Date: Jan 22, 2021 Version: 0.2.1

What is this?

This is Synth! A fast and highly configurable NoSQL synthetic data engine. It reconciles the two worlds of synthetic data and test data by letting users generate realistic synthetic data for testing their applications and ML models.

What can I do with this?

With Synth you can:

  • Anonymize sensitive data easily.

    As simple as JSON-in/JSON-out. If you’re not happy with the result, simply tweak the synthetic data model with a custom JSON metadata format and Synth will adjust everything on the fly, no additional ETL required.

  • Augment your datasets with synthetic data.

    For those times when you already have some data but just not enough of it to do what you need to do. It can extrapolate from patterns it finds in your data, so you can create as much of it as you want.

  • Create entirely new fake data declaratively.

    You can even add you own set of constraints and logic to create completely unseen scenario.

How does it work?

It has two components:

  • synthd: a persistent process that ingests raw (usually sensitive) training data and trains and builds synthetic data models from it. Think of it as a NoSQL datastore that never persists actual data, only anonymized model parameters.

  • synthpy: a reference Python implementation for the synthd API. This lets you leverage synthd in custom scripts and test harnesses.


Here is an end-to-end example using the Python client, synthpy.

from synthpy import Synth

# Assuming `synthd` is running on `localhost` with default settings
client = Synth("localhost:8182")

with open("my_users_data.json", "r") as data_f:
    documents = json.load(data_f)

# Submit your JSON documents to `synthd` for training
client.put_documents(namespace="app", collection="users", batch=documents)

# Generate 100 new synthetic users
synthetic_users = client.get_documents(namespace="app", collection="users", size=100)