Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> On 2018-Sep-26, Tom Lane wrote:
>> I agree that it would be surprising for transaction timestamp to be newer
>> than statement timestamp. So for now at least, I'd be satisfied with
>> documenting the behavior.
> Really? I thought it was practically obvious that for transaction-
> controlling procedures, the transaction timestamp would not necessarily
> be aligned with the statement timestamp. The surprise would come
> together with the usage of the new feature, so existing users would not
> be surprised in any way.
Nope. That's the same poor reasoning we've fallen into in some other
cases, of assuming that "the user" is a point source of knowledge.
But DBMSes tend to interact with lots of different code. If some part
of application A starts using intraprocedure transactions, and then
application B breaks because it wasn't expecting to see xact_start
later than query_start in pg_stat_activity, you've still got a problem.
I'm also a bit hesitant to invent new semantics here based on the
assumption that we've got only one, nonoverlapping, top-level transaction
at a time. It's not terribly hard to imagine suspend-and-resume-
transaction features coming down the pike at some point. What will
we do then? We'll already have a definitional issue for xact_start,
but it'll get worse the more different kinds of xact_start we have.
regards, tom lane