TransactionSequence was emitting result() twice when rolling back.
- How did this happen?
The TransactionRollbackJob is (automatically) added as a subjob of the
TransactionSequence, so when it finishes, slotResult is called (like for
all subjobs), as well as rollbackResult().
Since the latter emits result() already [mostly for symmetry with
commitResult()], we don't need to do that in slotResult (which doesn't do
it for the case of committing, either).
- Why is it a problem to emit result() twice?
Well, first, it's against the law in KJob world. In practice,
ItemSyncPrivate::slotTransactionResult was called twice (for the same
TransactionSequence job) which made it decrement mTransactionJobs one
time too many.
As a result, checkDone() finished too early and didn't go into the
"commit transaction" branch for other transactions.
Leaving a transaction "open" is a good recipe for database deadlocks further
down the line.
- Why did the TransactionSequence roll back in the first place?
In my case because of the infamous and not-yet fixed "Multiple merge
candidates" problem, but it seems that it can also happen when having
items without a part, according to Volker's investigations.
All of these issues still need to be fixed, but at least akonadi seems
to be still usable after they happen.