jianf/ModelRouter
1
0
Fork 0
mirror of https://gitcode.com/JianFeeeee/ModelRouter.git synced 2026-10-07 08:03:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf(gateway): 配额桶扫描改为窗口化 + pinned 桶惰性创建

Browse Source 
审查本特性线的性能时发现两个问题，均有实测数据。

## 1. 窗口查询是全扫，代价落在每个请求上

sumBuckets 原来遍历整个 map（最多 960 个小时桶），实测 5.9us/op。
配额检查在每个请求上跑 2-3 次（key 总 token、key 请求数、scope token），
于是单请求多付约 18us。

注意这**不是本改动引入的成本**：main 上既有的 WindowTokens 同样是
5907ns/op（全扫）。是本改动让它在请求路径上被调用得更多。

改为只遍历窗口可能覆盖的桶（键是整点小时，范围是精确的，不是采样）：
- 24h 窗口 200ns -> 55ns
- 1h  窗口  80ns -> 49ns
- 30d 窗口 5.9us -> 3.9us（720 次查找，只有配 month 配额时才走到）

等价性由 TestSumBucketsMatchesFullScan 保证（400 组随机桶位置 x 6 种
窗口，对全扫逐项比对）。★ 第一次写错成 floor，判据立刻抓到：
30 天窗口报 8878 而全扫是 8649 —— 正确是 ceil。

## 2. pinned 桶无条件创建，内存最坏 26.7MB

每条记录写两个桶：裸 model 与 "source::model"。但 pinned 桶只有
「配额里显式 pin 了 source」时才会被查。

实测最坏情况（20 key x 8 model x 3 source x 40 天全 retention）：
HeapAlloc 26.67MB —— 而 README 宣传「16 源生产实例 ~32-35MB」，
等于吃掉 80% 内存预算。

改为惰性：只有 KeyWindowModelTokens 带 source 查询过某个 (key, model)
之后，才开始维护它的 pinned 桶。

  20key x 8model x 3src   26.67MB -> 8.87MB  (-67%)
  5key x 6model（真实）     3.36MB -> 2.26MB  (-33%)
  5key x 12model            5.82MB -> 3.59MB  (-38%)

代价：配 pinned 配额之前发生的用量无法事后按源拆分（记录里虽然有
Source，但桶只存了裸 model），所以 pinned 配额的首个窗口可能少算。
已在代码注释与判据中写明。

## 其余实测

  Record        main 基线 275ns/429B/3allocs -> 283ns/429B/3allocs
                （+8ns，分配数不变；3 allocs 来自 ring buffer）
  配额检查全路径  115ns / 0 allocs（每请求新增）
  纯读路径      6.5ns / 0 allocs

## 100 并发调度/拒绝压测（真实进程 + 可报并发峰值的假上游）

  容量 100（4+96），0.15s/请求，100 并发  ok=100 fail=0   上游峰值 42
  容量 100，0.15s/请求，200 并发          ok=200 fail=0   上游峰值 97
  容量 8，3s/请求，100 并发               ok=8   fail=92  上游峰值 8
  容量 8，0.6s/请求，100 并发             ok=32  fail=68  上游峰值 8
  容量 8，0.6s/请求，40 并发              ok=32  fail=8   上游峰值 8

上游峰值恒定不超过 max_concurrent，容量拒绝返回 503 + busyWait 有界
等待（约 2.6s）。main 基线在同条件下 ok=32 fail=68、上游峰值 8、
延迟分布相同 —— 配额改动没有触碰调度/拒绝路径。

配额拒绝单独验证（低并发避开容量拒绝）：req_quota=50 用尽后
100 并发全部 429 rate_limit_exceeded + Retry-After: 1661，
**延迟仅 19-21ms**、上游 total 未增加 —— 配额在入口廉价拒绝，
不占用任何上游槽位，与容量不足的昂贵等待形成明确分工。

## 判据

key_quota_perf_test.go：4 个基准 + 2 个判据（sumBuckets 等价性、
pinned 桶惰性）。3/3 变异全被抓（无条件建 pinned 桶、firstHour 用
floor、keyHour 不再写）。
This commit is contained in:
JianFeeeee
2026-09-27 18:34:13 +08:00
parent b5c3fea0bb
commit be11a06a46
3 changed files with 235 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
internal/gateway/stats.go
View File

@ -96,6 +96,9 @@ type Stats struct {
// key-wide request quota. Counted for every request type (chat, stream,
// image) including failed ones, so a failing client cannot loop for free.
keyReqHour map[string]map[int64]int64
// pinnedWanted records which (key, model) pairs have a quota that pins a
// source, so the pinned bucket is maintained only where it is read.
pinnedWanted map[string]bool
}
const hourSec = 3600
@ -138,6 +141,7 @@ func NewStats(maxRecords int) *Stats {
keyModelHour: map[string]map[string]map[int64]int64{},
keyHour: map[string]map[int64]int64{},
keyReqHour: map[string]map[int64]int64{},
pinnedWanted: map[string]bool{},
maxRecs: maxRecords,
}
}
@ -406,11 +410,16 @@ const quotaRetentionHours = 24 * 40
// addKeyTokenLocked adds tok to one key's (model[, source]) hourly bucket.
//
// It writes BOTH an unpinned and a pinned bucket. Every recorded request
// carries a resolved source, so the bucket key would otherwise always be
// "source::model" and a quota without a source pin would read an empty bucket
// and never trip. The unpinned bucket is the one a scope entry normally looks
// up; the pinned one serves entries that pin a source.
// The bare-model bucket is always written: it is what a scope entry without a
// source pin reads, and writing it is what makes that quota work at all (every
// recorded request carries a resolved source, so a source::model-only bucket
// would leave an unpinned quota reading an empty bucket forever).
//
// The pinned "source::model" bucket is written only for (key, model) pairs
// some scope entry actually pins a source for. Maintaining it unconditionally
// doubles the bucket count — measured at 26.7 MB for 20 keys x 8 models x 3
// sources at full retention, against a documented ~32 MB total budget — to
// serve a lookup that almost nobody performs.
func (s *Stats) addKeyTokenLocked(key, model, source string, h, tok int64) {
if model == "" {
return
@ -420,34 +429,45 @@ func (s *Stats) addKeyTokenLocked(key, model, source string, h, tok int64) {
byModel = map[string]map[int64]int64{}
s.keyModelHour[key] = byModel
}
// Write the bare-model bucket plus, when a source is known, the pinned
// one. The unpinned bucket is the aggregate the scope lookup normally
// reads; the pinned bucket serves entries that pin a source.
keys := [2]string{model, model}
n := 1
if source != "" {
if pinned := source + "::" + model; pinned != model {
keys[1] = pinned
n = 2
}
addBucket(byModel, model, h, tok)
if source == "" {
return
}
for i := 0; i < n; i++ {
hm := byModel[keys[i]]
if hm == nil {
hm = map[int64]int64{}
byModel[keys[i]] = hm
}
hm[h] += tok
if len(hm) > quotaRetentionHours {
for k := range hm {
if k < h-quotaRetentionHours {
delete(hm, k)
}
pinned := source + "::" + model
if pinned == model || !s.pinnedWanted[key+"\x00"+model] {
return
}
addBucket(byModel, pinned, h, tok)
}
// addBucket adds tok to one named hourly bucket and enforces retention.
func addBucket(byModel map[string]map[int64]int64, name string, h, tok int64) {
hm := byModel[name]
if hm == nil {
hm = map[int64]int64{}
byModel[name] = hm
}
hm[h] += tok
if len(hm) > quotaRetentionHours {
for k := range hm {
if k < h-quotaRetentionHours {
delete(hm, k)
}
}
}
}
// wantPinnedBuckets starts maintaining the pinned (source::model) bucket for a
// key's model, so a quota that pins a source counts traffic from the moment it
// is configured. Usage recorded before that cannot be attributed to a source
// after the fact, so a pinned quota may under-count its first window.
func (s *Stats) wantPinnedBuckets(key, model string) {
if key == "" || model == "" {
return
}
s.pinnedWanted[key+"\x00"+model] = true
}
// addKeyHourLocked adds tok to one key's all-model hourly total.
func (s *Stats) addKeyHourLocked(key string, h, tok int64) {
hm := s.keyHour[key]
@ -490,8 +510,15 @@ func (s *Stats) addKeyReqLocked(key string, h, n int64) {
}
}
// sumBuckets totals the hourly buckets at or after the cutoff given in unix
// seconds. sec <= 0 means "all retained history" (no reset).
// sumBuckets totals the hourly buckets inside the window ending at `now`
// (unix seconds). sec <= 0 means "all retained history" (no reset).
//
// It only visits the hour buckets a window can actually cover. The naive full
// scan measured 5.9us/op at the full 40-day retention, and a per-key quota
// check runs on every request (2-3 of these), so the scan cost lands directly
// on the request path; visiting at most sec/3600 + 1 buckets makes it
// proportional to the window instead of to the retained history. The map keys
// are whole unix hours, so the covered range is exact rather than a sample.
func sumBuckets(hm map[int64]int64, now, sec int64) int64 {
if len(hm) == 0 {
return 0
@ -503,12 +530,19 @@ func sumBuckets(hm map[int64]int64, now, sec int64) int64 {
}
return total
}
cut := now - sec
nowHour := now / hourSec
// A bucket at hour H covers [H*3600, (H+1)*3600). The pre-existing
// WindowTokens counts a bucket when h*3600 >= now-sec, so the first hour
// that qualifies is ceil((now-sec)/3600) — NOT floor, which would pull in
// one bucket too many (caught by TestSumBucketsMatchesFullScan: a 30-day
// window reported 8878 vs 8649).
firstHour := (now - sec + hourSec - 1) / hourSec
if firstHour < 0 {
firstHour = 0
}
var total int64
for h, v := range hm {
if h*hourSec >= cut {
total += v
}
for h := firstHour; h <= nowHour; h++ {
total += hm[h]
}
return total
}
@ -532,6 +566,9 @@ func (s *Stats) KeyWindowModelTokens(key, model, source string, sec int64) int64
mk := model
if source != "" {
mk = source + "::" + model
// from now on this pair keeps a pinned bucket, so a quota configured
// against this source counts the traffic it is meant to cap
s.wantPinnedBuckets(key, model)
}
return sumBuckets(byModel[mk], time.Now().Unix(), sec)
}