Ceph » CephFS

For the Linode tests, I've observed that after the clients finish removing their sandboxes (thus making the subtree completely unlinked) each MDS has a number of strays which are not eventually purged. The number varies between MDS:

8x8192 20000C MDS 64x4096 Client.2/stats$ sqlite3 mds-002/stats.db
-- Loading resources from /home/pdonnell/.sqliterc
SQLite version 3.17.0 2017-02-13 16:02:40
Enter ".help" for usage hints.
sqlite> SELECT timestamp, json_extract(status, '$.whoami') AS rank, json_extract(perf_dump, '$.mds_cache.num_strays')
   ...>     FROM MDSStatus
   ...>     WHERE rank >= 0 AND rank != 18446744073709551615
   ...>     ORDER BY timestamp DESC
   ...>     LIMIT 10
   ...>     ;
1488504791|4|130
1488504789|4|130
1488504787|4|130
1488504785|4|130
1488504783|4|130
1488504781|4|130
1488504779|4|130
1488504777|4|130
1488504775|4|130
1488504773|4|130

8x8192 20000C MDS 64x4096 Client.2/stats$ sqlite3 mds-003/stats.db
-- Loading resources from /home/pdonnell/.sqliterc
SQLite version 3.17.0 2017-02-13 16:02:40
Enter ".help" for usage hints.
sqlite> SELECT timestamp, json_extract(status, '$.whoami') AS rank, json_extract(perf_dump, '$.mds_cache.num_strays')
   ...>     FROM MDSStatus
   ...>     WHERE rank >= 0 AND rank != 18446744073709551615
   ...>     ORDER BY timestamp DESC
   ...>     LIMIT 10
   ...>     ;
1488504792|2|119
1488504790|2|119
1488504788|2|119
1488504786|2|119
1488504784|2|119
1488504782|2|119
1488504780|2|119
1488504778|2|119
1488504776|2|119
1488504774|2|119

I've attached a graph of the strays for this experiment. The dataset is here:

mira092.front.sepia.ceph.com:/mnt/pdonnell/vault/8x8192 20000C MDS 64x4096 Client.2