map(f, t): create Columns when f returns tuple / named tuple

[shashi]

Turned out to be a bit complicated...

addresses #52

[codecov-io]

Codecov Report

Merging #54 into master will decrease coverage by 0.37%.
The diff coverage is 81.57%.

@@            Coverage Diff             @@
##           master      #54      +/-   ##
==========================================
- Coverage   93.49%   93.11%   -0.38%     
==========================================
  Files           6        6              
  Lines         922      959      +37     
==========================================
+ Hits          862      893      +31     
- Misses         60       66       +6

Impacted Files	Coverage Δ
src/IndexedTables.jl	91.48% <81.57%> (-3.71%)	⬇️
src/columns.jl	96.12% <0%> (+0.77%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 01dde0b...d2e7a4f. Read the comment docs.

[davidanthoff]

Should a map that returns a Pair{NamedTuple{...}, NamedTuple{...}} return an IndexedTable where the columns in the key of the pair make up the index, and the columns in the value make up the data columns? This would essentially be #55, but it would use a normal map as the API.

In some sense, all we would need is a constructor for IndexedTable that takes an iterator of that element type. That is something I wanted to add for some while: queryverse/IterableTables.jl#26

[shashi]

Should a map that returns a Pair{NamedTuple{...}, NamedTuple{...}} return an IndexedTable where the columns in the key of the pair make up the index, and the columns in the value make up the data columns?

I don't think so... map should only work on values. You could have output keys that collide allowing you to implement filter with map which seems wrong.

a constructor for IndexedTable that takes an iterator of that element type

Good idea to add this... One could even convert any Associative to an IndexedTable just by calling this constructor.

[shashi]

An alternative approach using Base.promote_op is in #57

[shashi]

t = IndexedTable([1,2,3], Columns(x=[4,5,6]))
y = [1, 1//2, "x"]
t4 = map(x->(x.x, y[x.x-3]), t)

This doesn't work with the alternative approach as promote_op has difficulty figuring out the output type. Creating a function with return type annotation also didn't work...

[shashi]

has difficulty figuring out the output type

Turns out it was throwing away the fact that the results are tuples when any field is UnionAll... Fixed in #57