On 09/20/18 00:44, Tom Lane wrote:
> Chapman Flack <chap@anastigmatix.net> writes:
>> Would it be unprecedented / be unreasonable / break anything for the
>> install_jar function to simply force a CommandCounterIncrement
>> at the end of step 1 (after its temporary snapshot has been popped,
>> so the former/on-entry ActiveSnapshot gets the increment)?
>
> The risk you take there is changing the behavior of calling function(s).
>
>> DECISION TIME ...
>
>> 1. fiddle the loader to always pass read_only => false to SPI calls,
>> regardless of the volatility of the function it is loading for.
>> 2. leave the loader alone, and adjust install_jar (an infrequent
>> operation) to do something heretical with its on-entry snapshot.
>
> I suspect #1 is less likely to have bad side-effects. But I've not
> done any careful analysis.
Or, revisiting #2, what of install_jar first pushing a copied snapshot
(the current active one? a new one from GetTransactionSnapshot?) and
keeping that copy on the stack during both operations (loading the jar
and executing the deployment commands, with a CommandCounterIncrement
in between) and popping it before return?
Would that alleviate the concern of changing calling functions' behavior?
-Chap