Proposal for TimeTable scitype

The simplest kind of time series data consists of a table in which one "timestamp" column is a vector with some temporal eltype, and the other columns can have any type (but are named). Some ML models would require this column to be temporally ordered, and possibly uniform, but defining scitypes which need to inspect internal structure is problematic, from a performance point-of-view (detecting Missing data is a case-in-point we already encountered).

Some specific tabular formats do ensure sequential order (eg, TimeSeries.TimesArray) but we should not like to exclude ordinary DataFrames, for example, from being counted as a "TimeTable". So, checking internal structure may have to be delegated to the client (eg, ML model).

With this in mind I propose we start by defining

struct TimeTable{K,T,timestamp} <: Table{Union{K,AbstractVector{T}}} end

where K is the union of the column scitypes (excluding the timestamp column), T is the elscitype of the timestamp column, and timestamp is the name of the timestamp column (a symbol). To realise the sub-typing, we would need to change Table from struct to abstract type but I'm not immediately seeing any problems with that.

So, for example, we could then implement scitype (in MLJScientificTypes) to have this behaviour:

using TimeSeries
using Dates
dates = Date(2018, 1, 1):Day(1):Date(2018, 12, 31)
ta = TimeArray(dates, rand(length(dates)))

scitype(ta) <: TimeSeriesTable{ScientificDateTime, AbstractVector{Continuous}, :timestamp}  # true

Thoughts?

cc: @svollmer

Start by reviewing the existing Table type and the scitype interface, then compare the proposal with the TimeSeries.TimeArray example and the referenced MLJScientificTypes behavior. Done would require an agreed design for TimeTable, including how Table subtyping and the timestamp column are represented.