drill-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Paul Rogers <prog...@mapr.com>
Subject Re: JSON reader enhancement
Date Mon, 20 Nov 2017 00:01:42 GMT
Hi Ted,

Thanks for the suggestions.

To handle nested lists correctly, we need Drill’s List data type, which uses Drill’s Union
data type. (The List type is really mostly just a repeated Union, and so needs union support.)
But, the union type disabled by default. The case I was trying to handle is to avoid an exception
when union type is disabled, but a 2D array appears. You make a good case that we should leave
the existing behavior; which I’ll do.

What you seem to be saying is that the Union type and the List type should be completed and
enabled by default. Then we need to add the missing functionality. 

I wonder, in a application, how would all of this be used? If Tableau is the primary client,
then data is delivered via ODBC. But, ODBC understands only ordinary rows and columns; not
the JSON types. Would you use Drill to convert the JSON structures into simple rows?

What kinds of transforms (functions) would be needed to handle 2D or higher arrays? To handle
heterogeneous arrays? To handle multiple list columns within the same record? Some time back
you talked about a correlated flatten in which two arrays can be flattened side-by-side. Any
other use cases?

Or, would the JSON structures be kept in tact, and the data, say, exported to other JSON files
using CTAS?

Thanks,

- Paul

> On Nov 19, 2017, at 1:42 PM, Ted Dunning <ted.dunning@gmail.com> wrote:
> 
> I don't see the value with this suggestion. It isn't going to make things
> much better since the user will be totally stunned when the structure
> doesn't come through as an array.
> 
> A bigger issue is the fact that elements of maps aren't marked correctly as
> missing. That means that if I have these two records:
> 
> "a":{"first": 1, "second": 2}, b:3
> "a"={"second":20, "third":30}, b:5
> 
> it is nearly impossible for me to determine whether "first" is missing from
> the second value of "a". This makes Drill impossible to use for lots of
> variant structure work. The values that Drill provides will have all fields
> marked as present.
> 
> (at least, this was true at one time).
> 
> 
> 
> 
> 
> On Sun, Nov 19, 2017 at 4:33 AM, Paul Rogers <progers@mapr.com> wrote:
> 
>> Hi Arina,
>> 
>> The proposal is to represent 2D arrays as a string (using the original,
>> unparsed JSON.) That is, given this input:
>> 
>> {a: “fred”, b: [[10, 20, 30], [11, 21, 31]]}
>> 
>> The parsed columns are:
>> 
>> a, b
>> “fred”, "[[10, 20, 30], [11, 21, 31]]”
>> 
>> Notice that column b is just a string. It is a string of JSON, yes, but
>> still just a string.
>> 
>> So, the question about kvgen/flatten does not apply here since we are not
>> creating a Drill array.
>> 
>> There is a very interesting discussion to be had about how Drill
>> does/should handle “non-relational” JSON structures. But, here, the
>> suggestions is just for one very simple special case.
>> 
>> Thanks,
>> 
>> - Paul
>> 
>>> On Nov 18, 2017, at 7:15 AM, Arina Yelchiyeva <
>> arina.yelchiyeva@gmail.com> wrote:
>>> 
>>> In general sounds good.
>>> If user will apply kvgen / flatten over such 2-D array columns read as
>>> string, he will be able to normalize data in the format he wants? Right?
>> Or
>>> we need to come up with new function?
>>> 
>>> Kind regards
>>> Arina
>>> 
>>> On Fri, Nov 17, 2017 at 10:39 PM, Paul Rogers <progers@mapr.com> wrote:
>>> 
>>>> Hi All,
>>>> 
>>>> I’d like to propose a minor enhancement to the JSON reader to better
>>>> handle non-relational JSON structures. (See DRILL-5974 [1].)
>>>> 
>>>> As background, Drill handles simple tuples:
>>>> 
>>>> {a: 10, b: “fred”}
>>>> 
>>>> Drill also handles arrays:
>>>> 
>>>> {name: “fred”, hobbies: [“bowling”, “golf”]}
>>>> 
>>>> Drill even handles arrays of tuples:
>>>> 
>>>> {name: “fred”, orders: [
>>>> {id: 1001, amount: 12.34},
>>>> {id: 1002, amount: 56.78}]}
>>>> 
>>>> The above are termed "relational" because there is a straightforward
>>>> mapping to/from tables into the above JSON structures.
>>>> 
>>>> Things get interesting with non-relational types, such as 2-D arrays:
>>>> 
>>>> {id: 4, shape: “square”, points: [[0, 0], [0, 5], [5, 0], [5, 5]]}
>>>> 
>>>> Drill has two solutions:
>>>> 
>>>> * Turn on the experimental list and union support.
>>>> * Enable all-text mode to read all fields as JSON text.
>>>> 
>>>> Here, I’d like to propose a middle ground:
>>>> 
>>>> * Read fields with relational types into vectors.
>>>> * Read non-relational fields using text mode.
>>>> 
>>>> Thus, the first three examples would all result in the JSON data parsed
>>>> into Drill vectors. But, the fourth, non-relational example would
>> produce a
>>>> row that looks like this:
>>>> 
>>>> id, shape, points
>>>> 4, “shape”, “[[0, 0], [0, 5], [5, 0], [5, 5]]”
>>>> 
>>>> Although Drill can’t parse the 2-D array, Drill will pass the array
>> along
>>>> to the client, which can use its favorite JSON parser to parse the array
>>>> and do something useful (like draw the square in this case.)
>>>> 
>>>> In particular, the proposal:
>>>> 
>>>> * Apply this change only to the revised “batch size aware” JSON reader.
>>>> * Use the above parsing model by default.
>>>> * Use the experimental list-and-union support if the existing
>>>> `exec.enable_union_type` system/session option is set.
>>>> 
>>>> Existing queries should “just work.” In fact, now JSON with
>> non-relational
>>>> types will work “out-of-the-box” without all-text mode or the
>> experimental
>>>> types.
>>>> 
>>>> Thoughts?
>>>> 
>>>> - Paul
>>>> 
>>>> [1] https://issues.apache.org/jira/browse/DRILL-5974
>>>> 
>>>> 
>>>> 
>> 
>> 

Mime
View raw message