"Qingqing Zhou" <zhouqq@cs.toronto.edu> writes:
> What if we change to cost-based removal, i.e., remove the one whose cost is
> smaller. In this case, an two-days-to-be-done reindex should never get
> killed.
Instead, it would kill all useful work in your system :-(. An old
transaction would be the one invincible gunslinger in town, mowing
down every transaction that dared to go up against it, and getting
older and even less vincible with every win. Yet this would still
not give it any guarantee of ever being able to commit, as long as
there's a constant stream of competition.
Certainly it's interesting to think about a smarter deadlock-escape
algorithm, but that's no solution compared to not getting into a
deadlock in the first place.
regards, tom lane