Following up on this and other comments, there are 2 different time delays. One (1) is the time it takes from killing an MDS until a stand-by is made an active rank, and (2) the time it takes for the new active rank to restore all client sessions. My experience is that (1) takes close to 0 seconds while (2) can take between 20-30 seconds depending on how busy the clients are; the MDS will go through various states before reaching active. We usually have ca. 1600 client connections to our FS. With fewer clients, MDS fail-over is practically instantaneous. We are using latest mimic. From what you write, you seem to have a 40 seconds window for (1), which points to a problem different to MON config values. This is supported by your description including a MON election (??? this should never happen). Do you have have services co-located? Which of the times (1) or (2) are you referring to? How many FS clients do you have? Best regards, ================= Frank Schilder AIT Risø Campus Bygning 109, rum S14 ________________________________________ From: Patrick Donnelly <pdonnell@redhat.com> Sent: 03 May 2021 17:19:37 To: Lokendra Rathour Cc: Ceph Development; dev; ceph-users Subject: [ceph-users] Re: [ Ceph MDS MON Config Variables ] Failover Delay issue On Mon, May 3, 2021 at 6:36 AM Lokendra Rathour <lokendrarathour@gmail.com> wrote:
Hi Team, I was setting up the ceph cluster with
- Node Details:3 Mon,2 MDS, 2 Mgr, 2 RGW - Deployment Type: Active Standby - Testing Mode: Failover of MDS Node - Setup : Octopus (15.2.7) - OS: centos 8.3 - hardware: HP - Ram: 128 GB on each Node - OSD: 2 ( 1 tb each) - Operation: Normal I/O with mkdir on every 1 second.
T*est Case: Power-off any active MDS Node for failover to happen*
*Observation:* We have observed that whenever an active MDS Node is down it takes around* 40 seconds* to activate the standby MDS Node. on further checking the logs for the new-handover MDS Node we have seen delay on the basis of following inputs:
1. 10 second delay after which Mon calls for new Monitor election 1. [log] 0 log_channel(cluster) log [INF] : mon.cephnode1 calling monitor election
In the process of killing the active MDS, are you also killing a monitor? -- Patrick Donnelly, Ph.D. He / Him / His Principal Software Engineer Red Hat Sunnyvale, CA GPG: 19F28A586F808C2402351B93C3301A3E258DD79D _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io