Question About 'movingAverageParam' in MarginalHitsStrategy

[hazelnut-99]

Hi team,

I've been exploring different slab rebalancing strategies in CacheLib and had a question about the MarginalHitsStrategy.

Specifically, I was looking into the movingAverageParam, which is used to smooth rankings based on the marginal hit score.

When looking into the moving average equation: code reference

new_avg = avg * movingAverageParam + newValue * (1 - movingAverageParam)

Intuitively, it feels like this smoothing would preserve the original ranking anyways and wouldn’t impact which allocation class gets picked as a victim/receiver (please correct me if I’m wrong).

I’m curious—what’s the design reasoning behind this smoothing approach, and how does this parameter help in practice?

Thanks a lot!

[hazelnut-99]

Update, I think I've got to understand how this parameter works. The blurring average was with each class's past scores.