git-bayesect 1.1

Git bisection with Bayesian statistics

git bayesect

Bayesian git bisection!

Use this to detect changes in likelihoods of events, for instance, to isolate a commit where a slightly flaky test became very flaky.

You don't need to know the likelihoods (although you can provide priors), just that something has changed at some point in some direction

Installation

pip install git_bayesect

Or:

uv tool install git_bayesect

How it works

git_bayesect uses Bayesian inference to identify the commit introducing a change, with commit selection performed via greedy minimisation of expected entropy, and using a Beta-Bernoulli conjugacy trick while calculating posterior probabilities to make handling unknown failure rates tractable.

See https://hauntsaninja.github.io/git_bayesect.html for a write up.

Usage

Start a Bayesian bisection:

git bayesect start --old $COMMIT

Record an observation on the current commit:

git bayesect fail

Or on a specific commit:

git bayesect pass --commit $COMMIT

Check the overall status of the bisection:

git bayesect status

Reset:

git bayesect reset

More usage

Set the prior for a given commit:

git bayesect prior --commit $COMMIT --weight 10

Set prior for all commits based on filenames:

git bayesect priors_from_filenames --filenames-callback "return 10 if any('suspicious' in f for f in filenames) else 1"

Set prior for all commits based on the text in the commit message + diff:

git bayesect priors_from_text --text-callback "return 10 if 'timeout' in text.lower() else 1"

Set the beta priors:

# We expect "fail" observations 90% of the time at new commit, 5% of the time at old commit
git bayesect beta_priors --alpha-new 0.9 --beta-new 0.1 --alpha-old 0.05 --beta-old 0.95

Get a log of commands to let you reconstruct the state:

git bayesect log

Undo the last observation:

git bayesect undo

Run the bisection automatically using a command to make observations:

git bayesect run $CMD

Checkout the best commmit to test:

git bayesect checkout

Demo

This repository contains a little demo, in case you'd like to play around:

# Create a fake repository with a history to bayesect over
python scripts/generate_fake_repo.py
cd fake_repo

# The fake repo contains a script called flaky.py
# This is a simple script that fails some fraction of the time
# At some point in the history of the repo, that fraction was changed
python flaky.py
git log --oneline

# Start the bayesection
OLD_COMMIT=$(git rev-list HEAD --reverse | head -n 2 | tail -n 1)
git bayesect start --new main --old $OLD_COMMIT

# Run a bayesection to find the commit that introduced the change
git bayesect run python flaky.py