Cluster experiencing complete operational failure, various cephx authentication errors
Hi everyone, I have a serious problem which currently exists of my entire Ceph no longer being able to provide service. As if yesterday I added 10 OSD's total 2 per node, the rebalance started and took some IO but seemed to be doing its work. This morning the cluster was still processing the rebalance and taking so much IO that nearly all OSD's where marked as "slow ops" and from there everything went wrong. As attempt to clear as much IO for de rebalance I stoped all the clients and waited for the rebalance to finish. After it finished the cluster remained extremely slow and unusable. Whilst trying to debug I restarted several services and nodes trying to find the problem. Now the cluster has entered a state where multiple OSD's remain slow, various OSD's show a "BADAUTHORIZER" message and the mgr on all nodes also has issues "verify_authorizer". I verified all the clocks on all servers and they are sinked to the same NTP service and seem good. Please please please advise as straight 13 hours of debugging got me nowhere. Current version: ceph version 14.2.8 (2d095e947a02261ce61424021bb43bd3022d35cb) nautilus (stable) Mgr error example: 2020-08-24 20:19:12.865 7f2baf56a700 0 cephx: verify_authorizer could not get service secret for service mgr secret_id=12230 2020-08-24 20:19:13.043 7f2bb056c700 0 auth: could not find secret_id=12230 2020-08-24 20:19:13.043 7f2bb056c700 0 cephx: verify_authorizer could not get service secret for service mgr secret_id=12230 2020-08-24 20:19:13.210 7f2bb056c700 0 auth: could not find secret_id=12230 2020-08-24 20:19:13.210 7f2bb056c700 0 cephx: verify_authorizer could not get service secret for service mgr secret_id=12230 OSD error example: 2020-08-24 19:47:15.777 7f9957d79700 -1 osd.19 41255 get_health_metrics reporting 72 slow ops, oldest is osd_op(mds.0.1510:4 12.a6 12.4b2c82a6 (undecoded) ondisk+retry+read+known_if_redirected+full_force e41119) 2020-08-24 19:47:15.833 7f995c88b700 0 auth: could not find secret_id=12230 2020-08-24 19:47:15.833 7f995c88b700 0 cephx: verify_authorizer could not get service secret for service osd secret_id=12230 2020-08-24 19:47:15.833 7f995c88b700 0 --1- [v2:10.201.1.17:6814/1030299,v1:10.201.1.17:6815/1030299] >> v1:10.201.1.20:6823/1023281 conn(0x55affb8e7c00 0x55b007441000 :6815 s=ACCEPTING_WAIT_CONNECT_MSG_AUTH pgs=0 cs=0 l=0).handle_connect_message_2: got bad authorizer, auth_reply_len=0
On 2020-08-24 20:35, Mathijs Smit wrote:
Hi everyone,
I have a serious problem which currently exists of my entire Ceph no longer being able to provide service. As if yesterday I added 10 OSD's total 2 per node, the rebalance started and took some IO but seemed to be doing its work. This morning the cluster was still processing the rebalance and taking so much IO that nearly all OSD's where marked as "slow ops" and from there everything went wrong. As attempt to clear as much IO for de rebalance I stoped all the clients and waited for the rebalance to finish. After it finished the cluster remained extremely slow and unusable. Whilst trying to debug I restarted several services and nodes trying to find the problem. Now the cluster has entered a state where multiple OSD's remain slow, various OSD's show a "BADAUTHORIZER" message and the mgr on all nodes also has issues "verify_authorizer".
I verified all the clocks on all servers and they are sinked to the same NTP service and seem good.
Is there a warning of the ceph monitors that there is a time skew? Do your monitors log anything?
Please please please advise as straight 13 hours of debugging got me nowhere.
Can you paste "ceph -s"? Similar reports from users on this list where a bad behaving monitor (PG stuck peering - OSD cephx: verify_authorizer key problem), and an issue in Ceph Nautilus which has been fixed (BADAUTHORIZER in Nautilus). Messenger v2 was also disabled. There have been fixes is the messenger v1 and v2. Do you have any specific ceph auth settings (non default)? Can you check if the keys of the (new) OSD daemons match with the keys stored in ceph (ceph auth list, ceph auth export client.$id)? Gr. Stefan
On 2020-08-24 20:35, Mathijs Smit wrote:
Hi everyone,
I have a serious problem which currently exists of my entire Ceph no longer being able to provide service. As if yesterday I added 10 OSD's total 2 per node, the rebalance started and took some IO but seemed to be doing its work. This morning the cluster was still processing the rebalance and taking so much IO that nearly all OSD's where marked as "slow ops" and from there everything went wrong. As attempt to clear as much IO for de rebalance I stoped all the clients and waited for the rebalance to finish. After it finished the cluster remained extremely slow and unusable. Whilst trying to debug I restarted several services and nodes trying to find the problem. Now the cluster has entered a state where multiple OSD's remain slow, various OSD's show a "BADAUTHORIZER" message and the mgr on all nodes also has issues "verify_authorizer".
I verified all the clocks on all servers and they are sinked to the same NTP service and seem good.
Please please please advise as straight 13 hours of debugging got me nowhere.
If you can pause all client IO, I don't think it can harm to upgrade everything to 14.2.11, just to be sure you don't hit a bug that is already fixed. But it depends on the current health of your cluster I guess. Gr. Stefan
participants (2)
-
Mathijs Smit
-
Stefan Kooman