min/max fails on ordered factors when using by

[mcieslik-mctp]

It seems as this is a regression in the latest version (data.table_1.9.8), as it worked before

library(data.table)
test <- data.table(V1=factor(rep(c("a","b"), 10), levels=c("a", "b"), ordered=TRUE), 
V2=rep(c("c","d", "e", "f"), 5))
test[,min(V1)]           # (1)
test[,min(V1),by=V2]     # (2)

(1) correctly works but (2) returns an error:
Error in gmin(V1) : min is not meaningful for factors.

[mattdowle]

(1) dispatches to base R. (2) dispatches to GForce grouping.
(You can pass verbose=TRUE to queries to get more insight.)

GForce was changed here (a1b1c08) but I don't see any notes in NEWS or any tests added.

R's min treats ordered factors and non-ordered factors differently, as you nicely showed.

R --vanilla
> x = factor(letters)
> min(x)
Error in Summary.factor(1:26, na.rm = FALSE) : 
  ‘min’ not meaningful for factors
> x = factor(letters, ordered=TRUE)
> min(x)
[1] a
26 Levels: a < b < c < d < e < f < g < h < i < j < k < l < m < n < o < ... < z

Yes I guess we should be in line with base R in this regard. Thanks for highlighting.

[mbacou]

Is there a status update about data.table's support for summarizing ordered factors? Won't be implemented? Thx.

[MichaelChirico]

@mbacou as luck would have it I think the fix was trivial (#2944). Thanks for the impetus!!

In any case, you can always set options(datatable.optimize = 0) to prevent GForce from failing on ordered factors, it'll just be slower.

[franknarf1]

SO q to update: https://stackoverflow.com/questions/50916153/in-r-data-table-max-works-on-an-ordered-factor-and-fails-when-grouped

[mbacou]

Excellent thanks! Was surprised to see that feature work well in e.g. PostgreSQL and not in data.table. Glad that was an easy fix.

[MichaelChirico]

if I would venture a guess it's only survived so long because ordered factors are almost criminally underused as a data type (myself included) :)

…

On Wed, Jun 20, 2018, 11:35 PM Mel. Bacou ***@***.***> wrote: Excellent thanks! Was surprised to see that feature work well in e.g. PostgreSQL and not in data.table. Glad that was an easy fix. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1947 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHQQdYWaa_HjNmI7O9h6v_LBT-125l3iks5t-mvIgaJpZM4LBgqS> .