Our recommended approach is to keep an even split between variations for the entire experiment and instead adjust the overall traffic percent that is included to increase traffic.
For example, start with a 50/50 split on 10% of traffic, then increase that to 20%, etc. to send more traffic there.
When you do that, users in the experiment will not be reassigned, even without any stickiness or persistence. The hashing algorithm we use ensures that.
For more advanced changes beyond just increasing traffic, we did just launch a sticky bucketing solution you can check out.
https://docs.growthbook.io/app/sticky-bucketing