Skip to content

Making It Reproducible

Now let's solve the first major problem: lack of execution tracking. We'll transform our basic function into a reproducible, tracked job without changing the ML logic at all.

The Solution: PythonJob

Instead of calling our function directly, we'll wrap it with Runnable's PythonJob:

examples/tutorials/getting-started/02_making_it_reproducible.py
from runnable import PythonJob
from functions import train_ml_model_basic

def main():
    # Same function, now wrapped as a Job
    job = PythonJob(function=train_ml_model_basic)
    job.execute()
    return job

if __name__ == "__main__":
    main()

Try it:

uv run examples/tutorials/getting-started/02_making_it_reproducible.py

What Just Happened?

Your ML function ran exactly the same way, but Runnable automatically added powerful tracking capabilities:

📝 Execution Logging

Every run gets logged with complete details:

ls .run_log_store/
# Shows JSON files with run IDs: curious-jang-0442.json

Each run directory contains:

  • Execution metadata: when it ran, how long it took
  • Environment info: Python version, package versions
  • Results: function return values
  • Status: success/failure with any error details

♻️ Result Preservation

Unlike the basic version that overwrote model.pkl and results.json, each Runnable execution gets its own directory. Your results are never lost.

🔍 Full Reproducibility

Each run captures everything needed to reproduce it:

  • Exact timestamp
  • Code version (if using git)
  • Environment details
  • Input parameters (we'll add those next!)

🎯 Zero Code Changes

Notice that train_ml_model_basic() didn't change at all. Runnable works with your existing functions - no decorators, no API changes, no refactoring required.

Run It Multiple Times

Try running the script several times:

uv run examples/tutorials/getting-started/02_making_it_reproducible.py
uv run examples/tutorials/getting-started/02_making_it_reproducible.py
uv run examples/tutorials/getting-started/02_making_it_reproducible.py

Each run creates a separate log entry in .run_log_store/. You now have a complete history of all your experiments!

Compare: Before vs After

Before (Chapter 1):

  • ❌ Results overwritten each time
  • ❌ No execution history
  • ❌ No timestamps or metadata
  • ❌ Hard to track what worked

After (Chapter 2):

  • ✅ Every run preserved with timestamp
  • ✅ Complete execution history
  • ✅ Full metadata captured automatically
  • ✅ Easy to see what worked when

What's Still Missing?

We solved execution tracking, but we still have:

  • Parameters hardcoded in the function
  • No easy way to run experiments with different settings

Next chapter: We'll make the function configurable without changing the ML logic.

Next: Adding Flexibility - Configure experiments without touching code