Flatterer. Making JSON flatterer

Introduction

An opinionated JSON to CSV/XLSX/SQLITE/PARQUET/POSTGRES converter which tries to make a useful relational output for data analysis.

Rationale

When receiving a JSON file where the structure is deeply nested or not well specified, it is hard to determine what the data contains. Also, even after knowing the JSON structure, it requires a lot of time to work out how to flatten the JSON into a relational structure to do data analysis on and to be part of a data pipeline.

Flatterer aims to be the first tool to go to when faced with the above problem. It may not be the tool that you end up using to flatten the JSON in your data pipeline, as hand written flattening may be required, but it could be. It has many benefits over most hand written approaches:

• It is fast, written in rust but with python bindings for ease of use. It can be 10x faster than hand written python flattening.

• Memory efficient. Uses a custom streaming JSON parser to mean that long list of objects nested with the JSON will be streamed, so not much data needs to be loaded into memory at once.

• Fast memory efficient output to CSV/XLSX/SQLITE/PARQUET/POSTGRES

• Uses best practice that has been learnt from flattening JSON countless times, such as generating keys to link one-to-many tables to their parents.

• S3 Support

Install

pip install flatterer

Flatterer requires Python 3.6 or greater. It is written as a python extension in Rust but has binaries (wheels) for linux (x64 anylinux), macos (x64 and universal) and windows (x64, x86). On other platforms a rust toolchain will need to be installed.

Example JSON

Say you have a JSON data like this named games.json:

[
  {
    "id": 1,
    "title": "A Game",
    "releaseDate": "2015-01-01",
    "platforms": [
      {"name":"Xbox"},
      {"name":"Playstation"}
    ],
    "rating": {
      "code": "E",
      "name": "Everyone"
    }
  },
  {
    "id": 2,
    "title": "B Game",
    "releaseDate": "2016-01-01",
    "platforms": [
      {"name":"PC"}
    ],
    "rating": {
      "code": "E",
      "name": "Everyone"
    }
  }
]

Running Flatterer

CLI

Run the above file with flatterer.

flatterer games.json games_dir

See Option Reference for details of additional command line options, outputs for information on the output formats and inputs for more information on input sources.

Local web interface

Local web interface to explore flatterer features.

flatterer --web

See Web Interface for details of how to configure web interface.

As python library

import flatterer
output = flatterer.flatten('games.json', 'games_dir')

See Python Library for more details.

Output Files

By running the above you will get the following files:

tree games_dir

games_dir/
├── csv
│   ├── games.csv
│   └── platforms.csv
├── datapackage.json
├── fields.csv
└── ...

Main Table

games.csv contains:

_link	id	rating_code	rating_name	releaseDate	title
0	1	E	Everyone	2015-01-01	A Game
1	2	E	Everyone	2016-01-01	B Game

Special column _link is generated. _link is the primary key there unique per game.

Also the rating sub-object is promoted to this table it has a one-to-one relationship with games. Sub-object properties are separated by ‘_’.

One To Many Table

platforms is an array so is a one-to-many with games therefore needs its own table:

platforms.csv contains:

_link	_link_games	name
0.platforms.0	0	Xbox
0.platforms.1	0	Playstation
1.platforms.0	1	PC

Link Fields

_link is the primary key for the platforms table too. Every table except games table, contains a _link_games field to easily join to the main games table.

If there was a sub-array of platforms then that would have _link, _link_games and _link_platforms fields.

To generalize this the _link__<table_name> fields joins to the _link field of <table_name> i.e the _link__<table_name> are the foreign keys referencing <table_name>._link.

Fields CSV

fields.csv contains some metadata about the output tables:

table_name	field_name	field_type	count	field_title
platforms	_link	text	3	_link
platforms	_link_games	text	3	_link_games
platforms	name	text	3	name
games	_link	text	2	_link
games	id	number	2	id
games	rating_code	text	2	rating_code
games	rating_name	text	2	rating_name
games	releaseDate	date	2	releaseDate
games	title	text	2	title

The field_type column contains a type guess useful for inserting into a database. The field_title is the column heading in the CSV file or XLSX sheet, which is initially the same as the field_name. After editing this file then you can rerun the transform:

flatterer games.json new_games_dir -f myfields.csv --only-fields

This can be useful for renameing columns, rearranging the field order or if you want to remove some fields the --only-fields flag will only include the fields in the edited file.

datapackage.json contains metadata in the Tabular Datapackge Spec