parse with Missing

[bkamins]

I would like to put under consideration if we want to add methods to parse that accept Union{T, Missing} argument (or create a similar method). This would be something like tryparse but returning missing on failure.

I would think of something like:

• parse(Union{Int, Missing}, "123") == 123;
• parse(Union{Int, Missing}, "123a") == missing.

[nalimilan]

Actually with JuliaLang/julia#23642 tryparse will have return type Union{Some{T}, Void} (or maybe just Union{T, Void} depending on the outcome of discussions). I'm not sure we want to allow parse(Union{T, Missing}, ...) in addition to that. There have been lots of discussions about this kind of syntax around the introduction of tryparse in Base. One issue is that parse(Union{T, Missing}, x) could be syntax to parse "missing" as missing, numbers as Int, and throw and error in other cases.

[bkamins]

Close this issue then or keep for a reference?

[nalimilan]

I'd say it depends on whether you think this syntax is needed or not.

[bkamins]

I believe we need a function that returns missing if parsing fails. It does not have to be prase, actually tryparse is more natural, but now it returns Nullable (which behavior I earlier thought would be retained as computer scientist missingness).

So let me comment on a use case and you can decide which way to go (I have much less understanding of the whole ecosystem context). A typical scenario with DataFrames is to read some data which is messy but is supposed to contain integers. Assume that we already have it as ["1", "2", "?", "", "3", "unknown"].

A pure approach would be to handle the missing data carefully, but a practical workflow (and this is how R works) is to parse it as Vector{Int} and make all that is not an integer a missing.

Computer scientist now can do:

julia> tryparse.(Int, ["1", "2", "?", "", "3", "unknown"])
6-element Array{Nullable{Int64},1}:
 1
 2
 #NULL
 #NULL
 3
 #NULL

But data scientist has to do (or there is some simpler way?):

julia> [isnull(x) ? missing : get(x) for x in tryparse.(Int, ["1", "2", "?", "", "3", "unknown"])]
6-element Array{Any,1}:
 1
 2
  missing
  missing
 3
  missing

which not only is messy but also has wrong type (Any) of the produced array.

I would like to have this fixed (I understand that solving Any problem is an additional topic discussed in JuliaLang/julia#24332).

[nalimilan]

I see your point. I'm not sure we want to follow R in all aspects, in particular we have the ability to distinguish an invalid value from a missing value, so it wouldn't be absurd to be stricter, and require an explicit operation to convert an unparsable value to a missing value. That could be done quite easily using replace (see JuliaLang/julia#22324). Adding a function as you suggest also makes sense, but I'd rather wait for the design of tryparse to settle first (JuliaLang/julia#23642), we can easily extend it later.

[bkamins]

Thanks.

[alyst]

I also don't think parse() should return missing for invalid inputs.

But somewhat more conservative thing would be to allow

parse(Int, missing) = missing
#= or =# 
parse(Union{Int, Missing}, missing) = missing

I prefer the first alternative as the 2nd might imply: parse(Union{Int, Missing}, "missing") == missing.

Then it would be possible to parse.(Int, ::AbstractArray{Union{String, Missing}})

[quinnj]

The approach taken in CSV.jl, which @shashi and I have discussed generalizing to parsing functions in general is to have parse(Union{Int, Missing}, str) => Union{Int, Missing} and parse(Int, "missing") => throw(ParsingException()).

[alyst]

@quinnj Do you also throw ParsingException() on parse(Int, ::Missing)?

[quinnj]

I don't currently have a use-case for that in CSV as there's always guaranteed to be an input, though the input may be an empty string.

[nalimilan]

So you are using missing to indicate that parsing failed? Basically what tryparse does with Nullable() (and will soon do with nothing)? I think it'd prefer using parse(Union{Int, Void}, ...) for that, to distinguish clearly between "invalid" and "missing".

[quinnj]

@nalimilan was that in response to my comment or @alyst ?

[nalimilan]

To yours.

[quinnj]

No, the user has to provide, in the case of CSV, a "null" string and if the input matches this "null" string, then missing is returned. In the case of failure, e.g. non-numeric digits when parsing a number, an error is always thrown. That leaves try-catch behavior up to the caller, who would know what to do in case of failure anyway. I supposed you could keep a tryparse function that returned nothing instead of throwing.

[pdeffebach]

A user ran into this with Dates recently, where Dates throws an error on parsing missings.

JuliaLang/julia#28570

I would recommend a keyword argument, but if a constructor relies on parsing, we have to trust that they forward keyword arguments as well.