feat(openai): 极致优化 OAuth 链路并补齐性能守护

- 优化 /v1/responses 热路径,减少重复解析与不必要拷贝\n- 优化并发与 token 竞争路径并补齐运行指标\n- 补充 OpenAI/Ops 相关单元测试与回归用例\n- 新增灰度阈值守护与压测脚本,支撑发布验收
This commit is contained in:
yangjianbo
2026-02-12 09:41:37 +08:00
parent a88bb8684f
commit 61a2bf469a
16 changed files with 1519 additions and 135 deletions

View File

@@ -4,16 +4,74 @@ import (
"context"
"errors"
"log/slog"
"math/rand/v2"
"strings"
"sync/atomic"
"time"
)
const (
openAITokenRefreshSkew = 3 * time.Minute
openAITokenCacheSkew = 5 * time.Minute
openAILockWaitTime = 200 * time.Millisecond
openAITokenRefreshSkew = 3 * time.Minute
openAITokenCacheSkew = 5 * time.Minute
openAILockInitialWait = 20 * time.Millisecond
openAILockMaxWait = 120 * time.Millisecond
openAILockMaxAttempts = 5
openAILockJitterRatio = 0.2
openAILockWarnThresholdMs = 250
)
// OpenAITokenRuntimeMetrics 表示 OpenAI token 刷新与锁竞争保护指标快照。
type OpenAITokenRuntimeMetrics struct {
RefreshRequests int64
RefreshSuccess int64
RefreshFailure int64
LockAcquireFailure int64
LockContention int64
LockWaitSamples int64
LockWaitTotalMs int64
LockWaitHit int64
LockWaitMiss int64
LastObservedUnixMs int64
}
type openAITokenRuntimeMetricsStore struct {
refreshRequests atomic.Int64
refreshSuccess atomic.Int64
refreshFailure atomic.Int64
lockAcquireFailure atomic.Int64
lockContention atomic.Int64
lockWaitSamples atomic.Int64
lockWaitTotalMs atomic.Int64
lockWaitHit atomic.Int64
lockWaitMiss atomic.Int64
lastObservedUnixMs atomic.Int64
}
func (m *openAITokenRuntimeMetricsStore) snapshot() OpenAITokenRuntimeMetrics {
if m == nil {
return OpenAITokenRuntimeMetrics{}
}
return OpenAITokenRuntimeMetrics{
RefreshRequests: m.refreshRequests.Load(),
RefreshSuccess: m.refreshSuccess.Load(),
RefreshFailure: m.refreshFailure.Load(),
LockAcquireFailure: m.lockAcquireFailure.Load(),
LockContention: m.lockContention.Load(),
LockWaitSamples: m.lockWaitSamples.Load(),
LockWaitTotalMs: m.lockWaitTotalMs.Load(),
LockWaitHit: m.lockWaitHit.Load(),
LockWaitMiss: m.lockWaitMiss.Load(),
LastObservedUnixMs: m.lastObservedUnixMs.Load(),
}
}
func (m *openAITokenRuntimeMetricsStore) touchNow() {
if m == nil {
return
}
m.lastObservedUnixMs.Store(time.Now().UnixMilli())
}
// OpenAITokenCache Token 缓存接口(复用 GeminiTokenCache 接口定义)
type OpenAITokenCache = GeminiTokenCache
@@ -22,6 +80,7 @@ type OpenAITokenProvider struct {
accountRepo AccountRepository
tokenCache OpenAITokenCache
openAIOAuthService *OpenAIOAuthService
metrics *openAITokenRuntimeMetricsStore
}
func NewOpenAITokenProvider(
@@ -33,11 +92,27 @@ func NewOpenAITokenProvider(
accountRepo: accountRepo,
tokenCache: tokenCache,
openAIOAuthService: openAIOAuthService,
metrics: &openAITokenRuntimeMetricsStore{},
}
}
func (p *OpenAITokenProvider) SnapshotRuntimeMetrics() OpenAITokenRuntimeMetrics {
if p == nil {
return OpenAITokenRuntimeMetrics{}
}
p.ensureMetrics()
return p.metrics.snapshot()
}
func (p *OpenAITokenProvider) ensureMetrics() {
if p != nil && p.metrics == nil {
p.metrics = &openAITokenRuntimeMetricsStore{}
}
}
// GetAccessToken 获取有效的 access_token
func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Account) (string, error) {
p.ensureMetrics()
if account == nil {
return "", errors.New("account is nil")
}
@@ -64,6 +139,8 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
needsRefresh := expiresAt == nil || time.Until(*expiresAt) <= openAITokenRefreshSkew
refreshFailed := false
if needsRefresh && p.tokenCache != nil {
p.metrics.refreshRequests.Add(1)
p.metrics.touchNow()
locked, lockErr := p.tokenCache.AcquireRefreshLock(ctx, cacheKey, 30*time.Second)
if lockErr == nil && locked {
defer func() { _ = p.tokenCache.ReleaseRefreshLock(ctx, cacheKey) }()
@@ -82,14 +159,17 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
if expiresAt == nil || time.Until(*expiresAt) <= openAITokenRefreshSkew {
if p.openAIOAuthService == nil {
slog.Warn("openai_oauth_service_not_configured", "account_id", account.ID)
p.metrics.refreshFailure.Add(1)
refreshFailed = true // 无法刷新,标记失败
} else {
tokenInfo, err := p.openAIOAuthService.RefreshAccountToken(ctx, account)
if err != nil {
// 刷新失败时记录警告,但不立即返回错误,尝试使用现有 token
slog.Warn("openai_token_refresh_failed", "account_id", account.ID, "error", err)
p.metrics.refreshFailure.Add(1)
refreshFailed = true // 刷新失败,标记以使用短 TTL
} else {
p.metrics.refreshSuccess.Add(1)
newCredentials := p.openAIOAuthService.BuildAccountCredentials(tokenInfo)
for k, v := range account.Credentials {
if _, exists := newCredentials[k]; !exists {
@@ -106,6 +186,8 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
}
} else if lockErr != nil {
// Redis 错误导致无法获取锁,降级为无锁刷新(仅在 token 接近过期时)
p.metrics.lockAcquireFailure.Add(1)
p.metrics.touchNow()
slog.Warn("openai_token_lock_failed_degraded_refresh", "account_id", account.ID, "error", lockErr)
// 检查 ctx 是否已取消
@@ -126,13 +208,16 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
if expiresAt == nil || time.Until(*expiresAt) <= openAITokenRefreshSkew {
if p.openAIOAuthService == nil {
slog.Warn("openai_oauth_service_not_configured", "account_id", account.ID)
p.metrics.refreshFailure.Add(1)
refreshFailed = true
} else {
tokenInfo, err := p.openAIOAuthService.RefreshAccountToken(ctx, account)
if err != nil {
slog.Warn("openai_token_refresh_failed_degraded", "account_id", account.ID, "error", err)
p.metrics.refreshFailure.Add(1)
refreshFailed = true
} else {
p.metrics.refreshSuccess.Add(1)
newCredentials := p.openAIOAuthService.BuildAccountCredentials(tokenInfo)
for k, v := range account.Credentials {
if _, exists := newCredentials[k]; !exists {
@@ -148,9 +233,14 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
}
}
} else {
// 锁获取失败(被其他 worker 持有),等待 200ms 后重试读取缓存
time.Sleep(openAILockWaitTime)
if token, err := p.tokenCache.GetAccessToken(ctx, cacheKey); err == nil && strings.TrimSpace(token) != "" {
// 锁被其他 worker 持有:使用短轮询+jitter降低固定等待导致的尾延迟台阶。
p.metrics.lockContention.Add(1)
p.metrics.touchNow()
token, waitErr := p.waitForTokenAfterLockRace(ctx, cacheKey)
if waitErr != nil {
return "", waitErr
}
if strings.TrimSpace(token) != "" {
slog.Debug("openai_token_cache_hit_after_wait", "account_id", account.ID)
return token, nil
}
@@ -198,3 +288,64 @@ func (p *OpenAITokenProvider) GetAccessToken(ctx context.Context, account *Accou
return accessToken, nil
}
func (p *OpenAITokenProvider) waitForTokenAfterLockRace(ctx context.Context, cacheKey string) (string, error) {
wait := openAILockInitialWait
totalWaitMs := int64(0)
for i := 0; i < openAILockMaxAttempts; i++ {
actualWait := jitterLockWait(wait)
timer := time.NewTimer(actualWait)
select {
case <-ctx.Done():
if !timer.Stop() {
select {
case <-timer.C:
default:
}
}
return "", ctx.Err()
case <-timer.C:
}
waitMs := actualWait.Milliseconds()
if waitMs < 0 {
waitMs = 0
}
totalWaitMs += waitMs
p.metrics.lockWaitSamples.Add(1)
p.metrics.lockWaitTotalMs.Add(waitMs)
p.metrics.touchNow()
token, err := p.tokenCache.GetAccessToken(ctx, cacheKey)
if err == nil && strings.TrimSpace(token) != "" {
p.metrics.lockWaitHit.Add(1)
if totalWaitMs >= openAILockWarnThresholdMs {
slog.Warn("openai_token_lock_wait_high", "wait_ms", totalWaitMs, "attempts", i+1)
}
return token, nil
}
if wait < openAILockMaxWait {
wait *= 2
if wait > openAILockMaxWait {
wait = openAILockMaxWait
}
}
}
p.metrics.lockWaitMiss.Add(1)
if totalWaitMs >= openAILockWarnThresholdMs {
slog.Warn("openai_token_lock_wait_high", "wait_ms", totalWaitMs, "attempts", openAILockMaxAttempts)
}
return "", nil
}
func jitterLockWait(base time.Duration) time.Duration {
if base <= 0 {
return 0
}
minFactor := 1 - openAILockJitterRatio
maxFactor := 1 + openAILockJitterRatio
factor := minFactor + rand.Float64()*(maxFactor-minFactor)
return time.Duration(float64(base) * factor)
}