Gust Benchmark

This repository compares Gust and Apache Airflow under the same workflow and Docker-based setup.

TL;DR

Gust is significantly more resource-efficient than Apache Airflow, requiring up to $4.4\times$ less memory when idle and roughly half the peak RAM to handle identical parallel workloads, while maintaining a lower CPU footprint during orchestration.

The benchmark uses an equivalent three-step pipeline in both systems:

1. Fetch data from a mock HTTP server
2. Transform the payload and write intermediate data to disk
3. Send the transformed payload back to the mock server

The goal is to measure how each orchestrator behaves at rest and under bursty parallel execution.

Scenarios

The benchmark covers these scenarios:

• Idle system
• 50 parallel runs
• 80 parallel runs
• 100 parallel runs

In the environment used for the recorded results below, Airflow did not complete the 100-run scenario, while Gust did.

Each scenario was run independently and the Docker stack was restarted between runs.

Repository Layout

• gust/: Gust Docker stack, DAG, and mock server
• airflow/: Airflow Docker stack, DAGs, and mock server
• livebooks/gust_vs_airflow_benchmark.livemd: Livebook used to trigger runs and collect docker stats

Workflow Under Test

Both implementations run the same logical pipeline:

• extract_data: call GET /extract on the mock server
• transform_data: normalize and persist the returned data
• load_data: call POST /load with the transformed payload

The mock server is configured with:

• MOCK_DELAY_SECONDS=2
• MOCK_ITEM_COUNT=1000

Running The Benchmark

Gust

Start the stack:

cd gust
docker compose up --build

Trigger one run:

docker compose exec -T gust /app/bin/gust-cli trigger_run benchmark_mock_io

Gust Web is exposed at http://localhost:4000.

Airflow

Start the stack:

cd airflow
docker compose up --build

Trigger one run:

docker compose exec -T airflow-scheduler airflow dags trigger benchmark_mock_io

Airflow is exposed at http://localhost:8080.

Automated Runs

The benchmark orchestration used for the recorded measurements lives in:

• livebooks/gust_vs_airflow_benchmark.livemd

That notebook triggers concurrent runs and samples docker stats during execution.

Recorded Results

Versions used for the recorded results:

• Airflow 3.9
• Gust 0.1.31

Peak Memory Usage

Scenario	Airflow	Gust	Observation
Idle	1,710 MiB	387 MiB	Gust used about 4.4x less RAM
50 parallel runs	6,222 MiB	3,064 MiB	Gust used about 2.0x less RAM
80 parallel runs	7,239 MiB	5,508 MiB	Gust still used less RAM
100 parallel runs	Not completed	5,648 MiB	Gust completed this scenario in the recorded run

Peak CPU Usage

Scenario	Airflow	Gust	Observation
Idle	137%	~5%	Gust stayed close to idle at rest
50 parallel runs	1,419%	1,114%	Gust showed lower CPU pressure
80 parallel runs	1,313%	1,198%	Gust remained slightly lower

Takeaways

• Lower baseline cost: Gust used substantially less memory in the idle scenario.
• Lower burst overhead: Gust showed lower CPU and memory pressure during 50- and 80-run bursts.
• Better headroom in this setup: Gust completed the 100-run scenario; Airflow did not in the recorded environment.

Notes

• These numbers come from a local Docker benchmark and should be treated as comparative, not universal.
• The Airflow and Gust workflows are intended to be equivalent, but implementation details of each platform still matter.
• If you want to reproduce or extend the benchmark, start with the Livebook and the two Docker Compose stacks in this repo.