Generate and evaluate a MongoDB workload

Goal: connect schema inspection, sampling estimates, generation and execution. Install the toolkit and initialize MongoDB demo data first.

uv run --with jupyterlab --with ipykernel jupyter lab notebooks/workload_generation.ipynb

Run the notebook's offline examples first. They generate 10 patterns with 50 instances, inspect pattern membership, save/load JSONL and demonstrate seeded generation. Enable its RUN_LIVE cell to use demo.business on port 27017. That example builds a schema, prepares a temporary sample, constrains every query's selectivity and cleans up the sample when finished. It does not modify source documents.

Execute the generated workload

The notebook's final live section runs and profiles live_workload in two separate passes. This is the workload generated from the actual demo schema, not the earlier offline toy example. It checks outcomes and prints a few client timings and parsed index-examination counts.

Execution elapsed time and explain-derived server time measure different things. Keep concurrency at one for isolated query behavior; increase it only when concurrent load is part of the experiment.

The demo sample is deliberately tiny. Try changing the selectivity bounds and inspect generator.last_stats: tight bounds may make complete patterns impossible. For full-population guarantees, add exact filter counts as a separate verification step in your research project before publishing the workload.