Ceph

The first cut at fixing this problem is in wip-7116. Sage had comments on the pull request for this branch https://github.com/ceph/ceph/pull/1087.

Below is a script which created the orphaned pg_temps when run on my build machine in the dir "src" in a build tree.

#! /bin/sh

OSD=4 ./vstart.sh -l -n -d -o "osd_min_pg_log_entries=5" -o "osd_max_pg_log_entries=10" 
sleep 10

while(true)
do

  ./rados rmpool testpgtemp testpgtemp --yes-i-really-really-mean-it
  ./rados mkpool testpgtemp
  ./ceph osd pool set testpgtemp size 1
  ./rados -p testpgtemp bench 240 write --no-cleanup
  ./ceph osd pool set testpgtemp size 4

  while(true)
  do
    ./ceph osd dump > foo
    clear
    date
    cat foo
    grep pg_temp foo
    if [ $? = "0" ];
    then
      break
    fi
  done

  echo removing... 
  ./rados rmpool testpgtemp testpgtemp --yes-i-really-really-mean-it

  ./ceph osd dump | grep pg_temp > /dev/null
  if [ $? = "0" ];
  then
    break
  fi
done

wip-7116-joao ; https://github.com/ceph/ceph/pull/1153

Haven't been able to reproduce using David's test for well over an hour now, at it used to be a matter of minutes.