Marking osds down is not without risks. You are taking away one of the
copies of data for every PG on that osd. Also you are causing every PG on
that osd to peer. If that osd comes back up, every PG on it again needs to
peer and then they need to recover.

That is a lot of load and risks to automate into the system. Now let's take
into consideration other causes of slow requests like having more IO load
than your spindle can handle, backfilling settings set to aggressively
(related to the first option), or networking problems. If the mon is
detecting slow requests on OSDs and marking them down, you could end up
marking half of your cluster down or causing corrupt data by flapping OSDs.

The mon will mark osds down if those settings I mentioned are met. If the
osd isn't unresponsive enough to not respond to other OSDs or the mons,
then there really isn't much that ceph can do to automate this safely.
There are just so many variables. If ceph was a closed system on specific
hardware, it could certainly be monitoring that hardware closely for early
warning signs... But people are running Ceph on everything they can compile
it for including raspberry pis. The cluster admin, however, should be able
to add their own early detection for failures.

You can monitor a lot about disks including things such as average await in
a host to see if the disks are taking longer than normal to respond. That
particular check led us to find that we had several storage nodes with bad
cache batteries on the controllers. Finding that explained some slowness we
had noticed in the cluster. It also led us to a better method to catch that
scenario sooner.

On Tue, Mar 6, 2018, 11:22 PM shadow_lin <> wrote:

> Hi Turner,
> Thanks for your insight.
> I am wondering if the mon can detect slow/blocked request from certain osd
> why can't mon mark a osd with blocked request down if the request is
> blocked for a certain time.
> 2018-03-07
> ------------------------------
> shadow_lin
> ------------------------------
> *发件人:*David Turner <>
> *发送时间:*2018-03-06 23:56
> *主题:*Re: [ceph-users] Why one crippled osd can slow down or block all
> request to the whole ceph cluster?
> *收件人:*"shadow_lin"<>
> *抄送:*"ceph-users"<>
> There are multiple settings that affect this.  osd_heartbeat_grace is
> probably the most apt.  If an OSD is not getting a response from another
> OSD for more than the heartbeat_grace period, then it will tell the mons
> that the OSD is down.  Once mon_osd_min_down_reporters have told the mons
> that an OSD is down, then the OSD will be marked down by the cluster.  If
> the OSD does not then talk to the mons directly to say that it is up, it
> will be marked out after mon_osd_down_out_interval is reached.  If it does
> talk to the mons to say that it is up, then it should be responding again
> and be fine.
> In your case where the OSD is half up, half down... I believe all you can
> really do is monitor your cluster and troubleshoot OSDs causing problems
> like this.  Basically every storage solution is vulnerable to this.
> Sometimes an OSD just needs to be restarted due to being in a bad state
> somehow, or simply removed from the cluster because the disk is going bad.
> On Sun, Mar 4, 2018 at 2:28 AM shadow_lin <> wrote:
>> Hi list,
>> During my test of ceph,I find sometime the whole ceph cluster are blocked
>> and the reason was one unfunctional osd.Ceph can heal itself if some osd is
>> down, but it seems if some osd is half dead (have heart beat but can't
>> handle request) then all the request which are directed to that osd would
>> be blocked. If all osds are in one pool and the whole cluster would be
>> blocked due to that one hanged osd.
>> I think this is because ceph will try to distribute the request to all
>> osds and if one of the osd wont confirm the request is done then everything
>> is blocked.
>> Is there a way to let ceph to mark the the crippled osd down if the
>> requests direct to that osd are blocked more than certain time to avoid the
>> whole cluster is blocked?
>> 2018-03-04
>> ------------------------------
>> shadow_lin
>> _______________________________________________
>> ceph-users mailing list
ceph-users mailing list

Reply via email to