chaoslake 0.1.1

Synthetic relational dataset generator with schema-driven chaos controls.

chaoslake

Generate synthetic relational datasets from a YAML schema in seconds.

chaoslake preserves foreign-key integrity across tables, supports statistical distributions, and can deliberately inject controlled data-quality issues (nulls, duplicates, drift, mixed date formats) — making it ideal for testing data pipelines, ML models, and analytics dashboards.

---

Install

pip install chaoslake

For the optional DuckDB backend and database introspection:

pip install "chaoslake[dev]"

---

Quick Start

1. Try it instantly — no schema needed

chaoslake quick

Generates a demo_output/ folder with a users table and a FK-linked transactions table straight away.

2. Create a schema interactively

chaoslake init

Asks whether you want to answer a short questionnaire (table name, row count, columns, chaos settings) or drop a ready-made template. Either way it writes chaoslake.yaml in the current directory.

3. Validate your schema

chaoslake check chaoslake.yaml

4. Generate the dataset

chaoslake generate --schema chaoslake.yaml --output ./output --format csv

Options:

Flag	Default	Description
--schema	chaoslake.yaml	Path to YAML schema
--output	./output	Output directory
--format	csv	csv or parquet
--seed	(none)	Integer seed for reproducible output
--engine	auto	polars, duckdb, or auto
--verbose	false	Show per-column progress logs

---

Schema Reference

tables:
  - name: customers          # lowercase identifier
    rows: 1000
    columns:
      - name: id
        type: integer
        primary_key: true    # required — exactly one per table

      - name: full_name
        type: name           # pre-baked numpy name pool (~6M rows/sec)

      - name: email
        type: email
        unique: true

      - name: signup_date
        type: datetime
        range: [2020-01-01, 2024-12-31]

  - name: orders
    rows: 5000
    columns:
      - name: id
        type: integer
        primary_key: true

      - name: customer_id
        type: foreign_key
        references: customers.id   # always table.column

      - name: amount
        type: float
        range: [5, 5000]
        distribution: lognormal    # see distributions below

      - name: order_date
        type: datetime
        range: [2021-01-01, 2024-12-31]

    chaos:
      null_probability: 0.03         # 3 % of non-PK/FK values become null
      duplicate_probability: 0.01    # 1 % extra duplicate rows appended
      drift:                         # concept drift on a numeric column
        field: amount
        distribution: norm
        start_after_row: 4000
      format_inconsistency:          # randomly mix date formats
        fields: [order_date]
        style: mixed_us_dates

Column types

Type	Description
integer	Random integers. Add primary_key: true for sequential IDs.
float	Uniform floats. Combine with distribution for realistic shapes.
string	Random 12-char alphanumeric strings.
name	Full person names sampled from a pre-baked pool.
email	[e-mail] — add unique: true for deduplication.
datetime	Timestamps within an optional range.
foreign_key	Samples from a parent table's primary key. Requires references.

Distributions

Use distribution on float columns (or integer with drift):

Value	Shape
lognormal / lognorm	Right-skewed — good for prices, revenue
normal / norm	Bell curve
uniform	Flat
expon	Exponential decay
poisson	Integer-valued counts
Any scipy.stats name	Fallback to scipy, e.g. beta, gamma

---

All Commands

chaoslake --help

Command	What it does
generate	Generate tables from a YAML schema
init	Create chaoslake.yaml (interactive or static template)
check	Validate a schema file without generating data
quick	One-command demo — no schema file needed
introspect	Reflect a live database and emit a Chaoslake YAML schema

Introspect a database

chaoslake introspect --db sqlite:///mydb.db
chaoslake introspect --db postgresql://user:pass@localhost/mydb --tables customers,orders

Requires pip install "chaoslake[dev]" (SQLAlchemy).

---

Reproducible Output

Pass --seed to get bit-identical datasets every time:

chaoslake generate --schema chaoslake.yaml --output ./output --seed 42

---

Performance

Benchmarked on Apple M-series (1.1M rows, 3 columns):

Engine	Rows / sec
Polars (default)	~6 000 000
DuckDB	auto-selected above 500k total rows

Run the benchmark yourself:

chaoslake bench

---

License

MIT © Gourav Shokeen