Julia: use interned strings

[bkamins]

This change switches to use interned objects instead of standard strings (as standard strings are not interned) in joins.

@jangorecki - would it be possible to run the benchmarks with this change on your machine to see the impact? I changed it only for joins as it seems to be the most interesting case currently.

Thank you!

[bkamins]

cc @nalimilan @quinnj

[jangorecki]

Thanks, started juliadf now, only join task, so should be more early. From brief look, J1_1e7_NA_5_0 has exactly the same total timing 18s.

[bkamins]

Thank you. This change should only affect GC time. We have not implemented the changes in CSV.jl yet, but this is the current standard approach to reduce GC time.

[jangorecki]

taken from ./_utils/time.R

> tail.time("juliadf", "join", i=c(1L, 2L))
    in_rows               knasorted               question 20210515_66ee4fe 20210603_66ee4fe
 1:     1e7   0% NAs, unsorted data     small inner on int            0.751            0.797
 2:     1e7   0% NAs, unsorted data    medium inner on int            0.399            0.419
 3:     1e7   0% NAs, unsorted data    medium outer on int            1.965            1.890
 4:     1e7   0% NAs, unsorted data medium inner on factor            0.578            0.737
 5:     1e7   0% NAs, unsorted data       big inner on int            1.377            1.395
 6:     1e7   5% NAs, unsorted data     small inner on int            0.863            0.903
 7:     1e7   5% NAs, unsorted data    medium inner on int            0.703            0.754
 8:     1e7   5% NAs, unsorted data    medium outer on int            2.076            2.271
 9:     1e7   5% NAs, unsorted data medium inner on factor            0.741            0.746
10:     1e7   5% NAs, unsorted data       big inner on int            3.169            3.018
11:     1e7 0% NAs, pre-sorted data     small inner on int            0.662            0.643
12:     1e7 0% NAs, pre-sorted data    medium inner on int            0.412            0.410
13:     1e7 0% NAs, pre-sorted data    medium outer on int            1.878            1.834
14:     1e7 0% NAs, pre-sorted data medium inner on factor            0.459            0.600
15:     1e7 0% NAs, pre-sorted data       big inner on int            0.623            0.583
16:     1e8   0% NAs, unsorted data     small inner on int           31.547           37.773
17:     1e8   0% NAs, unsorted data    medium inner on int           27.440           20.118
18:     1e8   0% NAs, unsorted data    medium outer on int           58.285           53.417
19:     1e8   0% NAs, unsorted data medium inner on factor           37.202           31.975
20:     1e8   0% NAs, unsorted data       big inner on int           50.273           45.615
21:     1e8   5% NAs, unsorted data     small inner on int           37.778           48.157
22:     1e8   5% NAs, unsorted data    medium inner on int           28.907           28.997
23:     1e8   5% NAs, unsorted data    medium outer on int           47.130           62.510
24:     1e8   5% NAs, unsorted data medium inner on factor           37.742           47.860
25:     1e8   5% NAs, unsorted data       big inner on int           69.493           61.993
26:     1e8 0% NAs, pre-sorted data     small inner on int           29.921           31.294
27:     1e8 0% NAs, pre-sorted data    medium inner on int           25.802           21.033
28:     1e8 0% NAs, pre-sorted data    medium outer on int           63.371           39.580
29:     1e8 0% NAs, pre-sorted data medium inner on factor           35.144           28.662
30:     1e8 0% NAs, pre-sorted data       big inner on int           43.911           31.395

[bkamins]

Thank you for theis. In 1e7 case as expected there is no change (as GC is not triggered then anyway).
We are a bit better in 1e8 case, as we spend a bit less time in GC (but I expected that we would still be affected by GC, see JuliaLang/julia#40840 (comment) if you are interested in the reason).

@quinnj is finishing a PR to CSV.jl that should resolve GC issues without requiring changes in Julia Base. Then I will open another PR with a new setting of reading data in.

Thank you!