On Wed, Sep 5, 2012 at 9:42 AM, Smart Weblications GmbH - Florian Wiessner <f.wiessner@xxxxxxxxxxxxxxxxxxxxx> wrote: > Am 05.09.2012 18:22, schrieb Tommi Virtanen: >> On Tue, Sep 4, 2012 at 4:26 PM, Smart Weblications GmbH - Florian >> Wiessner <f.wiessner@xxxxxxxxxxxxxxxxxxxxx> wrote: >>> i set up a 3 node ceph cluster 0.48.1argonaut to test ceph-fs. >>> >>> i mount ceph via fuse, then i downloaded kernel tree and decompressed a few >>> times, then stopping one osd (osd.1), afer a while of recovering, suddenly: > >> >> Please provide English error messages when you share things with the >> list. In this case I can figure out what the message is, but really, >> we're all pattern matching animals and the specific strings in >> /usr/include/asm-generic/errno.h are what we know. >> > > OK, will change locales. > >>> no space left on device, but: >>> >>> 2012-09-04 18:46:38.242840 mon.0 [INF] pgmap v2883: 576 pgs: 512 active+clean, >>> 64 active+recovering; 1250 MB data, 14391 MB used, 844 MB / 15236 MB avail; >>> 36677/215076 degraded (17.053%) >>> >>> there is space left? >> >> Only 844 MB available, with the pseudo-random placement policies, >> means you practically are out of space. >> >> It looks like you had only 15GB to begin with, and with typical >> replication, that's <5GB usable space. That is dangerously small for >> any real use; Ceph currently does not cope very well with running out >> of space. >> > > It is a test-cluster running on my thinkpad, its main purpose is to test cephfs, > there is no need for real space. I added osd.1 again, then after recovery the > problem went away. I forced this situation to check how cephfs will behave when > cluster is near-full, osd fails and ceph tries to recover until backfill fills > up other osds so ceph is full. > > I observed on the client that no IO was possible anymore so that the client was > unusable. > > Is there a smarter way to handle this? It is bad that cephfs then stalls, it > would be better if it just returns that there is no space left, but still allow > read access... can this be tuned somewhere? What client were you using? I believe it does allow reads while full — but your client can pretty easily get itself into a situation where it needs to perform writes in order to continue doing reads. -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
