SDSTOR-21438: Adjust commit_quorum reduction ordering

conanfile.py

@@ -9,7 +9,7 @@
 
 class HomestoreConan(ConanFile):
     name = "homestore"
-    version = "7.5.2"
+    version = "7.5.3"
 
     homepage = "https://github.com/eBay/Homestore"
     description = "HomeStore Storage Engine"

src/lib/replication/repl_dev/raft_repl_dev.cpp

@@ -219,12 +219,40 @@ AsyncReplResult<> RaftReplDev::start_replace_member(std::string& task_id, const
             boost::uuids::to_string(member_out.id), boost::uuids::to_string(member_in.id));
 
     // Step1, validate request
+    bool is_leader = m_my_repl_id == get_leader_id();
+    // Check if leader itself is requested to move out.
+    if (m_my_repl_id == member_out.id) {
+        // immediate=false successor=-1, nuraft will choose an alive peer with highest priority as successor, and wait
+        // until the successor finishes the catch-up of the latest log, and then resign. Return NOT_LEADER and let
+        // client retry.
+        if (is_leader) {
+            RD_LOGI(trace_id, "Step1. Replace member, leader is the member_out so yield leadership, task_id={}", task_id);
+            raft_server()->yield_leadership(false /* immediate */, -1 /* successor */);
+        }
+        RD_LOGE(trace_id, "Step1. Replace member, I am not leader, can not handle the request, task_id={}", task_id);
+        return make_async_error<>(ReplServiceError::NOT_LEADER);
+    }
+    if (commit_quorum >= 1) {
+        // Reduce the quorum size BEFORE checking leadership. When two members are down, the raft leader will
+        // eventually yield its leadership after leadership_expiry (default: 20x heartbeat interval) because it
+        // cannot reach majority. Once leadership is lost, the remaining single node cannot elect itself without a
+        // reduced election quorum. By calling reset_quorum_size here first (which sets both custom_commit_quorum
+        // and custom_election_quorum to 1), the current leader is able to maintain leadership, and if leadership
+        // was already lost, the node will self-elect on the next election timeout. The caller should retry on
+        // NOT_LEADER to allow time for self-election to complete.
+        reset_quorum_size(commit_quorum, trace_id);
+    }
+
+    if (!is_leader) { return make_async_error<>(ReplServiceError::NOT_LEADER); }
+
+    // I am leader and not out_member
     // TODO support rollback, this could happen when the first task failed, and we want to launch a new task to
     // remediate it. Need to rollback the first task. And for the same task, it's reentrant and idempotent.
     auto existing_task_id = get_replace_member_task_id();
     if (!existing_task_id.empty() && existing_task_id != task_id) {
         RD_LOGE(trace_id, "Step1. Replace member, task_id={} is not the same as existing task_id={}", task_id,
                 existing_task_id);
+        reset_quorum_size(0, trace_id);
         return make_async_error<>(ReplServiceError::REPLACE_MEMBER_TASK_MISMATCH);
     }
 
@@ -239,18 +267,10 @@ AsyncReplResult<> RaftReplDev::start_replace_member(std::string& task_id, const
             return make_async_success<>();
         }
         RD_LOGE(trace_id, "Step1. Replace member invalid parameter, out member is not found, task_id={}", task_id);
+        reset_quorum_size(0, trace_id);
         return make_async_error<>(ReplServiceError::SERVER_NOT_FOUND);
     }
-    if (m_my_repl_id != get_leader_id()) { return make_async_error<>(ReplServiceError::NOT_LEADER); }
-    // Check if leader itself is requested to move out.
-    if (m_my_repl_id == member_out.id) {
-        // immediate=false successor=-1, nuraft will choose an alive peer with highest priority as successor, and wait
-        // until the successor finishes the catch-up of the latest log, and then resign. Return NOT_LEADER and let
-        // client retry.
-        raft_server()->yield_leadership(false /* immediate */, -1 /* successor */);
-        RD_LOGI(trace_id, "Step1. Replace member, leader is the member_out so yield leadership, task_id={}", task_id);
-        return make_async_error<>(ReplServiceError::NOT_LEADER);
-    }
+
     // quorum safety check. TODO currently only consider lsn, need to check last response time.
     auto active_peers = get_active_peers();
     // active_peers doesn't include leader itself.
@@ -272,18 +292,15 @@ AsyncReplResult<> RaftReplDev::start_replace_member(std::string& task_id, const
                 "Step1. Replace member, quorum safety check failed, active_peers={}, "
                 "active_peers_exclude_out/in_member={}, required_quorum={}, commit_quorum={}, task_id={}",
                 active_peers.size(), active_num, quorum, commit_quorum, task_id);
+        reset_quorum_size(0, trace_id);
         return make_async_error<>(ReplServiceError::QUORUM_NOT_MET);
     }
 
-    if (commit_quorum >= 1) {
-        // Two members are down and leader cant form the quorum. Reduce the quorum size.
-        reset_quorum_size(commit_quorum, trace_id);
-    }
-
     // Step 2: Handle out member.
 #ifdef _PRERELEASE
     if (iomgr_flip::instance()->test_flip("replace_member_set_learner_failure")) {
         RD_LOGE(trace_id, "Simulating set member to learner failure");
+        reset_quorum_size(0, trace_id);
         return make_async_error(ReplServiceError::FAILED);
     }
 #endif

src/tests/test_common/raft_repl_test_base.hpp

@@ -780,24 +780,33 @@ class RaftReplDevTestBase : public testing::Test {
     void truncate(int num_reserved_entries) { dbs_[0]->truncate(num_reserved_entries); }
     repl_lsn_t get_truncation_upper_limit() { return dbs_[0]->get_truncation_upper_limit(); }
 
+    void do_replace_member(std::shared_ptr< TestReplicatedDB > db, std::string& task_id, replica_id_t member_out,
+                           replica_id_t member_in, uint32_t commit_quorum,
+                           ReplServiceError error = ReplServiceError::OK) {
+        LOGINFO("Start replace member task_id={}, out={}, in={}", task_id, boost::uuids::to_string(member_out),
+                boost::uuids::to_string(member_in));
+        replica_member_info out{member_out, ""};
+        replica_member_info in{member_in, ""};
+        auto result =
+            hs()->repl_service().replace_member(db->repl_dev()->group_id(), task_id, out, in, commit_quorum).get();
+        if (error == ReplServiceError::OK) {
+            ASSERT_EQ(result.hasError(), false) << "Error in replacing member, err=" << result.error();
+        } else {
+            ASSERT_EQ(result.hasError(), true);
+            ASSERT_EQ(result.error(), error) << "Error in replacing member, err=" << result.error();
+        }
+    }
+
     void replace_member(std::shared_ptr< TestReplicatedDB > db, std::string& task_id, replica_id_t member_out,
                         replica_id_t member_in, uint32_t commit_quorum = 0,
                         ReplServiceError error = ReplServiceError::OK) {
-        this->run_on_leader(db, [this, error, db, &task_id, member_out, member_in, commit_quorum]() {
-            LOGINFO("Start replace member task_id={}, out={}, in={}", task_id, boost::uuids::to_string(member_out),
-                    boost::uuids::to_string(member_in));
-
-            replica_member_info out{member_out, ""};
-            replica_member_info in{member_in, ""};
-            auto result =
-                hs()->repl_service().replace_member(db->repl_dev()->group_id(), task_id, out, in, commit_quorum).get();
-            if (error == ReplServiceError::OK) {
-                ASSERT_EQ(result.hasError(), false) << "Error in replacing member, err=" << result.error();
-            } else {
-                ASSERT_EQ(result.hasError(), true);
-                ASSERT_EQ(result.error(), error) << "Error in replacing member, err=" << result.error();
-            }
-        });
+        if (commit_quorum == 0) {
+            this->run_on_leader(db, [this, error, db, &task_id, member_out, member_in, commit_quorum]() {
+                do_replace_member(db, task_id, member_out, member_in, commit_quorum, error);
+            });
+        } else {
+            do_replace_member(db, task_id, member_out, member_in, commit_quorum, error);
+        }
     }
 
     void remove_member(std::shared_ptr< TestReplicatedDB > db, replica_id_t member_id) {

src/tests/test_raft_repl_dev_dynamic.cpp

@@ -198,13 +198,33 @@ TEST_F(ReplDevDynamicTest, TwoMemberDown) {
         LOGINFO("Shutdown replica 2");
     }
 
+    LOGINFO("Sleep 10 seconds to waiting for leadership expiring");
+    sleep(10);
     std::string task_id = "task_id";
     if (g_helper->replica_num() == 0) {
         // Replace down replica 2 with spare replica 3 with commit quorum 1
         // so that leader can go ahead with replacing member.
+        // After reset_quorum_size(1) is applied, the node may need one election timeout
+        // to self-elect as leader, so retry on NOT_LEADER.
         LOGINFO("Replace member started, task_id={}", task_id);
-        replace_member(db, task_id, g_helper->replica_id(member_out), g_helper->replica_id(member_in),
-                       1 /* commit quorum*/);
+        constexpr int max_retries = 3;
+        bool succeeded = false;
+        for (int i = 0; i < max_retries; ++i) {
+            auto result = hs()->repl_service()
+                              .replace_member(db->repl_dev()->group_id(), task_id,
+                                              replica_member_info{g_helper->replica_id(member_out), ""},
+                                              replica_member_info{g_helper->replica_id(member_in), ""}, 1)
+                              .get();
+            if (!result.hasError()) {
+                succeeded = true;
+                break;
+            }
+            ASSERT_EQ(result.error(), ReplServiceError::NOT_LEADER)
+                << "Replace member failed with unexpected error: " << result.error();
+            LOGINFO("Replace member returned NOT_LEADER, retry {}/{}", i + 1, max_retries);
+            std::this_thread::sleep_for(std::chrono::seconds(2));
+        }
+        ASSERT_TRUE(succeeded) << "Replace member failed after " << max_retries << " retries";
         this->write_on_leader(num_io_entries, true /* wait_for_commit */);
         LOGINFO("Leader completed num_io={}", num_io_entries);
     }