intro

osd cluster map
 requirements
 desireable properties
 (c)rush

failure detection
 distributed ping or heartbeat
 central filter, notifier

design
 placement seed, class/superset, groups

normal operation
 reads
 writes

recovery
 triggers: failed disk, or total cluster reorganization

 notify
 peering
 pull
 push
 clean

writes during recovery

graceful data loss + recovery?