会在某一个1s突然出现过多失败次数，造成错误熔断

## Issue Description

Type: 利用dubbo rpc框架，采用的随机权重负载均衡策略，其中某个下游服务器有8个节点，突然有个节点频繁old gc出现响应超时。consumer单台机器 qps 13左右，设置的降级规则，为失败率50%进行熔断，按说1/8的节点，不会到达50的失败率。但是看sentinel metric日志，会在某一个1s窗口突然出现10次左右失败统计，造成失败率超过50%进行熔断，所以反而把服务质量变得更差了，求解


Start with Sentinel's metric logging and circuit-breaker failure-rate calculation; the issue names no source files or tests. Reproduce the one-second window using roughly 13 QPS, eight downstream nodes, one node timing out during old GC, and a 50% failure threshold. Done means explaining or correcting the unexpected burst of about 10 failures and its resulting circuit break.

Written by the indexing model from the issue text.