feature-pipeline-quality

A small, production-style CLI for validating tabular feature datasets against a contract before training or scoring jobs run.

It checks schema presence, row count, null thresholds, type parsing, duplicate keys, and freshness, then emits a JSON report and CI-friendly exit codes.

Validation flow

flowchart LR
    A["Feature dataset CSV"] --> C["feature-pipeline-quality validate"]
    B["Contract JSON"] --> C
    D["As-of date"] --> C
    C --> E["Structured JSON report"]
    C --> F["Exit code for CI or orchestration"]
    F --> G{"Passed?"}
    G -->|Yes| H["Training or scoring job proceeds"]
    G -->|No| I["Pipeline stops with explicit failure"]

Loading

Why this exists

Feature pipelines often fail silently until model training or scoring breaks downstream. This repo demonstrates a simple contract-check pattern that can run in scheduled jobs, CI, or orchestration tasks.

Features

• Required column checks
• Row count minimums
• Column-level null ratio thresholds
• Type parsing checks (str, int, float, bool, date, datetime)
• Duplicate detection on composite keys
• Freshness checks on a date/datetime column
• Structured JSON output + non-zero exit code on failure

Repository shape

• examples/contract.json shows the validation contract expected by the CLI.
• examples/features_good.csv and examples/features_bad.csv give passing and failing fixtures.
• tests/test_validator.py covers the core contract checks and exit behavior.
• pyproject.toml exposes the package metadata and console entry point.

Quickstart

python -m feature_pipeline_quality validate \
  --contract examples/contract.json \
  --data examples/features_good.csv \
  --as-of 2026-02-25

Failing example:

python -m feature_pipeline_quality validate \
  --contract examples/contract.json \
  --data examples/features_bad.csv \
  --as-of 2026-02-25

Write a JSON report:

python -m feature_pipeline_quality validate \
  --contract examples/contract.json \
  --data examples/features_good.csv \
  --report /tmp/feature-quality-report.json

Exit codes

• 0 = dataset passed contract checks
• 2 = one or more checks failed
• 1 = invalid input / runtime error

Where this fits

Use this pattern ahead of model training, batch scoring, or feature publication when you want explicit data contracts instead of silent downstream breakage. It pairs naturally with policy gates such as ml-release-gates and the broader operating patterns in ml-platform-reference.

Test

python -m unittest discover -s tests -v