we had some issues because of maintenance and hope someone can explain what happened.
Our cluster looks like this:
- Node A1 (Quorum vote)
- Node A2 (Quorum vote)
- Node A3 (Quorum vote)
- Node A4 (NO Quorum vote)
- Node B1 (Quorum vote)
- Node B2 (Quorum vote)
- Node B3 (Quorum vote)
- File Share Witness
Every node has 4 NICs that are configured for teaming:
- NIC 1 + 2 teamed | Configured for Cluster and Client
- NIC 3 + 4 teamed | Configured for Cluster
So in total we have 4 VLANs on each node.
All nodes run with Microsoft Windows Server 2016 Standard + Microsoft SQL Server 2016 Enterprise.
Our AGs have one automatic failover partner on both sites each and regularly the nodes in site A are primary. For example AG1:
- A1 - PRIMARY
- B1 - SECONDARY
We had to perform firmware updates on our servers. So we did a manual failover of all AGs from site A to site B. Everything was fine until the installation of the network drivers started. They started at the same time on all 4 server of site A.
When these 4 servers reported that they are not connected to the cluster networks the same happened for all 3 server on site B as well. Server A1 wrote into the cluster log that it lost quorum and it's going to shut down the cluster service. Some of the cluster resources (Availabilty Groups) also went down.
Server B1 for example tried to bring its AGs online again also trying to use the IPs from VLANs of site A.
After a few minutes when the servers on site A where back up again server B1 was able to bring up its resources. In the SQL log we saw that the states of the AGs changed from "PRIMARY_PENDING" to "PRIMARY_NORMAL". But at that time the dns registration wasn't finished. Every database changed from "RESOLVING" to "PRIMARY" except one. The AG itself was fine but the database was stuck in state "Not Synchronized" on all nodes. After the SQL server started listening to the Availabilty Group Listener again we could see the following line with no other notification stating this database inbetween: "Nonqualified transactions are being rolled back in database DB1 for an Always On Availability Groups state change. Estimated rollback completion: 0%. This is an informational message only. No user action is required." The database is very small with just a few hundred MB and there couldn't be much transactions since the application servers which use this database couldn't connect to it. But it was stuck in this rollback state. The percentage of completion didn't even change. Trying to resume it didn't work. I had to abort the query after 3 - 4 minutes. Because nothing changed within 30 minutes we performed a failover of the regarding AG vom B1 to A1 and the databse became healthy within seconds.
I have 3 question I hope someone can answer:
- Why where all 3 remaining nodes on site B not able to cummunicate any longer when the 4 servers on site A where down? They should have been able to communicate with eachother and also the File Share Witness on site C. Could it have something todo with the design described in the following article: https://docs.microsoft.com/en-us/troubleshoot/windows-server/high-availability/cluster-ip-resources-fail-2-node-2-site-fsw-cluster
- How do you prevent the cluster/cluster ressources from going down completely when there is just an issue with cross-vlan communication?
- Any idea what happened to DB1 that was stuck in "No Synchronized" and the roll back process until we triggered a failover?