Avoid diverged slave when migrating MariaDB master

When promoting one slave to the new master in a replication group,
previously the old master will be attached to the new one right after
the new master is on. For MariaDB, attaching the old master to the new
one, new GTID may be created on the old master and also may be synced
to some of the other replicas, as they're still connecting to the old
master. The new GTID does not exists in the new master, making these
slaves diverged from the master. After that, when the diverged slave
connects to the new master, 'START SLAVE' will fail with logs like:

    [ERROR] Error reading packet from server: Error: connecting slave
    requested to start from GTID X-XXXXXXXXXX-XX, which is not in the
    master's binlog. Since the master's binlog contains GTIDs with
    higher sequence numbers, it probably means that the slave has
    diverged due to executing extra erroneous transactions
    (server_errno=1236)

And these slaves will be left orphan and errored after
promote_to_replica_source finishs.

Attaching the other replicas to the new master before dealing with the
old master will fix this problem and the failure of the
trove-scenario-mariadb-multi Zuul job as well.

Closes-Bug: #1754539
Change-Id: Ib9c01b07c832f117f712fd613ae55c7de3561116
Signed-off-by: Zhao Chao <zhaochao1984@gmail.com>

1 files changed, 24 insertions, 2 deletions

diff --git a/trove/taskmanager/manager.py b/trove/taskmanager/manager.py
index e600e7ad..7559c1ec 100644
--- a/trove/taskmanager/manager.py
+++ b/trove/taskmanager/manager.py
@@ -99,6 +99,26 @@ class Manager(periodic_task.PeriodicTasks):
                                        replica_models):
             # First, we transition from the old master to new as quickly as
             # possible to minimize the scope of unrecoverable error
+
+            # NOTE(zhaochao): we cannot reattach the old master to the new
+            # one immediately after the new master is up, because for MariaDB
+            # the other replicas are still connecting to the old master, and
+            # during reattaching the old master as a slave, new GTID may be
+            # created and synced to the replicas. After that, when attaching
+            # the replicas to the new master, 'START SLAVE' will fail by
+            # 'fatal error 1236' if the binlog of the replica diverged from
+            # the new master. So the proper order should be:
+            # -1. make the old master read only (and detach floating ips)
+            # -2. make sure the new master is up-to-date
+            # -3. detach the new master from the old one
+            # -4. enable the new master (and attach floating ips)
+            # -5. attach the other replicas to the new master
+            # -6. attach the old master to the new one
+            #     (and attach floating ips)
+            # -7. demote the old master
+            # What we changed here is the order of the 6th step, previously
+            # this step took place right after step 4, which causes failures
+            # with MariaDB replications.
             old_master.make_read_only(True)
             master_ips = old_master.detach_public_ips()
             slave_ips = master_candidate.detach_public_ips()
@@ -106,10 +126,8 @@ class Manager(periodic_task.PeriodicTasks):
             master_candidate.wait_for_txn(latest_txn_id)
             master_candidate.detach_replica(old_master, for_failover=True)
             master_candidate.enable_as_master()
-            old_master.attach_replica(master_candidate)
             master_candidate.attach_public_ips(master_ips)
             master_candidate.make_read_only(False)
-            old_master.attach_public_ips(slave_ips)
 
             # At this point, should something go wrong, there
             # should be a working master with some number of working slaves,
@@ -138,6 +156,10 @@ class Manager(periodic_task.PeriodicTasks):
                     error_messages += "%s (%s)\n" % (
                         exc_fmt % msg_content, ex)
 
+            # dealing with the old master after all the other replicas
+            # has been migrated.
+            old_master.attach_replica(master_candidate)
+            old_master.attach_public_ips(slave_ips)
             try:
                 old_master.demote_replication_master()
             except Exception as ex: