Comment 2 for bug 592379

There are actually 2 ways to implement it and we might make it configurable. One to have transaction which performs update to fail, another is to kill transaction which is oldest (and preventing purge from process). In the former case we probably should kill it some time before the hard limit is reached.

Second approach can be better for a lot of multi user system which prevent application which started run away transaction from making the whole system non functional. I think just to kill such transaction instead of making system to be effectively read only is preferable.