[LU-2605] reduce the failover time when the MDS and OSS are on the same machine in High Availability

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.1.0
• Type: Improvement
• Priority: Trivial
• Reporter: gmsw
• Assignee: WC Triage
• Resolution: Duplicate
• Labels: server

Environment

Linux, 2 physical servers,
the 1st holds the active MDS and an active OSS and the 2nd holds the passive MDS and an active OSS

Description

In case of failure of the 1st server, the MDS1 and the OSS1 will fail in the same time.

Then the failover on the 2nd node will wait 5 more minutes because the moved OST will wait the timeout for the failed client MDS1.

This case of 2 servers with failover and 2 active OSS is important for little configurations.

(BULL)

Issue Links

• is related to: LU-12722 exclude local client mounted on MDS/O...

Comments

[Andreas Dilger]

This is somewhat expected behaviour if the MDT and OST are on the same node, which is why we do not recommend this configuration. However, the MDT connection to the OST1 (regardless of which node it is running on) should always use the same connection UUID even after a reboot (unlike a client node, which is always gets a new UUID for each mount), and so the failed-over MDT should be able to participate in recovery on node 2 as well.

Do you mount the MDT on node 2 at the same time as OST1?

Do you also have a lustre client mounted locally on the MDS/OSS server node? This would be more likely to be the cause of this recovery slowdown. Does this slow recovery also happen if you restart the MDT and OST on node 1 again, instead of failing over to node 2?

Having your actual console logs from during the recovery would make this more clear.

[Johann Lombardi (Inactive)]

This ticket is actually related to a question i got during a training class (i.e. is it safe to run the MDT on the same node as OSTs?).
I replied that the only problem (Andreas, can you think of anything else?) i could think of is this recovery issue since the MDS might not be able to join recovery. That said, Andreas is right, we reuse the same UUID and allow the MDS to reconnect. As a result, OSTs should be able to complete recovery.

[Andreas Dilger]

This should be fixed with patch https://review.whamcloud.com/36025 "LU-12722 target: disable recovery for local clients"