ml-goodput-measurement

What it is and what it does

ML Goodput Measurement is a library for quantifying the productive time (Goodput) and idle/overhead time (Badput) of machine learning training jobs running on cloud accelerators. It provides a GoodputRecorder to instrument your training code with timestamps for job start/end, individual training steps, device initialization, data loading, and training preparation. The recorded data is sent to Google Cloud Logging, where a separate GoodputCalculator can analyze it to compute overall productivity metrics and breakdowns of where time is lost. A GoodputMonitor component can asynchronously query and export these metrics to TensorBoard for real-time visibility.

The package is designed to work with Google Cloud accelerators and requires a GCP project with Cloud Logging enabled and appropriate access scopes. It depends on google-cloud-logging, google-cloud-monitoring, numpy, scipy, requests, and related libraries. The typical workflow is to instrument your training application with recorder calls, let it run, then run a separate analysis program to compute Goodput and understand where compute resources are underutilized.

Use it for:

• Identify bottlenecks in ML training pipelines by measuring productive computation versus data loading and overhead.
• Monitor training job efficiency in real-time via TensorBoard exports to catch performance regressions early.
• Analyze step-time deviation across distributed training to detect stragglers or synchronization issues.
• Quantify the impact of system changes on actual training productivity rather than just wall-clock time.
• Debug why a training job is slower than expected by breaking down time spent in initialization and preparation.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Measures and monitors the productive utilization (Goodput) and idle time (Badput) of ML training jobs running on cloud accelerators, logging metrics to Google Cloud and exporting analysis to TensorBoard.

Yes, if you are running ML training on Google Cloud accelerators and need visibility into actual compute utilization. The low install friction, active maintenance, and lack of known vulnerabilities make it safe to adopt. However, verify the license status before use in proprietary contexts, and confirm that your GCP project and cluster access scopes are properly configured—setup is non-trivial.

Install

ml-goodput-measurement on PyPI

pip install ml-goodput-measurement

uv add ml-goodput-measurement

poetry add ml-goodput-measurement

Installing ml-goodput-measurement

Before you install

Low friction installation with a pure Python wheel. Active maintenance as of 2026-08-13 with recent releases. Requires 8 runtime dependencies including google-cloud-logging, google-cloud-monitoring, numpy, scipy, and requests—all widely used packages.

License in practice

License treatment is unclear; the description excerpt references Apache License 2.0, but the fact sheet shows no license_spdx or license_raw value. Verify the actual license before use in proprietary or restricted contexts.

Quickstart

pip install ml-goodput-measurement

from ml_goodput_measurement import goodput

goodput_recorder = goodput.GoodputRecorder(
    job_name='my_training_run',
    logger_name='goodput_my_training_run',
    logging_enabled=True
)
goodput_recorder.record_job_start_time(datetime.datetime.now())

Requires a Google Cloud project with billing enabled, Cloud Logging API enabled, and appropriate access scopes on GPU/TPU and CPU node pools.

Verify before relying

• Whether the Apache License 2.0 reference in the description excerpt is the authoritative license.
• Whether the package works with non-Google cloud accelerators or only GCP TPU/GPU.
• Performance overhead of instrumentation on training step timing.

Package facts

Evidence: ml_goodput_measurement-0.2.2-py3-none-any.whl

Tags

More Monitoring packages