Improve speed of aggregates that use array_append as transition function.
In the previous coding, if an aggregate's transition function returned an expanded array, nodeAgg.c and nodeWindowAgg.c would always copy it and thus force it into the flat representation. This led to ping-ponging between flat and expanded formats, which costs a lot. For an aggregate using array_append as transition function, I measured about a 15X slowdown compared to the pre-9.5 code, when working on simple int[] arrays. Of course, the old code was already O(N^2) in this usage due to copying flat arrays all the time, but it wasn't quite this inefficient. To fix, teach nodeAgg.c and nodeWindowAgg.c to allow expanded transition values without copying, so long as the transition function takes care to return the transition value already properly parented under the aggcontext. That puts a bit of extra responsibility on the transition function, but doing it this way allows us to not need any extra logic in the fast path of advance_transition_function (ie, with a pass-by-value transition value, or with a modified-in-place pass-by-reference value). We already know that that's a hot spot so I'm loath to add any cycles at all there. Also, while only array_append currently knows how to follow this convention, this solution allows other transition functions to opt-in without needing to have a whitelist in the core aggregation code. (The reason we would need a whitelist is that currently, if you pass a R/W expanded-object pointer to an arbitrary function, it's allowed to do anything with it including deleting it; that breaks the core agg code's assumption that it should free discarded values. Returning a value under aggcontext is the transition function's signal that it knows it is an aggregate transition function and will play nice. Possibly the API rules for expanded objects should be refined, but that would not be a back-patchable change.) With this fix, an aggregate using array_append is no longer O(N^2), so it's much faster than pre-9.5 code rather than much slower. It's still a bit slower than the bespoke infrastructure for array_agg, but the differential seems to be only about 10%-20% rather than orders of magnitude. Discussion: <6315.1477677885@sss.pgh.pa.us>
Showing
Please register or sign in to comment