High Availability is just too unstable, nodes constantly turn off and on

[LunarMagic]

I don't know why HA keeps messing up so much. I've tried this twice now. After i enable HA on all of the virtual machines i then get nodes randomly rebooting nonstop. I tried changing settings in HA and i just can't get this to stop. When i turn off HA all of my nodes are perfectly fine

 

[LunarMagic]

screenshot of the errors i get

 

[UdoB]

The classic suspicion is that your corosync network is weak, has high latency or is not separate from a (congested) "main" network cable. Remember that VLANs do NOT have separate wires per LAN ;-)

Then the obvious recommendation would be to establish physically one (or two) independent cable-networks exclusively for corosync...


Disclaimer: pure guessing...

 

[LunarMagic]

The classic suspicion is that your corosync network is weak, has high latency or is not separate from a (congested) "main" network cable. Remember that VLANs do NOT have separate wires per LAN ;-)

Then the obvious recommendation would be to establish physically one (or two) independent cable-networks exclusively for corosync...


Disclaimer: pure guessing...

Each one of the machines are on their own switch together with their own connections

 

[UdoB]

Each one of the machines are on their own switch together with their own connections

Okay, great!

Of course there may be several causes of a cluster reboot - think a glitch on the power line.

If the reason was HA --> fencing you could find entries about it in the journal, similar like this:

~# journalctl   --grep fenc
-- Boot 2052e14b8b124d3e8cca747c9c998d64 --
-- Boot 8a67ec26cb82418592d5b22a0cd97c3d --
-- Boot 46c01a76137346028f356e323f5f8bcd --
Dec 18 19:03:27 pvem pve-ha-crm[2232]: node 'pvei': state changed from 'unknown' => 'fence'
Dec 18 19:04:28 pvem pve-ha-crm[2232]: fencing: acknowledged - got agent lock for node 'pvei'
Dec 18 19:04:28 pvem pve-ha-crm[2232]: node 'pvei': state changed from 'fence' => 'unknown'

Look for it on some nodes. If you find some examine the journal entries a few minutes before that timestamp.

To check live corosync connectivity I use

~# corosync-cfgtool -n
Local node ID 10, transport knet
nodeid: 2 reachable
   LINK: 0 udp (10.3.16.13->10.3.16.9) enabled connected mtu: 1397
   LINK: 1 udp (10.11.16.13->10.11.16.9) enabled connected mtu: 1397
...

Verify that all of the expected nodes are listed and have consistent settings.

 

[waltar]

Verify that all of your nodes have same time and synchronised eg with chrony.