Re: XTM & parallel search

On Fri, Jun 3, 2016 at 6:49 PM, Konstantin Knizhnik <
k(dot)knizhnik(at)postgrespro(dot)ru> wrote:

>
>
> On 03.06.2016 16:05, Amit Kapila wrote:
>
> On Fri, Jun 3, 2016 at 1:34 AM, Konstantin Knizhnik <
> <k(dot)knizhnik(at)postgrespro(dot)ru>k(dot)knizhnik(at)postgrespro(dot)ru> wrote:
>
>> We have to add three more functions to eXtensible Transaction Manager API
>> (XTM):
>>
>> /*
>> * Calculate transaction state size. This method is invoked by
>> EstimateTransactionStateSpace to copy transaction
>> * state to parallel workers
>> */
>> size_t (*GetTransactionStateSize)(void);
>>
>> /*
>> * Serialize transaction state
>> */
>> void (*SerializeTransactionState)(void* ctx);
>>
>> /*
>> * Deserialize transaction state
>> */
>> void (*DeserializeTransactionState)(void* ctx);
>>
>>
>
> In above proposal, are you suggesting to change the existing API's as
> well, because the parameters of function pointers don't match with exiting
> API's. I think it is better to consider this along with the overall XTM API.
>
>
> Sorry, but right now I have not replaced existed functions
>
> EstimateTransactionStateSpace/SerializeTransactionState/StartParallelWorkerTransaction
> with XTM indirect calls. If was my original intention, but these functions
> access static variable CurrentTransactionState defined in xact.c.
> So if user-defined TM wants to override this functions, it will have to
> invoke original functions to save/restore CurrentTransactionState.
> It is not convenient.
> This is why three XTM functions above are now called by existed xact
> funcations to save additional state, for example:
>
>
Sounds reasonable, although one might want to have a hook for
EndParallelWorkerTransaction as well. I think here larger question is
whether there is a consensus on getting XTM API in core.

With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com