Richard Loveland (rmloveland) commented:

Right now our docs on transaction records state that if a record is found not to exist,

If a transaction encounters a write intent whose transaction record doesn't exist, it uses the write intent's timestamp to determine how to proceed. If the write intent's timestamp is within the transaction liveness threshold, the write intent's transaction is treated as if it is PENDING, otherwise it's treated as if the transaction is ABORTED.

However, as Irfan pointed out in a discussion about a different docs change,

the reason a txn record is found to not exist is because we're lazy about creating txn records. For short txns we try to immediately go to the STAGING state, instead of first writing out and explicit PENDING state.

This information is captured in the section on parallel commits but should be described more explicitly here as well, since this is the section on txn records.

Jira Issue: DOC-479