* What is durability
  * D in ACID
  * persisted on non-volatile store
  * far from always required in strictest sense
* write ahead log / journaling
* expensive
  * at least one storage round-trip per commit
  * write latency => minimum commit latency
  * often at least one iop/commit
  * small commits can wear out flash somewhat fast
* synchronous_commit = off / local
  * per transaction / session / role / database / instance
  * full checkpoint overhead
* unlogged tables
  * not available on replica
  * vanish on crash
  * checkpoint overhead reduced
  * some WAL still written
* fsync = off
* guarantees of (streaming) replication
* synchronous replication
  * timeouts
  * < 9.6: one node
  * >= 9.6: multiple nodes
  * >= 10: quorum?
* synchronous_commit = remote_apply
* 2PC
  * coordination
  * faster in 9.6