[BugFix] fix schema change hang when data loading concurrent with sch…

…ema change (StarRocks#23456) Fixes StarRocks#23452 1. Assume there is a partition, current visible version is 13378, and all loading task is finished. 2. Trigger schema change, it will create new replica, and at the same time, one loading task is trigger and finished, and it's version is 13379. 3. Before send alter replica request, FE check that all loading task is finished, but loading task on new create replica haven't finish yet and V=1, LFV=13379. 4. Alter replica finish on BE, and BE send finish alter replica request back to FE. When call `handleFinishAlterTask`, `checkState` fail because LFV != -1. Signed-off-by: Moonm3n <[email protected]>
Moonm3n · May 23, 2023 · b1456ac · b1456ac
1 parent 3bd7081
commit b1456ac
Showing 1 changed file with 7 additions and 12 deletions.
diff --git a/fe/fe-core/src/main/java/com/starrocks/task/AlterReplicaTask.java b/fe/fe-core/src/main/java/com/starrocks/task/AlterReplicaTask.java
@@ -194,6 +194,9 @@ public TAlterTabletReqV2 toThrift() {
      *      There are new load jobs after alter task, and at least one of them is succeed on this replica.
      *      So the replica's version should be larger than X. So we don't need to modify the replica version
      *      because its already looks like normal.
+     * Case 3:
+     *      There are new load jobs after alter task, and their version and LFV is smaller or equal to X. 
+     *      And because alter request report success, it means that we can increase replica's version to X.
      */
     public void handleFinishAlterTask() throws Exception {
         Database db = GlobalStateMgr.getCurrentState().getDb(getDbId());
@@ -227,18 +230,10 @@ public void handleFinishAlterTask() throws Exception {
                         getVersion());
                 boolean versionChanged = false;
                 if (replica.getVersion() <= getVersion()) {
-                    if (replica.getLastFailedVersion() > getVersion()) {
-                        // Case 2.1
-                        replica.updateRowCount(getVersion(), replica.getDataSize(),
-                                replica.getRowCount());
-                        versionChanged = true;
-                    } else {
-                        // Case 1
-                        Preconditions.checkState(replica.getLastFailedVersion() == -1, replica.getLastFailedVersion());
-                        replica.updateRowCount(getVersion(), replica.getDataSize(),
-                                replica.getRowCount());
-                        versionChanged = true;
-                    }
+                    // Case 1, Case 2.1 or Case 3
+                    replica.updateRowCount(getVersion(), replica.getDataSize(),
+                            replica.getRowCount());
+                    versionChanged = true;
                 }
 
                 if (versionChanged) {