Solve the problem that when changing the cluster configuration through the joint consensus algorithm, a new leader cannot be elected under abnormal circumstances.

src/braft/node.cpp

@@ -70,6 +70,9 @@ DECLARE_bool(raft_enable_leader_lease);
 DEFINE_bool(raft_enable_witness_to_leader, false, 
             "enable witness temporarily to become leader when leader down accidently");
 
+DEFINE_bool(raft_enable_peer_not_in_conf_can_elec, false, 
+            "enable peer not in the conf can initiate elections");
+
 #ifndef UNIT_TEST
 static bvar::Adder<int64_t> g_num_nodes("raft_node_count");
 #else
@@ -1622,10 +1625,12 @@ void NodeImpl::pre_vote(std::unique_lock<raft_mutex_t>* lck, bool triggered) {
                         " configuration is possibly out of date";
         return;
     }
-    if (!_conf.contains(_server_id)) {
+    if (!FLAGS_raft_enable_peer_not_in_conf_can_elec) {
+      if (!_conf.contains(_server_id)) {
         LOG(WARNING) << "node " << _group_id << ':' << _server_id
                      << " can't do pre_vote as it is not in " << _conf.conf;
         return;
+      }
     }
 
     int64_t old_term = _current_term;
@@ -1681,10 +1686,12 @@ void NodeImpl::elect_self(std::unique_lock<raft_mutex_t>* lck,
                           bool old_leader_stepped_down) {
     LOG(INFO) << "node " << _group_id << ":" << _server_id
               << " term " << _current_term << " start vote and grant vote self";
-    if (!_conf.contains(_server_id)) {
+    if (!FLAGS_raft_enable_peer_not_in_conf_can_elec) {
+      if (!_conf.contains(_server_id)) {
         LOG(WARNING) << "node " << _group_id << ':' << _server_id
                      << " can't do elect_self as it is not in " << _conf.conf;
         return;
+      }
     }
     // cancel follower election timer
     if (_state == STATE_FOLLOWER) {
@@ -2393,6 +2400,16 @@ void NodeImpl::handle_append_entries_request(brpc::Controller* cntl,
     brpc::ClosureGuard done_guard(done);
     std::unique_lock<raft_mutex_t> lck(_mutex);
 
+    // for test
+    const int64_t reject_log_index = get_reject_log_index();
+    if (reject_log_index > 0 &&
+        request->prev_log_index() + 1 >= reject_log_index) {
+      // _last_leader_timestamp = butil::monotonic_time_ms();
+      // don't interfere check_dead_nodes
+      cntl->SetFailed(EBUSY, "handle_append_entries_request reject_log_index");
+      return;
+    }
+
     // pre set term, to avoid get term in lock
     response->set_term(_current_term);
 
@@ -2894,6 +2911,10 @@ void NodeImpl::get_status(NodeStatus* status) {
     }
 }
 
+void NodeImpl::get_log_mgr_status(LogManagerStatus* log_manager_status) {
+    _log_manager->get_status(log_manager_status);
+}
+
 void NodeImpl::stop_replicator(const std::set<PeerId>& keep,
                                const std::set<PeerId>& drop) {
     for (std::set<PeerId>::const_iterator

src/braft/node.h

@@ -214,6 +214,8 @@ friend class VoteBallotCtx;
     // see from the website, which is generated by |describe| actually.
     void get_status(NodeStatus* status);
 
+    void get_log_mgr_status(LogManagerStatus* log_manager_status);
+
     // Readonly mode func
     void enter_readonly_mode();
     void leave_readonly_mode();
@@ -241,6 +243,10 @@ friend class VoteBallotCtx;
 
     bool disable_cli() const { return _options.disable_cli; }
     bool is_witness() const { return _options.witness; }
+
+    // for test
+    void set_reject_log_index(const int64_t log_index) { reject_log_index_ = log_index; }
+    int64_t get_reject_log_index() const { return reject_log_index_; }
 private:
 friend class butil::RefCountedThreadSafe<NodeImpl>;
 
@@ -533,6 +539,9 @@ friend class butil::RefCountedThreadSafe<NodeImpl>;
 
     LeaderLease _leader_lease;
     FollowerLease _follower_lease;
+
+    // for test
+    int64_t reject_log_index_ {0};
 };
 
 }

src/braft/raft.cpp

@@ -230,6 +230,10 @@ void Node::get_status(NodeStatus* status) {
     return _impl->get_status(status);
 }
 
+void Node::get_log_mgr_status(LogManagerStatus* log_manager_status) {
+  return _impl->get_log_mgr_status(log_manager_status);
+}
+
 void Node::enter_readonly_mode() {
     return _impl->enter_readonly_mode();
 }
@@ -242,6 +246,14 @@ bool Node::readonly() {
     return _impl->readonly();
 }
 
+void Node::set_reject_log_index(const int64_t log_index) {
+    _impl->set_reject_log_index(log_index);
+}
+
+int64_t Node::get_reject_log_index() const {
+    return _impl->get_reject_log_index();
+}
+
 // ------------- Iterator
 void Iterator::next() {
     if (valid()) {

src/braft/raft.h

@@ -45,6 +45,7 @@ class LeaderChangeContext;
 class FileSystemAdaptor;
 class SnapshotThrottle;
 class LogStorage;
+class LogManagerStatus;
 
 const PeerId ANY_PEER(butil::EndPoint(butil::IP_ANY, 0), 0);
 
@@ -762,6 +763,8 @@ class Node {
     // see from the website.
     void get_status(NodeStatus* status);
 
+    void get_log_mgr_status(LogManagerStatus* log_manager_status);
+
     // Make this node enter readonly mode.
     // Readonly mode should only be used to protect the system in some extreme cases.
     // For example, in a storage system, too many write requests flood into the system
@@ -789,6 +792,10 @@ class Node {
     //        is less than the majority.
     bool readonly();
 
+    void set_reject_log_index(const int64_t log_index);
+
+    int64_t get_reject_log_index() const;
+
 private:
     NodeImpl* _impl;
 };