Experimentation
Automatic weight rebalancing
The realised split is measured and corrected on a schedule, so the traffic each variant actually gets converges on what you asked for.
Randomised assignment does not produce a perfectly even split, and small deviations accumulate. Rebalancing measures what each variant actually received and pushes the configured weights the other way, so the cumulative split converges on the one you configured rather than drifting from it.
The job runs every fifteen minutes against running experiments. It reads session counts per variant, computes a correction, writes the new weights, and patches the generated runtime file in place rather than rebuilding it, because a rebalance only changes a few integers.
Two guards keep it from doing harm. Below a minimum session count nothing is adjusted, since at low volume the deviation being corrected is just noise. An experiment with a deployed variant is skipped entirely: a deployment means the experiment has been concluded and its split is no longer a measurement question.
Everything it does is logged per experiment, so a split that looks surprising can be traced to the ticks that produced it.