The thing is that with 60m of unique values (which are hashes anyway) the grouping may not scale well. I've been doing tests on a single machine so far (32 cores, 64GB RAM) and the COUNT DISTINCT query wouldn't complete (while other queries gave very encouraging results). So my idea is to facilitate the run-time computation by doing so offline computation in the ETL job. Initially, I thought that a custom UDAF could implement this functionality, but I got the impression after going through the source code that it won't do. I may be wrong with my assumptions here, we're going to do testing on a 10-node cluster soon but my understanding is that grouping by so many distinct values is not really scalable. I'd be happy to be proven wrong :)
On Tue, Apr 7, 2015 at 5:11 PM, Aman Sinha <asi...@maprtech.com> wrote: > Drill already does most of this type of transformation. If you do an > 'EXPLAIN PLAN FOR <your count(distinct) query>' > you will see that it first does a grouping on the column and then applies > the COUNT(column). The first level grouping can be done either based on > sorting or hashing and this is configurable through a system option. > > Aman > > On Tue, Apr 7, 2015 at 3:30 AM, Marcin Karpinski <mkarpin...@opera.com> > wrote: > > > Hi Guys, > > > > I have a specific use case for Drill, in which I'd like to be able to > count > > unique values in columns with tens millions of distinct values. The COUNT > > DISTINCT method, unfortunately, does not scale both time- and memory-wise > > and the idea is to sort the data beforehand by the values of that column > > (let's call it ID), to have the row groups split at new a new ID boundary > > and to extend Drill with an alternative version of COUNT that would > simply > > count the number of times the ID changes through out the entire table. > This > > way, we could expect that counting unique values of pre-sorted columns > > could have complexity comparable to that of the regular COUNT operator (a > > full scan). So, to sum up, I have three questions: > > > > 1. Can such a scenario be realized in Drill? > > 2. Can it be done in a modular way (eg, a dedicated UDAF or an operator), > > so without heavy hacking throughout entire Drill? > > 3. How to do it? > > > > Our initial experience with Drill was very good - it's an excellent tool. > > But in order to be able to adopt it, we need to sort out this one central > > issue. > > > > Cheers, > > Marcin > > >