Skip to content

Latest commit

 

History

History
22 lines (21 loc) · 1.09 KB

File metadata and controls

22 lines (21 loc) · 1.09 KB

Limitations

  • V1 is a discrete-event simulator, not a live service.
  • Logical throughput is not wall-clock throughput.
  • Request service time is known from the offline trace. DRR and SCFQ therefore use oracle cost information that a production system would need to estimate or require clients to declare.
  • Workers are identical and execution is non-preemptive.
  • Network delay, retries, failures, autoscaling, and distributed coordination are not modeled.
  • SCFQ's classical fairness analysis is single-server; multi-worker runs are an experimental deterministic extension.
  • SCFQ finish tags assigned to later requests are not recomputed when an earlier queued request times out.
  • Jain fairness is descriptive when tenant demand differs.
  • The starvation count is a configured finite-horizon proxy.
  • Built-in workloads are synthetic and intentionally small enough for tests and source review.
  • Matplotlib plots summarize recorded outputs but do not add statistical confidence intervals.
  • V1 does not claim to outperform Netflix concurrency-limits, Envoy, or any production rate limiter.