improve group performance

[aplavin]

create 1-vector instead of creating empty vector and pushing to it

[aplavin]

Failures on 1.0 seem totally unrelated.

[andyferris]

Yes, unfortunately the push! route can be a little bit slow (last I checked there should be micro-optimization opportunates for Julia there). Have you got a benchmark for these improvements @aplavin?

I find that injecting the type T is really useful for inferring cases like Union{Missing, X}.

[aplavin]

There are minor performance improvements for small groups, and no difference for large groups. Eg:

julia> A = 1:100
1:100

# before
julia> @btime groupview(a -> a % 100, $A);
  8.927 μs (216 allocations: 23.52 KiB)

# after
julia> @btime groupview(a -> a % 100, $A);
  7.812 μs (116 allocations: 15.70 KiB)

Something along these lines #28 would make much more difference, I think.

[andyferris]

Awesome, that’s still a nice improvement.