Performance Tuning

Goals

• Keep decision path latency low and predictable
• Avoid shared-dict pressure and hot-key contention

Key levers

1) Shared dict sizing

FAIRVISOR_SHARED_DICT_SIZE controls in-memory limiter state capacity.

• Start at 128m
• Increase to 256m/512m for high cardinality keys
• Monitor for eviction-like behavior and unexpected limiter resets

2) Descriptor cardinality

Avoid unbounded key dimensions in limit_keys.

• Prefer stable IDs (jwt:org_id, header:x-api-key)
• Avoid raw per-request unique values as keys

3) Gateway timeout budget

Keep gateway->edge auth timeout tight (typical 200-500ms budget) and observe p99.

4) Worker and runtime tuning

• FAIRVISOR_WORKER_PROCESSES=auto usually works well
• Increase only after measuring CPU saturation and queueing

Practical baseline

• FAIRVISOR_SHARED_DICT_SIZE=128m
• deterministic retry jitter enabled by default
• short gateway timeout and clear fail-open/fail-closed policy

Validate after changes

• Reject/allow mix remains stable
• No spike in decision latency metrics
• No abnormal descriptor-missing metrics