We wanted to drive a lot more control into our A/B testing process.

Basic requirements were:

1. Ability to create multiple experiments and multiple variants for an experiment.
2. Should be able to distribute requests as configured (by percentage) between different variants of an experiments.
3. Should do so in a deterministic way. That is, until the rollout of an experiment is changed, a user should always get the same variant of an experiment.
4. Overrides based on user properties.
5. Should have very low latency.
6. Should have data logging capabilities.

Flagr fits the bill for all the above requirements mentioned and does it with a very low latency (it's written in Golang and caches data in memory to avoid it's own DB calls).