Bug #19291
Updated by Patrick Donnelly about 7 years ago
If an MDS respawns then its "comm" name becomes "exe" which confuses logrotate since it relies on killlall. What ends up happening is that logrotate will rename the current log to e.g. "ceph-mds.li1015-93.log.1", send SIGHUP to all processes named "ceph-mds" (there are none!), gzip the ceph-mds.li1015-93.log.1 to ceph-mds.li1015-93.log.1.gz and then unlink the original log ceph-mds.li1015-93.log.1. Unfortunately because ceph-mds does not get the SIGHUP, it will keep writing to the unlinked log inode until the disk space is consumed or ceph-mds dies.
This problem was introduced in 66a122025f6cf023cf7b2f3d8fbe4964fb7568a7 and discussed on ceph-devel here: https://www.spinics.net/lists/ceph-devel/msg33780.html
I'm going to suggest the simple (but very slightly racy) solution of calling prctl on startup to unconditionally change the name of the ceph-mds daemon to "ceph-mds". Thoughts?