Improve Tables.istable by adding Tables.astable

[bkamins]

@quinnj The issue that Tables.istable can return false and the object is still a table resurfaces often. See e.g. JuliaEarth/geospatial-data-science-with-julia#3 from today.

What I propose:

1. allow packages to assume that only something for which Tables.istable returns true is a table; (stricter rule)
2. still other packages can take anything to be a table; (loose rule)

For packages that want to follow the stricter rule it would be nice to add a method to e.g. Tables.table:

Tables.astable(t) = GenericTable(t)

where GenericTable would be a thin wrapper for which Tables.istable would return true, and all other methods would just unwrap t from GenericTable and call it.

In this way packages following stricter rule could easily distinguish between general objects that are not tables and something that user explicitly asked to be a table.

[ronisbr]

I fully agree with the proposal of @bkamins ! I would just change how we call those scenarios.

In my opinion, I would call an object that returns Tables.istable == true something that follows Tables.jl public API and not just a table. Because we can have a table (in the generic sense) that does not completely implement Tables.jl API but can be used as tabular data.

Hence, packages that want to follow the loose rule, can use the Tables.jl functions to fetch the data, even though the object does not comply with the Tables.jl public API. Packages that want to follow the stricter rule, just need to mention that they are expecting something compliant with the Tables.jl public API.

Does it make sense or is it even more complicated?

[ronisbr]

One good example pointed out by @bkamins is the vector of Dict. It does not comply with Tables.jl API. Hence, PrettyTables.jl, which expects it, prints them as a vector:

julia> pretty_table([Dict(1=>2), Dict(3=>4)])
┌────────────┐
│     Col. 1 │
├────────────┤
│ Dict(1=>2) │
│ Dict(3=>4) │
└────────────┘

DataFrames, on the other hand, parse the data as follows:

julia> DataFrame([Dict(1=>2), Dict(3=>4)])
2×8 DataFrame
 Row │ slots                              keys                               vals                               ndel   count  age     idxfloor  maxprobe
     │ Array…                             Array…                             Array…                             Int64  Int64  UInt64  Int64     Int64
─────┼───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
   1 │ UInt8[0x00, 0x00, 0x00, 0x00, 0x…  [0, 4585988768, 5056425184, 4585…  [0, 5776188496, 5056425184, 4585…      0      1       1        15         0
   2 │ UInt8[0x00, 0x00, 0x00, 0x00, 0x…  [0, 4585988768, 5056425184, 4585…  [0, 4585988768, 5056425184, 4585…      0      1       1        14         0

[bkamins]

Does it make sense or is it even more complicated?

I think, if we add Tables.astable, we do not even need to use name for the behavior. Packages should just announce that they either:

• only treat as tables objects that produce true from Tables.istable (and thus require use of Tables.astable for objects that do not fully implement Tables.jl API). PrettyTables.jl is an example of such a package.
• or assume that anything can be potentially a table, regardless of what Tables.istable returns. DataFrames.jl is such a package (and probably will stay so, because changing this would be breaking)

[ronisbr]

Perfect! I think it will make easier to handle those types internally in PrettyTables.jl since now I need to check for custom table-like types if Tables.istable is false (vectors, matrices, and dictionaries).

[ronisbr]

Btw, from PrettyTables.jl side, I think that with @bkamins proposal we will not break depending on how Tables.jl would treat vectors, matrices, and dictionaries after Tables.astable. However, if it breaks, I am fine to release a new major version. There are some breaking changes I would want to implement anyway.

[bkamins]

There are matrices that are tables, but AFAICT all such types implement Tables.istable, so things should be fine.