Merge pull request #24388 from vespa-engine/jonmv/abandon-always-failing-apps

Jonmv/abandon always failing apps

4 files changed, 74 insertions, 5 deletions

diff --git a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTrigger.java b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTrigger.java
index cf09afa7181..d5a31a07408 100644
--- a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTrigger.java
+++ b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTrigger.java
@@ -91,10 +91,31 @@ public class DeploymentTrigger {
                                                                                status,
                                                                                false));
             }
+
+            // If app has been broken since it was first submitted, and not fixed for a long time, we stop managing it until a new submission comes in.
+            if (applicationWasAlwaysBroken(status))
+                application = application.withProjectId(OptionalLong.empty());
+
             applications().store(application);
         });
     }
 
+    private boolean applicationWasAlwaysBroken(DeploymentStatus status) {
+        // If application has a production deployment, we cannot forget it.
+        if (status.application().instances().values().stream().anyMatch(instance -> ! instance.productionDeployments().isEmpty()))
+            return false;
+
+        // Then, we need a job that always failed, and failed on the last revision for at least 30 days.
+        RevisionId last = status.application().revisions().last().get().id();
+        Instant threshold = clock.instant().minus(Duration.ofDays(30));
+        for (JobStatus job : status.jobs().asList())
+            for (Run run : job.runs().descendingMap().values())
+                if (run.hasEnded() && ! run.hasFailed() || ! run.versions().targetRevision().equals(last)) break;
+                else if (run.start().isBefore(threshold)) return true;
+
+        return false;
+    }
+
     /**
      * Records information when a job completes (successfully or not). This information is used when deciding what to
      * trigger next.
@@ -339,8 +360,8 @@ public class DeploymentTrigger {
     /** Returns the set of all jobs which have changes to propagate from the upstream steps. */
     private List<Job> computeReadyJobs() {
         return jobs.deploymentStatuses(ApplicationList.from(applications().readable())
-                                                      .withProjectId() // Need to keep this, as we have applications with deployment spec that shouldn't be orchestrated. // Maybe not any longer?
-                                                      .withDeploymentSpec())
+                                                      .withProjectId() // Need to keep this, as we have applications with deployment spec that shouldn't be orchestrated.
+                                                      .withJobs())
                    .withChanges()
                    .asList().stream()
                    .filter(status -> ! hasExceededQuota(status.application().id().tenant()))
diff --git a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/OutstandingChangeDeployer.java b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/OutstandingChangeDeployer.java
index b33a43a2031..37b06fea066 100644
--- a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/OutstandingChangeDeployer.java
+++ b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/OutstandingChangeDeployer.java
@@ -26,9 +26,8 @@ public class OutstandingChangeDeployer extends ControllerMaintainer {
     protected double maintain() {
         double ok = 0, total = 0;
         for (Application application : ApplicationList.from(controller().applications().readable())
-                                                      .withProductionDeployment()
                                                       .withProjectId()
-                                                      .withDeploymentSpec()
+                                                      .withJobs()
                                                       .asList())
             try {
                 ++total;
diff --git a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/Upgrader.java b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/Upgrader.java
index 037dacfcac9..d49cb244e47 100644
--- a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/Upgrader.java
+++ b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/Upgrader.java
@@ -70,7 +70,8 @@ public class Upgrader extends ControllerMaintainer {
 
     private DeploymentStatusList deploymentStatuses(VersionStatus versionStatus) {
         return controller().jobController().deploymentStatuses(ApplicationList.from(controller().applications().readable())
-                                                                              .withProjectId(),
+                                                                              .withProjectId()
+                                                                              .withJobs(),
                                                                versionStatus);
     }
 
diff --git a/controller-server/src/test/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTriggerTest.java b/controller-server/src/test/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTriggerTest.java
index d8cef45f124..537090c6d68 100644
--- a/controller-server/src/test/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTriggerTest.java
+++ b/controller-server/src/test/java/com/yahoo/vespa/hosted/controller/deployment/DeploymentTriggerTest.java
@@ -2700,6 +2700,54 @@ public class DeploymentTriggerTest {
     }
 
     @Test
+    void testBrokenApplication() {
+        DeploymentContext app = tester.newDeploymentContext();
+        app.submit().runJob(systemTest).failDeployment(stagingTest).failDeployment(stagingTest);
+        tester.clock().advance(Duration.ofDays(31));
+        tester.outstandingChangeDeployer().run();
+        assertEquals(OptionalLong.empty(), app.application().projectId());
+
+        app.assertNotRunning(stagingTest);
+        tester.triggerJobs();
+        app.assertNotRunning(stagingTest);
+        assertEquals(4, app.deploymentStatus().jobsToRun().size());
+
+        app.submit().runJob(systemTest).failDeployment(stagingTest);
+        tester.clock().advance(Duration.ofDays(20));
+        app.submit().runJob(systemTest).failDeployment(stagingTest);
+        tester.clock().advance(Duration.ofDays(20));
+        tester.outstandingChangeDeployer().run();
+        assertEquals(OptionalLong.of(1000), app.application().projectId());
+        tester.clock().advance(Duration.ofDays(20));
+        tester.outstandingChangeDeployer().run();
+        assertEquals(OptionalLong.empty(), app.application().projectId());
+
+        app.assertNotRunning(stagingTest);
+        tester.triggerJobs();
+        app.assertNotRunning(stagingTest);
+        assertEquals(4, app.deploymentStatus().jobsToRun().size());
+
+        app.submit().runJob(systemTest).runJob(stagingTest).failDeployment(productionUsCentral1);
+        tester.clock().advance(Duration.ofDays(31));
+        tester.outstandingChangeDeployer().run();
+        assertEquals(OptionalLong.empty(), app.application().projectId());
+
+        app.assertNotRunning(productionUsCentral1);
+        tester.triggerJobs();
+        app.assertNotRunning(productionUsCentral1);
+        assertEquals(3, app.deploymentStatus().jobsToRun().size());
+
+        app.submit().runJob(systemTest).runJob(stagingTest).timeOutConvergence(productionUsCentral1);
+        tester.clock().advance(Duration.ofDays(31));
+        tester.outstandingChangeDeployer().run();
+        assertEquals(OptionalLong.of(1000), app.application().projectId());
+
+        app.assertNotRunning(productionUsCentral1);
+        tester.triggerJobs();
+        app.assertRunning(productionUsCentral1);
+    }
+
+    @Test
     void testJobNames() {
         ZoneRegistryMock zones = new ZoneRegistryMock(SystemName.main);
         List<ZoneApi> existing = new ArrayList<>(zones.zones().all().zones());