Hi,
while working on the bloom filters for hashjoins, I've started using the
HLL library committed as part of the sorting improvements for 9.5. I
propose adding two more functions to the API, which I think are quite
useful:
1) initHyperLogLogError(hyperLogLogState *cState, double error)
Instead of specifying bwidth (essentially the number of bits used for addressing in the counter), this allows
specifyingthe expected error rate for the counter, which is
error_rate = 1.04 / sqrt(2^bwidth)
So for 5% we get bwidth=5, and so on. This makes the API a bit easier the use, because there are pretty much no
commentsabout the meaning of bwidth, and the existing callers simply use 10 without details.
2) freeHyperLogLog(hyperLogLogState *cState)
I think it's a good idea to provide function "undoing" what init does, i.e. freeing the internal memory etc.
Currentlythat's trivial to do, but perhaps we'll make the structure more complicated in the future (albeit that
mightbe unlikely).
FWIW I've already used this in the patch marrying hash joins and bloom
filters.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services