Add itemType constraint to array type

[akariv]

This attribute will allow specifying the exact items that are allowed in an array.

e.g.

{
   "name": "years",
   "type": "array",
   "constraints": {
      "typeOf": "year"
  }
}

Related

Doing this with appropriate semantics would also resolve

• Enum constraint with arrays in table schema Enum constraint with arrays in table schema #549

[rufuspollock]

@akariv really like it. Again may be v1.1

[danfowler]

Related: #381

[pwalsh]

+1 on this, and we should do a PR once v1 settles.

[rufuspollock]

@akariv would you be up for doing a PR to implement this?

I also think that rather than just have the type we might as well go the whole hog and just have the whole field schema be reused for the individual item. In this case the property name might be something like an itemSchema or itemType.

Alternatively, since most of the constraints in constraints only make sense for individual values we could implicitly apply them to the individual values in the array e.g. max, min, enum etc in which case we do only need typeOf.

Thoughts @roll ...

[akariv]

@rufuspollock that's my suggestion in #410

[rufuspollock]

Isn't the difference that for array all entries are implicitly of the same type?

…

On Thu, Apr 30, 2020 at 9:44 PM Adam Kariv ***@***.***> wrote: @rufuspollock <https://github.com/rufuspollock> that's my suggestion in #410 <#410> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#409 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABMDMWJTBM4NABT36V6N53RPHIKXANCNFSM4DJSDHSQ> .

[akariv]

table type could be seen as syntactic sugar for an array field of itenType=object and a mandatory schema property.

In both cases the datatype is the same in all array entries.

[roll]

I think itemType should be as simple as possible thing. I would move all the complexity to something like #410 and v2 of the specs

[rufuspollock]

@roll generally agree - though could you clarify what "as simple as possible" looks like for you - and would it address the related stuff e.g. #549?

[roll]

@rufuspollock

I think there are two options

Simple

https://specs.frictionlessdata.io/table-schema/#array

array
The field contains data that is a valid JSON format arrays.

Because it's a JSON we can probably treat items just as native JSON values (itemType):

• string
• number
• boolean
• null

It will mean no processing on the TableSchema level but just checking that a JSON has items with the same given type e.g. [1,2,3] or [true,false] - just re-using native JSON types.

We can also add a itemConstraints property.

This option will keep things simple.

Complex

Another option is to have full itemSchema or something like as you mentioned. This adds Table Schema level processing to array's items e.g. an ability to extract parsed vales from items as strings ["10$", "30$"] or ["2020-02-01", "2020-04-05"].

With the second option, we add an ability to process array's items with all the power of Table Schema.

This option is much more complex though.

[akariv]

I think that `itemType` should have the same semantics as `type` (i.e.read a tableschema type name and not a json). Array items would then be considered as physical representation and would be processed to extract the logical value. But then, we would need also `itemConstraints`, `itemFormat`, `itemDecimalChar` and there's no end to this. Therefore, a different proposal for this would be to create a single `arrayItem` configuration, which is parsed exactly as a tableschema field. Like so: ``` ... { "name": "futurePaymentDates", "type": "array", "constraints": { "maxLength": 12 }, "arrayItem": { "type": "date", "format": "%Y/%m/%d" } } ``` This also means that `schema`, as a field specification, might be added to the spec as an optional parameter of an `object` field - but is unrelated to the `array` item issue.

[roll]

But then, we would need also itemConstraints, itemFormat,
itemDecimalChar and there's no end to this.

Yea exactly, without all the options a Table Schema type is partially useless.

I think that itemType should have the same semantics as type (i.e.read a tableschema
type name and not a json).

It can be itemJsonType. It will

• solve "80%" of use cases
• keep array very simple
• give an ability to design a proper nested type in e.g. Create new type table #410

I'm not sure that it's the best option just trying to figure out the pros.

[JDziurlaj]

There is something elegant in reusing the existing schema structures when adding this feature. I don't think constraining the arrays to simple JSON types would provide a lot of value. I favor @akariv's proposal.

[rufuspollock]

OK, it looks like we have convergence here.

[pwalsh]

@roll @rufuspollock

I see frictionlessdata/frictionless-py#627 has been referenced here but it solves a fundamentally different issue.

It would be great to see this original request added. It is very common that members of an array be same typed, and declaring it as part of table schema would allow:

• In the SQL driver, use of Array fields for backends that support it (which would be a more logically correct mapping than JSONB fields, and allow round-tripping data)
• In the Elasticsearch driver, use of array fields
• Most likely most other data backends that support an array as a type would require members to have the same type

Note that this behavior ( also #410 ) is already implemented in some form in https://github.com/frictionlessdata/tableschema-elasticsearch-py and has a bunch of successful usage in production systems.

[rufuspollock]

Happy to have PRs on this - or even start with a pattern and implement into say python. Remember the complex thing with the specs now is that one then needs all drivers to upgrade to be compliant.