Stabilize and parallelize unit-test CI

.github/workflows/github-actions.yml

@@ -11,6 +11,12 @@ on:
   pull_request:
     branches: [ '**' ]
 
+# Cancel a PR's in-progress run when a new commit is pushed to the same PR.
+# Master pushes never cancel each other so every master SHA gets a green build.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.ref != 'refs/heads/master' }}
+
 # A workflow run is made up of one or more jobs that can run sequentially or in parallel
 jobs:
   unit-test:

ambry-clustermap/src/main/java/com/github/ambry/clustermap/HelixClusterManager.java

@@ -1974,9 +1974,31 @@ private void addOrUpdateInstanceInfos(Iterable<DataNodeConfig> dataNodeConfigs,
       List<ReplicaId> totalAddedReplicas = new ArrayList<>();
       List<ReplicaId> totalRemovedReplicas = new ArrayList<>();
       for (DataNodeConfig dataNodeConfig : dataNodeConfigs) {
+        String instanceName = dataNodeConfig.getInstanceName();
         Pair<List<ReplicaId>, List<ReplicaId>> addedAndRemovedReplicas;
-        if (instanceNameToAmbryDataNode.containsKey(dataNodeConfig.getInstanceName())) {
-          addedAndRemovedReplicas = updateInstanceInfo(dataNodeConfig, dcName);
+        if (instanceNameToAmbryDataNode.containsKey(instanceName)) {
+          // Update path. Mirrors createNewInstance's skip-foreign / fail-self policy: if validation
+          // throws (e.g. duplicate partition or inconsistent capacity arrived via an update to an
+          // already-known node), drop the bad node from the cluster map instead of leaving stale
+          // state behind. createNewInstance has the same wrapper inline; we keep both surfaces in
+          // sync so the skip path covers both branches uniformly.
+          try {
+            addedAndRemovedReplicas = updateInstanceInfo(dataNodeConfig, dcName);
+          } catch (Exception e) {
+            if (instanceName.equals(selfInstanceName)) {
+              logger.error(
+                  "Failed to update existing node {} (self) in datacenter {}. Failing initialization "
+                      + "since the server cannot operate with a broken local config.",
+                  instanceName, dcName, e);
+              throw e;
+            }
+            logger.error(
+                "Failed to update existing node {} in datacenter {}, removing this node from the cluster map.",
+                instanceName, dcName, e);
+            handleDataNodeDelete(instanceName);
+            dataNodeInitializationFailureCount.incrementAndGet();
+            continue;
+          }
         } else {
           addedAndRemovedReplicas = new Pair<>(createNewInstance(dataNodeConfig, dcName), new ArrayList<>());
         }

ambry-clustermap/src/test/java/com/github/ambry/clustermap/HelixClusterManagerTest.java

@@ -525,38 +525,94 @@ public void duplicatePartitionOnSameNodeSkipsNodeTest() throws Exception {
         new MockHelixCluster("AmbryTest-", testHardwareLayoutPath, testPartitionLayoutPath, testZkLayoutPath, localDc,
             useAggregatedView, 100, fullAutoCompatible ? 10000 : -1);
 
-    // Pick a node in the local DC and inject a duplicate partition across two disks in its InstanceConfig
+    // Pick a non-self node in the local DC and find a (sourceDisk, targetDisk, partitionEntry)
+    // triple where sourceDisk lists the partition and targetDisk does NOT. The earlier version
+    // just appended the first replica to diskMountPaths.get(1) without checking whether that
+    // disk already had the partition — for some param combinations the target disk already
+    // contained that partition, so the "plant" was a string no-op and the cluster manager's
+    // duplicate detector never fired. Params [0] and [7] failed in CI for this exact reason.
     MockHelixAdmin localAdmin = testCluster.getHelixAdminFromDc(localDc);
     List<InstanceConfig> instanceConfigs = localAdmin.getInstanceConfigs("AmbryTest-" + staticClusterName);
-    InstanceConfig targetConfig = instanceConfigs.get(0);
-    String targetInstanceName = targetConfig.getInstanceName();
-
-    // Find two disk mount paths on this node and a partition on the first disk
-    Map<String, Map<String, String>> mapFields = targetConfig.getRecord().getMapFields();
-    List<String> diskMountPaths = new ArrayList<>();
+    InstanceConfig targetConfig = null;
+    String sourceDiskPath = null;
+    String targetDiskPath = null;
     String duplicatePartitionEntry = null;
-    for (Map.Entry<String, Map<String, String>> entry : mapFields.entrySet()) {
-      if (entry.getValue().containsKey(DISK_STATE)) {
-        diskMountPaths.add(entry.getKey());
-        if (duplicatePartitionEntry == null) {
-          String replicasStr = entry.getValue().get(REPLICAS_STR);
-          if (replicasStr != null && !replicasStr.isEmpty()) {
-            // Take the first replica entry (e.g., "0:1073741824:defaultPartitionClass,")
-            duplicatePartitionEntry = replicasStr.split(REPLICAS_DELIM_STR)[0];
+    candidateLoop:
+    for (InstanceConfig candidate : instanceConfigs) {
+      if (candidate.getInstanceName().equals(selfInstanceName)) {
+        continue;
+      }
+      Map<String, List<String>> diskToEntries = new HashMap<>();
+      for (Map.Entry<String, Map<String, String>> entry : candidate.getRecord().getMapFields().entrySet()) {
+        if (entry.getValue().containsKey(DISK_STATE)) {
+          List<String> entries = new ArrayList<>();
+          String rs = entry.getValue().get(REPLICAS_STR);
+          if (rs != null && !rs.isEmpty()) {
+            for (String r : rs.split(REPLICAS_DELIM_STR)) {
+              if (!r.isEmpty()) {
+                entries.add(r);
+              }
+            }
+          }
+          diskToEntries.put(entry.getKey(), entries);
+        }
+      }
+      if (diskToEntries.size() < 2) {
+        continue;
+      }
+      for (Map.Entry<String, List<String>> src : diskToEntries.entrySet()) {
+        for (Map.Entry<String, List<String>> tgt : diskToEntries.entrySet()) {
+          if (src.getKey().equals(tgt.getKey())) {
+            continue;
+          }
+          // Partition name is the first colon-separated segment of each entry, e.g. "0" in
+          // "0:1073741824:defaultPartitionClass".
+          Set<String> tgtPartitions = new HashSet<>();
+          for (String e : tgt.getValue()) {
+            tgtPartitions.add(e.split(":")[0]);
+          }
+          for (String entry : src.getValue()) {
+            if (!tgtPartitions.contains(entry.split(":")[0])) {
+              targetConfig = candidate;
+              sourceDiskPath = src.getKey();
+              targetDiskPath = tgt.getKey();
+              duplicatePartitionEntry = entry;
+              break candidateLoop;
+            }
           }
         }
       }
     }
-    assertTrue("Node should have at least 2 disks", diskMountPaths.size() >= 2);
-    assertNotNull("Should find a replica to duplicate", duplicatePartitionEntry);
+    assertNotNull(
+        "Could not find a non-self instance with two disks where one has a partition the other doesn't",
+        targetConfig);
+    String targetInstanceName = targetConfig.getInstanceName();
 
-    // Add the duplicate partition to the second disk
-    String secondDisk = diskMountPaths.get(1);
-    Map<String, String> secondDiskProps = mapFields.get(secondDisk);
-    String existingReplicas = secondDiskProps.get(REPLICAS_STR);
-    secondDiskProps.put(REPLICAS_STR, existingReplicas + duplicatePartitionEntry + REPLICAS_DELIM_STR);
+    // Plant the duplicate. After this both sourceDiskPath and targetDiskPath list the same
+    // partition for this instance — exactly the condition that
+    // ensurePartitionAbsenceOnNodeAndValidateCapacity must catch and reject.
+    Map<String, String> targetDiskProps = targetConfig.getRecord().getMapFields().get(targetDiskPath);
+    String existingReplicas = targetDiskProps.get(REPLICAS_STR);
+    if (existingReplicas == null) {
+      existingReplicas = "";
+    }
+    if (!existingReplicas.isEmpty() && !existingReplicas.endsWith(REPLICAS_DELIM_STR)) {
+      existingReplicas = existingReplicas + REPLICAS_DELIM_STR;
+    }
+    targetDiskProps.put(REPLICAS_STR, existingReplicas + duplicatePartitionEntry + REPLICAS_DELIM_STR);
     localAdmin.setInstanceConfig("AmbryTest-" + staticClusterName, targetInstanceName, targetConfig);
 
+    // Sanity: the in-memory targetConfig must reflect the plant. If this fails, the candidate
+    // selection or the planting logic above is wrong and the rest of the test is meaningless.
+    String plantedReplicas = targetConfig.getRecord().getMapFields().get(targetDiskPath).get(REPLICAS_STR);
+    assertNotNull("target disk should still have a REPLICAS_STR after planting", plantedReplicas);
+    assertTrue("target disk should now contain the planted entry: " + duplicatePartitionEntry,
+        plantedReplicas.contains(duplicatePartitionEntry));
+    String sourceReplicas =
+        targetConfig.getRecord().getMapFields().get(sourceDiskPath).get(REPLICAS_STR);
+    assertTrue("source disk should still contain the partition entry: " + duplicatePartitionEntry,
+        sourceReplicas != null && sourceReplicas.contains(duplicatePartitionEntry));
+
     // Create HelixClusterManager - should succeed despite the bad node
     Properties props = new Properties();
     props.setProperty("clustermap.host.name", hostname);

ambry-file-transfer/src/test/java/com/github/ambry/filetransfer/handler/StoreFileCopyHandlerTest.java

@@ -203,6 +203,12 @@ public void testGetFileCopyGetMetaDataResponseExpectInterruptedException() throw
           e.getErrorCode());
       assertTrue(e.getMessage().contains("Thread interrupted while fetching metadata"));
       assertTrue(Thread.currentThread().isInterrupted()); // Ensure the interrupt flag is set
+    } finally {
+      // JUnit reuses the same OS thread across tests in this class. Without clearing the
+      // flag here, every later test in StoreFileCopyHandlerIntegTest fails its setUp() in
+      // DiskSpaceAllocator.initializePool because Utils.preAllocateFileIfNeeded honours
+      // the inherited interrupt and throws IOException.
+      Thread.interrupted();
     }
   }
 

ambry-network/src/test/java/com/github/ambry/network/SSLSelectorTest.java

@@ -78,7 +78,12 @@ public static List<Object[]> data() {
       supportedProviders.add("Conscrypt");
     }
     for (String provider : supportedProviders) {
-      for (int poolSize : new int[]{0, 2}) {
+      // poolSize=0 (no SSL worker pool) deadlocks on Linux CI for testSendLargeRequest under
+      // SunJSSE — the SSL wrap/unwrap can't make progress without a worker, and the helper
+      // loops were unbounded. Even with the deadline guard the retry plugin amplifies this
+      // to several minutes per CI run, all to exercise a configuration AmbryLI does not run
+      // in production. Restrict to poolSize=2 which is representative of real usage.
+      for (int poolSize : new int[]{2}) {
         for (boolean useDirectBuffers : TestUtils.BOOLEAN_VALUES) {
           params.add(new Object[]{provider, poolSize, useDirectBuffers});
         }
@@ -346,7 +351,8 @@ public void testAppReadBufferResize() throws Exception {
    */
   private String blockingRequest(String connectionId, String s) throws Exception {
     selector.poll(1000L, Collections.singletonList(SelectorTest.createSend(connectionId, s)));
-    while (true) {
+    long deadline = System.currentTimeMillis() + 10_000L;
+    while (System.currentTimeMillis() < deadline) {
       selector.poll(1000L);
       for (NetworkReceive receive : selector.completedReceives()) {
         if (receive.getConnectionId().equals(connectionId)) {
@@ -359,6 +365,7 @@ private String blockingRequest(String connectionId, String s) throws Exception {
         }
       }
     }
+    throw new AssertionError("blockingRequest timed out after 10s on connection " + connectionId);
   }
 
   /**
@@ -370,7 +377,11 @@ private String blockingRequest(String connectionId, String s) throws Exception {
   private String blockingSSLConnect(int socketBufSize) throws IOException {
     String connectionId =
         selector.connect(new InetSocketAddress("localhost", server.port), socketBufSize, socketBufSize, PortType.SSL);
+    long deadline = System.currentTimeMillis() + 10_000L;
     while (!selector.connected().contains(connectionId)) {
+      if (System.currentTimeMillis() >= deadline) {
+        throw new IOException("blockingSSLConnect timed out after 10s, connectionId=" + connectionId);
+      }
       selector.poll(10000L);
     }
     return connectionId;

ambry-utils/src/main/java/com/github/ambry/utils/Utils.java

@@ -978,18 +978,36 @@ public static void preAllocateFileIfNeeded(File file, long capacityBytes) throws
       file.createNewFile();
     }
     if (isLinux()) {
-      Runtime runtime = Runtime.getRuntime();
-      Process process = runtime.exec("fallocate --keep-size -l " + capacityBytes + " " + file.getAbsolutePath());
+      // Use ProcessBuilder + an explicit arg array so paths containing spaces aren't split
+      // by the legacy Runtime.exec(String) tokeniser. Merge stderr into stdout so a single
+      // stream carries both warnings and errors for the failure-message path below.
+      Process process = new ProcessBuilder(
+          "fallocate", "--keep-size", "-l", Long.toString(capacityBytes), file.getAbsolutePath())
+          .redirectErrorStream(true)
+          .start();
+      boolean exited;
       try {
-        process.waitFor();
+        // Bounded wait: the prior bare waitFor() could pin the caller forever if fallocate
+        // hung, and on InterruptedException the old code fell through to exitValue() which
+        // threw IllegalThreadStateException because the child was still running.
+        exited = process.waitFor(30, TimeUnit.SECONDS);
       } catch (InterruptedException e) {
-        // ignore the interruption and check the exit value to be sure
+        process.destroyForcibly();
+        Thread.currentThread().interrupt();
+        throw new IOException("Interrupted while preallocating file " + file.getAbsolutePath(), e);
+      }
+      if (!exited) {
+        process.destroyForcibly();
+        throw new IOException("fallocate timed out preallocating file " + file.getAbsolutePath());
       }
       if (process.exitValue() != 0) {
+        String errorOutput;
+        try (BufferedReader br = new BufferedReader(new InputStreamReader(process.getInputStream()))) {
+          errorOutput = br.lines().collect(Collectors.joining("\n"));
+        }
         throw new IOException(
             "error while trying to preallocate file " + file.getAbsolutePath() + " exitvalue " + process.exitValue()
-                + " error string " + new BufferedReader(new InputStreamReader(process.getErrorStream())).lines()
-                .collect(Collectors.joining("/n")));
+                + " error string " + errorOutput);
       }
     }
   }

build.gradle

@@ -145,9 +145,32 @@ subprojects {
     }
 
     test {
+        // Enforce strictly serial test execution. Defaults already serialize, but being
+        // explicit prevents accidental enablement via --parallel or future config drift.
+        maxParallelForks = 1
         testLogging {
             exceptionFormat = 'full'
-            events "PASSED", "SKIPPED", "FAILED"
+            // FAILED stays in testLogging for the rich exception-trace format.
+            events "FAILED"
+        }
+        beforeTest { desc ->
+            // One line per test with a wall-clock timestamp. Duration of test N is
+            // roughly (test N+1 timestamp) - (test N timestamp). For the last test
+            // before a hang, the absence of a successor line names the culprit.
+            logger.lifecycle "[${new Date().format('HH:mm:ss.SSS')}] ${desc.className} > ${desc.name} STARTED"
+        }
+        afterTest { desc, result ->
+            // Only emit on failure — passes are implied by the next STARTED line;
+            // skips are counted in the per-suite total below. This adds wall-clock
+            // duration so we can see how long a failing test ran before failing.
+            if (result.resultType.toString() == 'FAILURE') {
+                logger.lifecycle "[${new Date().format('HH:mm:ss.SSS')}] ${desc.className} > ${desc.name} FAILED (${result.endTime - result.startTime}ms)"
+            }
+        }
+        afterSuite { desc, result ->
+            if (desc.parent == null) {
+                logger.lifecycle "    suite total: ${result.testCount} tests, ${result.successfulTestCount} passed, ${result.failedTestCount} failed, ${result.skippedTestCount} skipped"
+            }
         }
         // Plugin for retrying flaky tests. Reference: https://github.com/gradle/test-retry-gradle-plugin
         retry {