Re: delta relations in AFTER triggers

On Tue, Sep 23, 2014 at 1:51 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> Add a new p_tableref_hook function pointer, similar to p_paramref_hook.
>>> Whenever the parser sees a RangeVar that it doesn't recognize (or actually,
>>> I think it should call it *before* resolving regular tables, but let's
>>> ignore that for now), it calls the p_tableref_hook. It can return a new
>>> RelationParam node (similar to regular Param), which contains a numeric ID
>>> for the table/tuplestore, as well as its tuple descriptor.
>>>
>>> For the execution phase, add a new array of Tuplestorestates to
>>> ParamListInfo. Similar to the existing array of ParamExternalDatas.
>
>> I haven't been following this issue closely, but this sounds like a
>> really nice design.
>
> I'm on board with the parser hooks part of that. I don't especially agree
> with the idea of a new sub-structure for ParamListInfo: if you do that you
> will need a whole bunch of new boilerplate infrastructure to allocate,
> copy, and generally manage that structure, for darn little gain. What I'd
> suggest is just saying that some Params might have type INTERNAL with
> Datum values that are pointers to tuplestores; then all you need to do is
> remember which Param number has been assigned to the particular tuplestore
> you want. There is already precedent for that in the recursive CTE code,
> IIRC.

Studying this proposed design a bit further, I am a little fuzzy on
what is supposed to happen in this design during parse analysis. In
Kevin's current code, after checking whether a RangeVar might be a CTE
reference and before deciding that it must be a table reference,
scanNameSpaceForTsr() is used to check whether there's a tuplestore by
that name and, if so, then we insert a RTE with type RTE_TUPLESTORE
which references the tuplestore by name. To me, the obvious thing to
do here seems to be to instead call p_tableref_hook and let it return
a suitable RangeTblRef (or NULL if it wishes to take no action). In
the case where the hook wants to redirect the use of that name to a
tuplestore, it can add a range-table entry of type RTE_TUPLESTORE, and
that entry can store a parameter-index rather than, as in the current
design, a name. But then where does Heikki's notion of a
RelationParam get used?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company