Bubble Sort’s Early Exit: code and data

Read the published article on Medium.

Cover illustration: the timing labels refer to the 2,000-value experiment. The six-item examples are illustrative and are not the paired benchmark inputs.

Gokul Krishna · educational computational experiment.

This bundle accompanies “Bubble Sort’s Early Exit: Why Input Order Matters.” It contains only the Bubble Sort implementation, synthetic paired inputs, its recorded timings/calibration, analysis and figure. These observations illustrate an early exit and measurement practice; they are not a novelty or peer-reviewed publication claim. See RIGHTS.md for licensing: MIT for code; CC BY 4.0 for data, documentation and figures. Release: v1.0.1.

What is here

• bubble_sort.py: the measured copying implementation, extracted without a body change.
• workloads.py: the paired generator and stable seed/hash helpers.
• measurement.py: the original timer/calibrator functions, without workspace tooling.
• benchmark.py: a new, portable Bubble-only rerun driver.
• data_io.py, analyze.py, plot.py, verify.py: validation, summaries, chart and integrity checks.
• tests/test_bundle.py: correctness, count, GC-state, input/provenance and aggregation checks.
• data/: both datasets' original Bubble-only timing/calibration records and all exact inputs.
• results/: regenerated input medians, summaries and the article table.
• figures/: the measured chart in PNG/SVG, its provenance, and the supplied article cover.
• METHODS.md, DATA_DICTIONARY.md, PROVENANCE.json, config.json: interpretation and settings.
• requirements-figures.txt: pinned optional plotting dependencies; analysis uses only the standard library.
• MANIFEST.json: exact payload inventory, byte counts and SHA-256 hashes (excluding itself).
• LICENSE, LICENSE-CC-BY-4.0.txt, RIGHTS.md: license terms and their scope.

Verify and reproduce saved results

Use the recorded CPython 3.12.13 environment when possible. Run from this directory. The first four commands need no third-party packages or network access:

python3.12 -m venv /tmp/bubble-review-env
/tmp/bubble-review-env/bin/python verify.py --regenerate-inputs
/tmp/bubble-review-env/bin/python -m unittest discover -s tests -v
/tmp/bubble-review-env/bin/python analyze.py --output /tmp/bubble-summary

Use new output paths: existing output directories are never replaced. Seed regeneration is an additional check; verify.py without that flag checks the exact saved fixtures and hashes without depending on a future PRNG sampling implementation.

To regenerate the chart:

/tmp/bubble-review-env/bin/python -m pip install -r requirements-figures.txt
/tmp/bubble-review-env/bin/python plot.py --summary /tmp/bubble-summary/summary.csv --output /tmp/bubble-figures

The pinned figure environment was used on macOS ARM64/Python 3.12.13. Other platforms may need compatible wheels; record any dependency changes. The supplied figure can also be viewed without installing plotting dependencies.

Collect a new timing run

A small engineering run:

/tmp/bubble-review-env/bin/python benchmark.py --dataset dataset-1 --size 100 --inputs 1 --repeats 2 --output /tmp/bubble-smoke
/tmp/bubble-review-env/bin/python analyze.py --run /tmp/bubble-smoke --output /tmp/bubble-smoke-summary

Omit --size, --inputs and --repeats for all 600 Bubble batches in that dataset. Use dataset-2 and a different output directory for the second dataset. New wall clock times will differ. Do not overwrite or pool them with historical observations.

Collection context matters: historical Bubble observations were interleaved with four other sorting methods. Their original order positions are retained. This driver runs only Bubble Sort and does not reconstruct that cross-method schedule. It reproduces the exact input values, function, calibration rules and timing boundaries. See METHODS.md for this limitation and the summary procedure.

Article-size results

At n=2,000, median milliseconds per call (rounded):

Order	Dataset 1	Dataset 2
Shuffled	112.7	112.6
Sorted	0.06917	0.07046
Reversed	140.6	141

The linear best-case comparison count is n−1 for an already sorted input of n≥2 values; descending distinct inputs make n(n−1)/2 comparisons. Those formulas are not elapsed-time ratios. Swaps, copying and interpreter/machine costs also matter.

AI tools assisted with code, experiments, analysis and documentation. The named author remains responsible for reviewing the scientific content and any publication.