I had an image cache issue with Ceph backend: because of a previous incorrect image property I had images locally cached on the compute nodes (shared storage), this disabled the advantages of fast copy-on-write clones. Fixing the image's property and deleting the cache (/var/lib/instances/_base) didn't solve anything. To let others find it faster than me, this answer solved my problem. Setting cache_images = none in /etc/nova/nova.conf on all compute nodes solved my issue.

To affect the behavior of the image cache, The following settings are in nova.conf

# Number of seconds to wait between runs of the image cache
# manager. Set to -1 to disable. Setting this to 0 will
# disable, but this will change in the K release to mean "run
# at the default rate". (integer value)
#image_cache_manager_interval=2400

# Where cached images are stored under $instances_path. This
# is NOT the full path - just a folder name. For per-compute-
# host cached images, set to _base_$my_ip (string value)
#image_cache_subdirectory_name=_base

# Should unused base images be removed? (boolean value)
#remove_unused_base_images=true

# Unused unresized base images younger than this will not be
# removed (integer value)
#remove_unused_original_minimum_age_seconds=86400