Watch k self-tune via gradient descent — thresholds learn from hit rate feedback
Current k
2
Graduation Rate
0%
CloxCache
0%
LRU
0%
Cache Pressure:
Medium
Learned Thresholds:
25% - 50%
k value Graduation % CloxCache % LRU %
Synthetic workload with power-law distribution (random α 0.80-1.20)
10,000 / 100,000 keys (10%)
How it works: Track the graduation rate — what fraction of items cross from freq=k to freq=k+1 (becoming protected).
Low graduation rate? Items die before proving their worth → lower k to make protection easier.
High graduation rate? Protection is working well → raise k to be more selective.
Self-tuning: The thresholds themselves are learned via gradient descent. If changing k improved hit rate, make that change easier; if it hurt, make it harder.