Capacity Planning

Input dimensions

• active unique limit keys per window
• algorithms used (token_bucket, cost_based, token_bucket_llm)
• expected concurrency and burst profile
• retention/reset windows

Practical sizing workflow

1. Estimate unique active keys per critical rule
2. Start with FAIRVISOR_SHARED_DICT_SIZE=128m
3. Load test with realistic cardinality
4. Increase to 256m/512m if signs of state pressure appear

Rule-of-thumb table

Active key cardinality	Starting dict size
<= 100k	128m
100k-250k	256m
250k-500k	512m
> 500k	profile and shard strategy

Validate after sizing

• stable allow/reject semantics under load
• no abnormal limiter resets
• latency remains within SLO at p95/p99