groupby with dogroups (R expression) performance regression #4200
Description
There is a performance regression (AFAIU) when doing by group computation where we run R's C eval by each group (q7 and q8 in db-benchmark).
in_rows question_group question 20181206_da98fb2 20190913_35b0de3 20191115_92abb70 20191205_eba8704 20191209_6808d2c 20191212_e0140ea 20191229_d52b0d8 20200124_c005296
1: 1e9 basic sum v1 by id1 21.144 10.026 8.375 9.362 9.060 9.082 9.366 9.271
2: 1e9 basic sum v1 by id1:id2 38.914 11.746 9.243 9.327 9.331 9.978 10.813 9.220
3: 1e9 basic sum v1 mean v3 by id3 99.517 14.487 12.044 12.291 13.496 14.325 13.191 13.169
4: 1e9 basic mean v1:v3 by id4 26.593 17.357 15.135 15.157 15.278 16.761 16.724 16.754
5: 1e9 basic sum v1:v3 by id6 122.214 14.569 13.454 14.035 14.046 14.842 14.400 15.019
6: 1e9 advanced median v3 sd v3 by id4 id5 NA NA 121.742 110.925 106.340 113.837 111.984 123.411
7: 1e9 advanced max v1 - min v2 by id3 NA NA 98.680 91.596 87.005 93.749 91.294 299.863
8: 1e9 advanced largest two v3 by id6 NA NA 234.926 215.574 213.824 216.152 211.295 411.241
9: 1e9 advanced regression v1 v2 by id2 id4 NA 72.466 81.297 76.121 75.311 74.769 72.571 41.157
10: 1e9 advanced sum v3 count by id1:id6 NA 180.257 196.403 187.816 177.257 187.702 190.282 187.403
worth to note that at the same time q9 x[, .(r2=cor(v1, v2)^2), by=.(id2, id4)] got nice speed up