Faster computation of quantiles in describe (#2909)

Computing all quantiles when we only need the median is significantly slower.
Also avoid trying to compute quantiles for columns that do not support basic arithmetic
since the failure only happens after sorting the vector, which is almost all of the work.

Author: nalimilan

src/abstractdataframe/abstractdataframe.jl

@@ -703,10 +703,23 @@ function get_stats(@nospecialize(col::Union{AbstractVector, Base.SkipMissing}),
     d = Dict{Symbol, Any}()
 
     if :q25 in stats || :median in stats || :q75 in stats
-        q = try quantile(col, [.25, .5, .75]) catch; (nothing, nothing, nothing) end
-        d[:q25] = q[1]
-        d[:median] = q[2]
-        d[:q75] = q[3]
+        # types that do not support basic arithmetic (like strings) will only fail
+        # after sorting the data, so check this beforehand to fail early
+        T = eltype(col)
+        if isconcretetype(T) && !hasmethod(-, Tuple{T, T})
+            d[:q25] = d[:median] = d[:q75] = nothing
+        else
+            mcol = Base.copymutable(col)
+            if :q25 in stats
+                d[:q25] = try quantile!(mcol, 0.25) catch; nothing; end
+            end
+            if :median in stats
+                d[:median] = try quantile!(mcol, 0.50) catch; nothing; end
+            end
+            if :q75 in stats
+                d[:q75] = try quantile!(mcol, 0.75) catch; nothing; end
+            end
+        end
     end
 
     if :min in stats || :max in stats

test/dataframe.jl

@@ -668,7 +668,7 @@ end
                                        nothing, nothing, nothing],
                                 min = [1.0, 1.0, "a", "a", Date(2000), 1],
                                 q25 = [1.75, 1.5, nothing, nothing, nothing, nothing],
-                                median = [2.5, 2.0, nothing, nothing, nothing, nothing],
+                                median = [2.5, 2.0, nothing, nothing, VERSION >= v"1.7.0-beta1.2" ? Date(2002) : nothing, nothing],
                                 q75 = [3.25, 2.5, nothing, nothing, nothing, nothing],
                                 max = [4.0, 3.0, "d", "c", Date(2004), 2],
                                 nunique = [nothing, nothing, 4, 3, 4, 2],