M4a:神识全量网关能力——OpenAI 兼容适配器(元炁共用)、候选路由+凭证/绑定级冷却+首字超时切道+粘性会话、用量计量与 /me/usage,故障注入测试四场景

This commit is contained in:
chenxin committed 2026-08-27 09:32:37 +08:00
1 parent df5caa69e7
commit 7279904043
8 files changed
+717 -35

No files matched your search

+128 -19
View File
@@ -1,33 +1,42 @@
// Package server 神识 HTTP 服务:数据面(IR 入站 → SSE 出站)与健康面。
// Package server 神识 HTTP 服务:数据面(IR 入站 → SSE 出站)、健康面与用量面。
//
// 故障切换红线:首帧到达前不提交 SSE 响应——失败可整体换道并返回 JSON 错误;
// 已写出字节后不再换道,如实以 error 终帧收尾(shenshi-gateway.md §4/§5)。
package server
import (
"context"
"encoding/json"
"net/http"
"time"
"hunjing/shenshi/internal/ir"
"hunjing/shenshi/internal/metering"
"hunjing/shenshi/internal/relay"
"hunjing/shenshi/internal/route"
"hunjing/shenshi/internal/transform"
)
// Config 服务配置。
type Config struct {
ServiceToken string // 灵台→神识服务令牌;为空时开发档不校验
ServiceToken string
}
// Server 神识 HTTP 服务。
type Server struct {
cfg Config
routes *route.Registry
usage *metering.Log
mux *http.ServeMux
}
// New 构建服务并注册路由。
func New(cfg Config, routes *route.Registry) *Server {
s := &Server{cfg: cfg, routes: routes, mux: http.NewServeMux()}
s := &Server{cfg: cfg, routes: routes, usage: metering.New(10_000), mux: http.NewServeMux()}
s.mux.HandleFunc("GET /healthz", s.handleHealth)
s.mux.HandleFunc("GET /readyz", s.handleHealth)
s.mux.Handle("POST /v1/chat/completions", s.auth(http.HandlerFunc(s.handleChat)))
s.mux.Handle("GET /me/usage", s.auth(http.HandlerFunc(s.handleUsage)))
return s
}
@@ -41,8 +50,7 @@ func (s *Server) auth(next http.Handler) http.Handler {
next.ServeHTTP(w, r)
return
}
credential := r.Header.Get("Authorization")
if credential != "Bearer "+s.cfg.ServiceToken {
if r.Header.Get("Authorization") != "Bearer "+s.cfg.ServiceToken {
writeError(w, http.StatusUnauthorized, "HJ-2003", "服务令牌无效")
return
}
@@ -50,7 +58,7 @@ func (s *Server) auth(next http.Handler) http.Handler {
})
}
// handleChat 数据面:解析 IR → 路由 → 适配器流出 SSE。
// handleChat 数据面:IR → 候选迭代(首字超时/冷却标记/粘性)→ SSE。
func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
var req ir.ChatRequest
if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
@@ -61,35 +69,136 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
writeError(w, http.StatusBadRequest, "HJ-2002", err.Error())
return
}
endpoints, err := s.routes.Resolve(req.Slot)
candidates, err := s.routes.Candidates(req.Slot, req.RunID)
if err != nil {
writeError(w, http.StatusServiceUnavailable, "HJ-2001", err.Error())
return
}
relay.SSEHeaders(w)
w.WriteHeader(http.StatusOK)
var lastErr error
for _, endpoint := range candidates {
res := s.tryEndpoint(w, r, endpoint, &req)
s.usage.Record(res.Attempt)
if res.Committed {
// 已向客户端写出:无论成败都不能再换道
if res.Status == "ok" {
s.routes.Sticky.Remember(req.RunID, endpoint.Name)
}
return
}
// 未提交:分类冷却后换下一候选
if res.Upstream != nil {
if res.Upstream.StatusCode == 401 || res.Upstream.StatusCode == 403 {
s.routes.Cooldowns.Mark(route.CredentialCooldownKey(endpoint), true)
} else {
s.routes.Cooldowns.Mark(route.BindingCooldownKey(endpoint), false)
}
} else {
s.routes.Cooldowns.Mark(route.BindingCooldownKey(endpoint), false)
}
lastErr = res.Err
}
writeError(w, http.StatusServiceUnavailable, "HJ-2001", "全部候选端点失败:"+errString(lastErr))
}
type attempt struct {
metering.Attempt
Committed bool // 是否已向客户端写出 SSE
Upstream *transform.UpstreamError // 上游分类(可空)
Err error // 原始错误
}
func errString(err error) string {
if err == nil {
return "未知错误"
}
return err.Error()
}
// tryEndpoint 单端点尝试。
func (s *Server) tryEndpoint(w http.ResponseWriter, r *http.Request, endpoint route.Endpoint, req *ir.ChatRequest) attempt {
start := time.Now()
ctx, cancel := context.WithTimeout(r.Context(), s.routes.FirstByteTimeout)
defer cancel()
deltas := make(chan ir.Delta, 16)
streamErr := make(chan error, 1)
errCh := make(chan error, 1)
go func() {
defer close(deltas)
streamErr <- endpoints[0].Adaptor.Stream(r.Context(), &req, deltas)
errCh <- endpoint.Adaptor.Stream(ctx, req, deltas)
}()
res := attempt{Attempt: metering.Attempt{
RunID: req.RunID, CharacterID: req.CharacterID,
Model: req.Model, Slot: req.Slot, Endpoint: endpoint.Name,
}}
// 阶段一:等首帧或失败——此阶段未提交响应,可整体回退
var first ir.Delta
select {
case d, ok := <-deltas:
if !ok {
res.Err = <-errCh
res.Upstream, _ = res.Err.(*transform.UpstreamError)
res.Status, res.TotalMS = failStatus(res.Err), time.Since(start).Milliseconds()
return res
}
first = d
case <-ctx.Done():
res.Status, res.Err = "first_byte_timeout", ctx.Err()
res.TotalMS = time.Since(start).Milliseconds()
return res
}
// 阶段二:首帧已到——提交 SSE,流式转发
relay.SSEHeaders(w)
w.WriteHeader(http.StatusOK)
res.Committed = true
res.FirstTokenMS = time.Since(start).Milliseconds()
_ = relay.WriteDelta(w, first)
for d := range deltas {
if err := relay.WriteDelta(w, d); err != nil {
return // 客户端断开:取消上游由 r.Context() 传播
if d.Delta != "" || d.FinishReason != "" {
if err := relay.WriteDelta(w, d); err != nil {
break // 客户端断开
}
}
if d.FinishReason != "" {
res.Attempt.PromptTokens = d.UsagePromptTokens
res.Attempt.CompletionTokens = d.UsageCompletionTokens
}
}
if err := <-streamErr; err != nil {
// 流已开始,只能如实中断:写一条错误帧后终止(不伪装 [DONE])
_ = relay.WriteDelta(w, ir.Delta{FinishReason: "error", Delta: ""})
err := <-errCh
res.TotalMS = time.Since(start).Milliseconds()
if err == nil {
_ = relay.WriteDone(w)
return
res.Status = "ok"
return res
}
if err := relay.WriteDone(w); err != nil {
return
// 流中失败:如实以 error 终帧收尾,不伪装完成
res.Err = err
res.Upstream, _ = err.(*transform.UpstreamError)
res.Status = "error"
res.ErrorCode = "stream_written"
_ = relay.WriteDelta(w, ir.Delta{FinishReason: "error"})
_ = relay.WriteDone(w)
return res
}
func failStatus(err error) string {
if up, ok := err.(*transform.UpstreamError); ok {
if up.StatusCode == 0 {
return "timeout_or_network"
}
return "upstream_error"
}
return "error"
}
// handleUsage 用量聚合(/me/usage;BYOK 归属维度随 M4b 接入)。
func (s *Server) handleUsage(w http.ResponseWriter, _ *http.Request) {
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(s.usage.Aggregate())
}
// handleHealth 健康面。