Today, Impala does not evaluate "<col> != <constant>" against stats, but as Zoltan pointed out there is a way to reasonably do that. It does not work if we ignore NaN though, so we need to be careful.
On Tue, Feb 20, 2018 at 9:24 AM, Zoltan Ivanfi <z...@cloudera.com> wrote: > In parquet-mr, if you are looking for a value that is not equal to some > reference value r and stats are min = r and max = r then that row group is > discarded, because there can not be any other values in that row group. > > On Tue, Feb 20, 2018 at 6:21 PM Jim Apple <jbap...@cloudera.com> wrote: > > > For that predicate in particular, does Impala use stats already? > > > > Let's say a column contains only the intuitive notion of floats: no > > NaNs, no infs, no -0.0. If we are filtering for $COL != a and the > > row-group stats are b <= $COL <= c, were a < b, we can know that the > > whole row group can be included. The addition of NaNs doesn't change > > that. > > > > OTOH, if b <= a <= c, then we have to check the whole row group, and > > the addition of NaNs doesn't change that. > > > > On Tue, Feb 20, 2018 at 9:14 AM, Alexander Behm <alex.b...@cloudera.com> > > wrote: > > > On Mon, Feb 19, 2018 at 8:04 AM, Zoltan Ivanfi <z...@cloudera.com> > wrote: > > > > > >> Hi, > > >> > > >> Tim, I added your suggestion to introduce a new ColumnOrder to > > PARQUET-1222 > > >> <https://issues.apache.org/jira/browse/PARQUET-1222> as the preferred > > >> solution. > > >> > > >> Alex, not writing min/max if there is a NaN is indeed a feasible > > quick-fix, > > >> but I think it would be better to just ignore NaN-s for the pruposes > of > > >> min/max stats. For reading, we can ignore stats that contain a NaN. We > > also > > >> shouldn't use stats when looking for a NaN. -0 and +0 will still be > > >> problematic, though. > > >> > > > > > > I don't think ignoring NaNs is correct. Consider a predicate <col> != > > > <constant> that would evaluate to true against NaN. We cannot reliable > > use > > > stats for such a predicate. > > > > > > > > >> > > >> Jim, fmax is indeed very close to IEEE-754's maxNum, but -0 and +0 are > > >> implementation-dependent, az Zoltan Borok-Nagy pointed it out to me: > > "This > > >> function is not required to be sensitive to the sign of zero, although > > some > > >> implementations additionally enforce that if one argument is +0 and > the > > >> other is -0, then +0 is returned." [1 > > >> <http://en.cppreference.com/w/c/numeric/math/fmax>] > > >> > > >> Br, > > >> > > >> Zoltan > > >> > > >> > > >> > > >> On Fri, Feb 16, 2018 at 6:57 PM Jim Apple <jbap...@cloudera.com> > wrote: > > >> > > >> > On Fri, Feb 16, 2018 at 9:44 AM, Zoltan Borok-Nagy > > >> > <borokna...@cloudera.com> wrote: > > >> > > I would just like to mention that the fmax() / fmin() functions in > > >> C/C++ > > >> > > Math library follow the aforementioned IEEE 754-2008 min and max > > >> > > specification: > > >> > > http://en.cppreference.com/w/c/numeric/math/fmax > > >> > > > > >> > > I think this behavior is also the most intuitive and useful > > regarding > > >> to > > >> > > statistics. If we want to select the max value, I think it's > > reasonable > > >> > to > > >> > > ignore nulls and not-numbers. > > >> > > > >> > It should be noted that this is different than the total ordering > > >> > predicate. With that predicate, -NaN < -inf < negative numbers < > -0.0 > > >> > < +0.0 < positive numbers < +inf < +NaN > > >> > > > >> > fmax appears to be closest to IEEE-754's maxNum, but not quite > > >> > matching for some corner cases (-0.0, signalling NaN), but I'm not > > >> > 100% sure on that. > > >> > > > >> > > >