically retries until
+     * the worker is assigned, then return true; otherwise, return false.
+     */
+    boolean schedulerHandlesAllocationRetries();
 }
diff --git a/...-server/src/main/java/io/mantisrx/server/master/scheduler/MantisSchedulerFactoryImpl.java b/...-server/src/main/java/io/mantisrx/server/master/scheduler/MantisSchedulerFactoryImpl.java
@@ -63,7 +63,8 @@ public MantisScheduler forClusterID(ClusterID clusterID) {
                                 executeStageRequestFactory,
                                 jobMessageRouter,
                                 metricsRegistry),
-                            "scheduler-for-" + cid.getResourceID()));
+                            "scheduler-for-" + cid.getResourceID()),
+                            masterConfiguration.getSchedulerHandlesAllocationRetries());
                     });
         } else {
             log.error("Scheduler gets unexpected null clusterID");

diff --git a/...rver/src/main/java/io/mantisrx/server/master/scheduler/ResourceClusterAwareScheduler.java b/...rver/src/main/java/io/mantisrx/server/master/scheduler/ResourceClusterAwareScheduler.java
@@ -32,6 +32,7 @@
 public class ResourceClusterAwareScheduler implements MantisScheduler {
 
     private final ActorRef schedulerActor;
+    private final boolean handlesAllocationRetries;
 
     @Override
     public void scheduleWorkers(BatchScheduleRequest scheduleRequest) {
@@ -67,4 +68,9 @@ public void initializeRunningWorker(ScheduleRequest scheduleRequest, String host
             new InitializeRunningWorkerRequestEvent(scheduleRequest, TaskExecutorID.of(hostID)),
             null);
     }
+
+    @Override
+    public boolean schedulerHandlesAllocationRetries() {
+        return handlesAllocationRetries;
+    }
 }
diff --git a/...l-plane-server/src/test/java/io/mantisrx/master/jobcluster/job/JobTestMigrationTests.java b/...l-plane-server/src/test/java/io/mantisrx/master/jobcluster/job/JobTestMigrationTests.java
@@ -209,6 +209,11 @@ public void initializeRunningWorker(ScheduleRequest scheduleRequest, String host
             // TODO Auto-generated method stub
 
         }
+
+        @Override
+        public boolean schedulerHandlesAllocationRetries(){
+            return false;
+        }
     }
 
     public static void main(String[] args) {

diff --git a/...-control-plane-server/src/test/java/io/mantisrx/master/scheduler/FakeMantisScheduler.java b/...-control-plane-server/src/test/java/io/mantisrx/master/scheduler/FakeMantisScheduler.java
@@ -64,4 +64,9 @@ public void updateWorkerSchedulingReadyTime(final WorkerId workerId, final long
     public void initializeRunningWorker(final ScheduleRequest scheduleRequest, final String hostname, final String hostID) {
         // no-op
     }
+
+    @Override
+    public boolean schedulerHandlesAllocationRetries() {
+        return false;
+    }
 }