ceph - cephfs - 2024-06-26

Timestamp (UTC)Message
2024-06-26T04:48:22.054Z
<Venky Shankar> You could attach the logs in the tracker @Erich Weiler
2024-06-26T11:19:00.794Z
<Dhairya Parmar> i have the mds log downloaded from them
2024-06-26T11:51:06.048Z
<Dhairya Parmar> @Venky Shankar
2024-06-26T11:52:25.606Z
<Dhairya Parmar> is it possible to dump `pre_allocated` inos set using ceph-dencoder?
2024-06-26T12:57:21.660Z
<Dhairya Parmar> @Kotresh H R from your today's status
> I found that's because of the OSD full of metadata pool and data pool. Because of this fs connection was hung during mount.
Usually client would bail out with the `Client::_handle_full_flag` , if it hung looks like the osd full/nearfull/backfillfull ratio was closer to defaults?
2024-06-26T13:01:25.519Z
<Dhairya Parmar> @Leonid Usov <https://github.com/ceph/ceph/pull/58281>
2024-06-26T13:01:33.682Z
<Dhairya Parmar> do you think pr is in the right direction?
2024-06-26T13:02:38.305Z
<Dhairya Parmar> at least from my runs, that was the case. since the metadata pool is full, mds starts reporting slow metadata io, client can't trim the cache and the mount is hung
2024-06-26T13:03:16.530Z
<Dhairya Parmar> but if you reduce to some fair values like
```mon osd nearfull ratio: 0.6
mon osd backfillfull ratio: 0.6
mon osd full ratio: 0.7```
it works fine and the client should bail out
2024-06-26T13:03:26.759Z
<Kotresh H R> It was default
2024-06-26T13:03:36.970Z
<Kotresh H R> in vstart cluster
2024-06-26T13:03:40.555Z
<Dhairya Parmar> ok so yes it should be these values then
2024-06-26T13:07:27.402Z
<Leonid Usov> does that compile 🙂
2024-06-26T13:07:28.225Z
<Leonid Usov> ?
2024-06-26T13:08:18.601Z
<Dhairya Parmar> haven' tried, this is just to get an initial opinion from you whether my approach is correct
2024-06-26T13:08:40.981Z
<Leonid Usov> it looks like in the right direction, yes
2024-06-26T13:08:43.513Z
<Leonid Usov> I’ll add comments
2024-06-26T13:09:05.014Z
<Dhairya Parmar> okay
2024-06-26T13:10:29.069Z
<Dhairya Parmar> AFAICK, the behaviour is quite random when the ratios are default. If you start the vstart with setting these values
```index 59a3798744d..e86a820c190 100755
--- a/src/vstart.sh
+++ b/src/vstart.sh
@@ -804,10 +804,11 @@ prepare_conf() {
 
 [global]
         fsid = $(uuidgen)
-        osd failsafe full ratio = .99
-        mon osd full ratio = .99
-        mon osd nearfull ratio = .99
-        mon osd backfillfull ratio = .99
+        osd failsafe full ratio = 1.0
+        osd mon report interval = 5
+        mon osd full ratio = 0.7
+        mon osd nearfull ratio = 0.6
+        mon osd backfillfull ratio = 0.6
         mon_max_pg_per_osd = ${MON_MAX_PG_PER_OSD:-1000}
         erasure code dir = $EC_PATH
         plugin dir = $CEPH_LIB
@@ -866,10 +867,11 @@ EOF
         bluestore_spdk_mem = 2048"
         else
             BLUESTORE_OPTS="        bluestore block db path = $CEPH_DEV_DIR/osd\$id/block.db.file
-        bluestore block db size = 1073741824
+        bluestore block db size = 53687091
         bluestore block db create = true
         bluestore block wal path = $CEPH_DEV_DIR/osd\$id/block.wal.file
-        bluestore block wal size = 1048576000
+        bluestore block wal size = 52428800
+        bluestore block size = 5368709120
         bluestore block wal create = true"
             if [ ${#block_devs[@]} -gt 0 ] || \
                [ ${#bluestore_db_devs[@]} -gt 0 ] || \```
you'd get consistent results
2024-06-26T13:42:23.048Z
<Kotresh H R> Yeah, you are right. but my intention was to test the clone_progress PR and review. So didn't notice the storage space and kept on creating clones, which resulted in the hung.
2024-06-26T14:05:49.465Z
<Erich Weiler> The logs are a 300 MB tarball - would it allow such a large attachment?
2024-06-26T14:06:06.602Z
<Erich Weiler> The logs are a 300 MB tarball - would it allow such a large attachment @Venky Shankar?
2024-06-26T14:07:07.409Z
<Erich Weiler> Can anyone upload items or update an issue in redmine there?
2024-06-26T16:11:20.725Z
<Dhairya Parmar> oh okay. if that's a repetitive workload, i'd suggest you to tune in those values otherwise you're bound to land in such scenario.

Any issue? please create an issue here and use the infra label.