[tarantool-patches] Re: [PATCH v3] replication: fix assertion with duplicate connection

  • From: Vladimir Davydov <vdavydov.dev@xxxxxxxxx>
  • To: Olga Arkhangelskaia <arkholga@xxxxxxxxxxxxx>
  • Date: Tue, 25 Sep 2018 13:26:14 +0300

On Tue, Sep 25, 2018 at 11:53:52AM +0300, Olga Arkhangelskaia wrote:

diff --git a/src/box/applier.cc b/src/box/applier.cc
index 604119e9d..1f99f8090 100644
--- a/src/box/applier.cc
+++ b/src/box/applier.cc
@@ -630,7 +630,7 @@ applier_f(va_list ap)
                              return -1;
                      }
              } catch (FiberIsCancelled *e) {
-                     applier_disconnect(applier, APPLIER_OFF);
+                     applier_disconnect(applier, APPLIER_DISCONNECTED);

APPLIER_DISCONNECTED is a state that denotes temporary errors.
It means that the applier will reconnect soon. In case when an
applier is permanently stopped, we should use APPLIER_OFF.

                      break;
              } catch (SocketError *e) {
                      applier_log_error(applier, e);
diff --git a/src/box/box.cc b/src/box/box.cc
index f25146d01..d3aeb5de0 100644
--- a/src/box/box.cc
+++ b/src/box/box.cc
@@ -667,8 +667,10 @@ box_sync_replication(bool connect_quorum)
              diag_raise();
 
      auto guard = make_scoped_guard([=]{
-             for (int i = 0; i < count; i++)
+             for (int i = 0; i < count; i++) {
+                     applier_stop(appliers[i]);
                      applier_delete(appliers[i]); /* doesn't affect diag */
+             }

As I mentioned in the previous review, this makes replicaset_connect
protocol obscure: turns out it may or may not leave appliers running.
I think that replicaset_connect should revert all appliers to the state
they were in before this function was called, i.e. stopped.

      });
 
      replicaset_connect(appliers, count, connect_quorum);
diff --git a/src/box/replication.cc b/src/box/replication.cc
index 5755ad45e..0f205212e 100644
--- a/src/box/replication.cc
+++ b/src/box/replication.cc
@@ -427,6 +427,7 @@ replicaset_update(struct applier **appliers, int count)
      auto uniq_guard = make_scoped_guard([&]{
              replica_hash_foreach_safe(&uniq, replica, next) {
                      replica_hash_remove(&uniq, replica);
+                     replica_clear_applier(replica);
                      replica_delete(replica);
              }
      });
@@ -454,6 +455,9 @@ replicaset_update(struct applier **appliers, int count)
              replica->uuid = applier->uuid;
 
              if (replica_hash_search(&uniq, replica) != NULL) {
+                     applier = replica->applier;

This line is pointless.

+                     replica_clear_applier(replica);
+                     replica_delete(replica);
                      tnt_raise(ClientError, ER_CFG, "replication",
                                "duplicate connection to the same replica");
              }
diff --git a/test/replication/misc.test.lua b/test/replication/misc.test.lua
index 56e1bab69..95677e12b 100644
--- a/test/replication/misc.test.lua
+++ b/test/replication/misc.test.lua
@@ -162,4 +162,22 @@ test_run:cmd("stop server replica_auth")
 test_run:cmd("cleanup server replica_auth")
 test_run:cmd("delete server replica_auth")
 
+
+--
+-- Test case for gh-3610. Before the fix replica would fail with the 
assertion
+-- when trying to connect to the same master twice.
+--
+box.schema.user.grant('guest', 'replication')
+test_run:cmd("create server replica with rpl_master=default, 
script='replication/replica.lua'")
+test_run:cmd("start server replica")
+test_run:cmd("switch replica")
+replication = box.cfg.replication
+box.cfg{replication = {replication, replication}}
+
+test_run:cmd("switch default")
+box.schema.user.revoke('guest', 'replication')
+test_run:cmd("stop server replica")
+test_run:cmd('cleanup server replica')
+test_run:cmd("delete server replica")
+
 box.schema.user.drop('cluster')

Your test case should go before this line, because user 'cluster' was
created by the previous test case.

Also, there's still no test for the case when applier detects duplicate
connection asynchronously. It should be pretty easy to implement:

 1. On the master (default instance). Disable replication by clearing
    box.cfg.listen.
 2. On the replica. Set box.cfg.replication_connect_quorum to 0 and
    replication_connect_timeout to 0.001. Then set box.cfg.replication
    so that the replica tries to connect to the master twice. Wait for
    box.cfg() to return.
 3. On the master. Enable replication by setting box.cfg.listen back to
    its original value.
 4. On the replica. Check that replication connection is established and
    that 'duplicate connection' error is printed to the log.

Other related posts:

  1. Home
  2. tarantool-patches
  3. Archive
  4. 09-2018