I've been assisting Atri with development of an implementation of GROUPING SETS, beginning with a one-pass implementation of ROLLUP. Two issues have arisen regarding the API for calling aggregate transition and final functions that I think need answering now, since they relate to changes in 9.4.
1. Assumptions about the relationship between aggcontext and fn_extra Tom's version of the ordered-set aggregate code makes the assumption that it is safe to store the aggcontext returned by AggCheckCallContext in a structure hung off flinfo->fn_extra. This implicitly assumes that the function will be called in only one aggcontext, or at least only one per flinfo. Doing rollup via GroupAggregate by maintaining multiple transition values at a time (one per grouping set) means that the transfn is being called interleaved for transition values in different contexts. So the question becomes: is it wrong for the transition function to assume that aggcontext can be cached this way, or is it necessary for the executor to use a separate flinfo for each concurrent grouping set? 2. AggGetPerAggEContext The comment documents this as returning the per-output-tuple econtext, and the ordered-set code assumes that the rescan of this context implies that the aggcontext is also about to be reset (so cleanup functions can be called). Since it seems that the cleanup callback is the sole reason for this function to exist, is it acceptable to remove any implication that the context returned is the overall per-output-tuple one, and simply state that it is a context whose cleanup functions are called at the appropriate time before aggcontext is reset? (In fact what I'm considering is making aggcontext be the per-tuple memory context of an ExprContext created for each grouping set, and having AggGetPerAggEContext return this - but it won't be the actual context in which the result projection happens.) -- Andrew (irc:RhodiumToad) -- Sent via pgsql-hackers mailing list (firstname.lastname@example.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers