raftstore: only persist merge target if the merge is known to be succeeded (#12251)

components/raftstore/src/store/peer.rs

@@ -960,7 +960,13 @@ where
             &mut kv_wb,
             &region,
             PeerState::Tombstone,
-            self.pending_merge_state.clone(),
+            // Only persist the `merge_state` if the merge is known to be succeeded
+            // which is determined by the `keep_data` flag
+            if keep_data {
+                self.pending_merge_state.clone()
+            } else {
+                None
+            },
         )?;
         // write kv rocksdb first in case of restart happen between two write
         let mut write_opts = WriteOptions::new();

components/test_raftstore/src/cluster.rs

@@ -1120,6 +1120,28 @@ impl<T: Simulator> Cluster<T> {
             .unwrap()
     }
 
+    pub fn must_peer_state(&self, region_id: u64, store_id: u64, peer_state: PeerState) {
+        for _ in 0..100 {
+            let state = self
+                .get_engine(store_id)
+                .c()
+                .get_msg_cf::<RegionLocalState>(
+                    engine_traits::CF_RAFT,
+                    &keys::region_state_key(region_id),
+                )
+                .unwrap()
+                .unwrap();
+            if state.get_state() == peer_state {
+                return;
+            }
+            sleep_ms(10);
+        }
+        panic!(
+            "[region {}] peer state still not reach {:?}",
+            region_id, peer_state
+        );
+    }
+
     pub fn wait_last_index(
         &mut self,
         region_id: u64,

tests/failpoints/cases/test_merge.rs

@@ -1302,3 +1302,81 @@ fn test_merge_election_and_restart() {
     // If logs from different term are process correctly, store 2 should have latest updates.
     must_get_equal(&cluster.get_engine(2), b"k14", b"v14");
 }
+
+// Testing that when the source peer is destroyed while merging, it should not persist the `merge_state`
+// thus won't generate gc message to destroy other peers
+#[test]
+fn test_destroy_source_peer_while_merging() {
+    let mut cluster = new_node_cluster(0, 5);
+    configure_for_merge(&mut cluster);
+    let pd_client = Arc::clone(&cluster.pd_client);
+    pd_client.disable_default_operator();
+
+    cluster.run();
+
+    cluster.must_put(b"k1", b"v1");
+    cluster.must_put(b"k3", b"v3");
+    for i in 1..=5 {
+        must_get_equal(&cluster.get_engine(i), b"k1", b"v1");
+        must_get_equal(&cluster.get_engine(i), b"k3", b"v3");
+    }
+
+    cluster.must_split(&pd_client.get_region(b"k1").unwrap(), b"k2");
+    let left = pd_client.get_region(b"k1").unwrap();
+    let right = pd_client.get_region(b"k3").unwrap();
+    cluster.must_transfer_leader(right.get_id(), new_peer(1, 1));
+
+    let schedule_merge_fp = "on_schedule_merge";
+    fail::cfg(schedule_merge_fp, "return()").unwrap();
+
+    // Start merge and wait until peer 5 apply prepare merge
+    cluster.must_try_merge(right.get_id(), left.get_id());
+    cluster.must_peer_state(right.get_id(), 5, PeerState::Merging);
+
+    // filter heartbeat and append message for peer 5
+    cluster.add_send_filter(CloneFilterFactory(
+        RegionPacketFilter::new(right.get_id(), 5)
+            .direction(Direction::Recv)
+            .msg_type(MessageType::MsgHeartbeat)
+            .msg_type(MessageType::MsgAppend),
+    ));
+
+    // remove peer from target region to trigger merge rollback.
+    pd_client.must_remove_peer(left.get_id(), find_peer(&left, 2).unwrap().clone());
+    must_get_none(&cluster.get_engine(2), b"k1");
+
+    // Merge must rollbacked if we can put more data to the source region
+    fail::remove(schedule_merge_fp);
+    cluster.must_put(b"k4", b"v4");
+    for i in 1..=4 {
+        must_get_equal(&cluster.get_engine(i), b"k4", b"v4");
+    }
+
+    // remove peer 5 from peer list so it will destroy itself by tombstone message
+    // and should not persist the `merge_state`
+    pd_client.must_remove_peer(right.get_id(), new_peer(5, 5));
+    must_get_none(&cluster.get_engine(5), b"k3");
+
+    // so that other peers will send message to store 5
+    pd_client.must_add_peer(right.get_id(), new_peer(5, 6));
+    // but it is still in tombstone state due to the message filter
+    let state = cluster.region_local_state(right.get_id(), 5);
+    assert_eq!(state.get_state(), PeerState::Tombstone);
+
+    // let the peer on store 4 have a larger peer id
+    pd_client.must_remove_peer(right.get_id(), new_peer(4, 4));
+    pd_client.must_add_peer(right.get_id(), new_peer(4, 7));
+    must_get_equal(&cluster.get_engine(4), b"k4", b"v4");
+
+    // if store 5 have persist the merge state, peer 2 and peer 3 will be destroyed because
+    // store 5 will response their request vote message with a gc message, and peer 7 will cause
+    // store 5 panic because peer 7 have larger peer id than the peer in the merge state
+    cluster.clear_send_filters();
+    cluster.add_send_filter(IsolationFilterFactory::new(1));
+
+    cluster.must_put(b"k5", b"v5");
+    assert!(!state.has_merge_state(), "{:?}", state);
+    for i in 2..=5 {
+        must_get_equal(&cluster.get_engine(i), b"k5", b"v5");
+    }
+}