Christophe Pettus <[email protected]> writes:
>> On Apr 9, 2018, at 07:33, Thomas Poty <[email protected]> wrote:
>> ok, and long answer ? is it random?
> It's not literally random, but from the application point of view, it's not
> predictable. For example, it's not always the one that opened first, or any
> other consistent measure.
It's whichever one runs the deadlock detector first after the circular
wait becomes established. For instance:
* Process A takes lock L1
* Process B takes lock L2
* Process A tries to take lock L2, blocks
* Process B tries to take lock L1, blocks (now a deadlock exists)
Process A will run the deadlock detector one deadlock_timeout after
blocking. If that happens before B has blocked, then A will see
no deadlock and will go back to waiting. In that case, when B's
own deadlock_timeout expires and it runs the deadlock detector,
it will see the deadlock and fix it by canceling its own wait.
On the other hand, if B started to wait less than one deadlock_timeout
after A did, then A will be first to observe the deadlock and it will
cancel itself, not B.
So you can't predict it unless you have a lot of knowledge about
the timing of events. You could probably make it more predictable
by making deadlock_timeout either very short or very long, but
neither of those are desirable things to do.
regards, tom lane
