HBASE-28850 Only return from ReplicationSink.replicationEntries while all background tasks are finished (#6263)

Signed-off-by: Andrew Purtell <apurtell@apache.org>
(cherry picked from commit 52082bc)

Author: Apache9

hbase-server/src/main/java/org/apache/hadoop/hbase/replication/regionserver/ReplicationSink.java

@@ -509,17 +509,33 @@ private void batch(TableName tableName, Collection<List<Row>> allRows, int batch
       }
       futures.addAll(batchRows.stream().map(table::batchAll).collect(Collectors.toList()));
     }
-
+    // Here we will always wait until all futures are finished, even if there are failures when
+    // getting from a future in the middle. This is because this method may be called in a rpc call,
+    // so the batch operations may reference some off heap cells(through CellScanner). If we return
+    // earlier here, the rpc call may be finished and they will release the off heap cells before
+    // some of the batch operations finish, and then cause corrupt data or even crash the region
+    // server. See HBASE-28584 and HBASE-28850 for more details.
+    IOException error = null;
     for (Future<?> future : futures) {
       try {
         FutureUtils.get(future);
       } catch (RetriesExhaustedException e) {
+        IOException ioe;
         if (e.getCause() instanceof TableNotFoundException) {
-          throw new TableNotFoundException("'" + tableName + "'");
+          ioe = new TableNotFoundException("'" + tableName + "'");
+        } else {
+          ioe = e;
+        }
+        if (error == null) {
+          error = ioe;
+        } else {
+          error.addSuppressed(ioe);
         }
-        throw e;
       }
     }
+    if (error != null) {
+      throw error;
+    }
   }
 
   /**