The science of crash consistency

barriers are not used
but we usually get away with it becuase journal is sequentially written

consistency is not a binary thing -- probabilistic crashconsistency
What's the probability of inconsistency upon a crash?

P = time in (inconsistency) window / total time
(They assume/observe that for a give storage configuation, there are not many ordering obseved. So they average these out)

Different re-ordering:
commit-commit: second transaction before the first transaction
commit-data: data block before transaction commits
commit-metadata: metadata block before transaction commits
checkpoint-commit
checkpoint-data
checkpoint-metadata