Once we surveyed the market, we noticed the necessity for an answer that would carry out quick SQL queries on fluid JSON information, together with arrays and nested objects:
The Problem of SQL on JSON
Some type of ETL to rework JSON to tables in SQL databases could also be workable for primary JSON information with mounted fields which are identified up entrance. Nevertheless, JSON with nested objects or new fields that “can spring up each 2-4 weeks,” as the unique Stack Overflow poster put it, is not possible to deal with in such a inflexible method.
Relational databases supply different approaches to accommodate extra complicated JSON information. SQL Server shops JSON in varchar columns, whereas Postgres and MySQL have JSON information sorts. In these situations, customers can ingest JSON information with out conversion to SQL fields, however take a efficiency hit when querying the info as a result of these columns assist minimal indexing at finest.
SQL on Nested JSON Utilizing Rockset
With plenty of fields that change, get added/eliminated, and so on, it may be reasonably cumbersome to take care of ETL pipelines. Rockset was designed to assist with this drawback—by indexing all fields in JSON paperwork, together with all sort info, and exposing a SQL API on prime of it.
For instance, with a Rockset assortment named new_collection, I can begin by including a single doc to an empty assortment that appears like:
{
"my-field": "doc1",
"my-other-field": "some textual content"
}
… after which question it.
rockset> choose "my-field", "my-other-field"
from new_collection;
+------------+------------------+
| my-field | my-other-field |
|------------+------------------|
| doc1 | some textual content |
+------------+------------------+
Now, if a brand new JSON doc is available in with some new fields – perhaps with some arrays, nested JSON objects, and so on, I can nonetheless question it with SQL.
{
"my-field": "doc2",
"my-other-field":[
{
"c1": "this",
"c2": "field",
"c3": "has",
"c4": "changed"
}
]
}
I add that to the identical assortment and may question it simply as earlier than.
rockset> choose "my-field", "my-other-field"
from new_collection;
+------------+---------------------------------------------------------------+
| my-field | my-other-field |
|------------+---------------------------------------------------------------|
| doc1 | some textual content |
| doc2 | [{'c1': 'this', 'c2': 'field', 'c3': 'has', 'c4': 'changed'}] |
+------------+---------------------------------------------------------------+
I can additional flatten nested JSON objects and array fields at question time and assemble the desk I wish to get to – with out having to do any transformations beforehand.
rockset> choose mof.*
from new_collection, unnest(new_collection."my-other-field") as mof;
+------+-------+------+---------+
| c1 | c2 | c3 | c4 |
|------+-------+------+---------|
| this | area | has | modified |
+------+-------+------+---------+
Along with this, there’s robust sort info saved, which implies I will not get tripped up by having blended sorts, and so on. Including a 3rd doc:
{
"my-field": "doc3",
"my-other-field":[
{
"c1": "unexpected",
"c2": 99,
"c3": 100,
"c4": 101
}
]
}
It nonetheless provides my doc as anticipated.
rockset> choose mof.*
from new_collection, unnest(new_collection."my-other-field") as mof;
+------------+-------+------+---------+
| c1 | c2 | c3 | c4 |
|------------+-------+------+---------|
| sudden | 99 | 100 | 101 |
| this | area | has | modified |
+------------+-------+------+---------+
… and the fields are strongly typed.
rockset> choose typeof(mof.c2)
from new_collection, unnest(new_collection."my-other-field") as mof;
+-----------+
| ?typeof |
|-----------|
| int |
| string |
+-----------+
If with the ability to run SQL on complicated JSON, with none ETL, information pipelines, or mounted schema, sounds fascinating to you, it’s best to give Rockset a strive.
