Transaction Optimizations

[qpfiffer]

Idle thoughts on ways to improve the transaction implementation:

• dirty state: if a transaction didn't actually do anything (quite common for ol_unjar operations) then we should just discard the transaction when we're done instead of writing back an empty AOL file. Done.
• Don't read the transaction's AOL file from disk: This is probably the slowest part of the whole procedure. I did it in a really dumb way. The transaction is written to disk, and then read back in from disk during commit.

[Hamcha]

Keeping all them in memory at all time isn't the best either, can't we have either a timeout or a "when I'm holding more than X transactions I'll put N oldest away" thing?

[qpfiffer]

@Hamcha Wow I guess I never saw your comment, but yeah with the current architecture we should be able to do that.