Files
easyai-ai-gateway/apps/api/internal/securityevents/metrics.go
T
wangbo c28bf74230 feat(worker): 实现集群限流与自适应负载
保留平台模型 RPM、TPM 和并发策略语义,增加 PostgreSQL 集群级租约、饱和候选重选和多平台自动负载,避免突发任务固定等待首个平台。\n\n新增 Worker 实时负载采样、自适应 active/heavy 容量、心跳与管理端指标,并扩展本地 acceptance runner,覆盖三 Worker、同模型三平台 2/4/6 并发和 48 个带图视频突发任务。\n\n验证:go test ./...、go vet ./...、PostgreSQL 跨 Store 集成测试、gofmt、bash -n、ShellCheck 及本地集群 provider-burst 验收通过;48/48 成功,无越限、重复提交、重复计费、重复回调或终态资源泄漏。
2026-08-03 00:13:46 +08:00

574 lines
27 KiB
Go

package securityevents
import (
"context"
"errors"
"fmt"
"net/http"
"strings"
"sync/atomic"
"time"
"github.com/easyai/easyai-ai-gateway/apps/api/internal/store"
)
type MetricsSnapshotProvider interface {
SecurityEventMetrics(context.Context, string, string) (string, *time.Time, error)
}
type DynamicMetricsSnapshotProvider interface {
MetricsSnapshotProvider
SecurityEventConnection(context.Context) (store.SecurityEventConnection, error)
BillingMetrics(context.Context) (store.BillingMetricsSnapshot, error)
}
type Metrics struct {
accepted atomic.Uint64
rejected atomic.Uint64
duplicate atomic.Uint64
sessionsDeleted atomic.Uint64
watermarkRejected atomic.Uint64
verificationAccepted atomic.Uint64
heartbeatAccepted atomic.Uint64
heartbeatFailed atomic.Uint64
introspectionActive atomic.Uint64
introspectionInactive atomic.Uint64
introspectionFailed atomic.Uint64
jwksSSFFailed atomic.Uint64
jwksOIDCFailed atomic.Uint64
processingCount atomic.Uint64
processingNanos atomic.Uint64
processingBuckets [6]atomic.Uint64
billingSettlementCompleted atomic.Uint64
billingSettlementRetry atomic.Uint64
billingManualReview atomic.Uint64
billingEstimateFailed atomic.Uint64
billingIdempotentReplay atomic.Uint64
billingPricingUnavailable atomic.Uint64
asyncWorkerCapacity atomic.Int64
asyncWorkerDesiredCapacity atomic.Int64
asyncWorkerHardLimit atomic.Int64
asyncWorkerCapacityCapped atomic.Int64
asyncWorkerActiveInstances atomic.Int64
asyncWorkerGlobalCapacity atomic.Int64
asyncWorkerAllocated atomic.Int64
workerSafeCapacity atomic.Int64
workerHeavyCapacity atomic.Int64
workerActiveTasks atomic.Int64
workerPreparingTasks atomic.Int64
workerWaitingTasks atomic.Int64
workerFinalizingTasks atomic.Int64
workerPressureState atomic.Int64
providerQuotaWaitRPM atomic.Uint64
providerQuotaWaitTPM atomic.Uint64
providerQuotaWaitConcurrent atomic.Uint64
providerQuotaWaitOther atomic.Uint64
asyncWorkerResizeSuccess atomic.Uint64
asyncWorkerRefreshFailed atomic.Uint64
asyncWorkerCreateFailed atomic.Uint64
asyncWorkerStartFailed atomic.Uint64
leaseRenewalSuccess atomic.Uint64
leaseRenewalFailure atomic.Uint64
leaseRenewalLost atomic.Uint64
taskEventDuplicate atomic.Uint64
taskEventUnknownType atomic.Uint64
taskEventBudgetExceeded atomic.Uint64
taskAdmissionAdmitted atomic.Uint64
taskAdmissionQueueFull atomic.Uint64
taskAdmissionTimeout atomic.Uint64
taskAdmissionCancelled atomic.Uint64
taskAdmissionExpired atomic.Uint64
taskAdmissionMigrated atomic.Uint64
taskAdmissionWaitBuckets [11]atomic.Uint64
taskAdmissionWaitMicros atomic.Uint64
}
var processingDurationBounds = [...]time.Duration{
10 * time.Millisecond, 50 * time.Millisecond, 100 * time.Millisecond,
500 * time.Millisecond, time.Second, 3 * time.Second,
}
var taskAdmissionWaitBounds = [...]time.Duration{
time.Second,
5 * time.Second,
15 * time.Second,
30 * time.Second,
time.Minute,
2 * time.Minute,
5 * time.Minute,
10 * time.Minute,
30 * time.Minute,
time.Hour,
}
func (m *Metrics) ObserveReceipt(outcome string, duration time.Duration, sessionsDeleted int64) {
switch outcome {
case "accepted":
m.accepted.Add(1)
case "duplicate":
m.duplicate.Add(1)
default:
m.rejected.Add(1)
}
if sessionsDeleted > 0 {
m.sessionsDeleted.Add(uint64(sessionsDeleted))
}
if duration < 0 {
duration = 0
}
m.processingCount.Add(1)
m.processingNanos.Add(uint64(duration))
for index, bound := range processingDurationBounds {
if duration <= bound {
m.processingBuckets[index].Add(1)
}
}
}
func (m *Metrics) ObserveWatermarkRejection() { m.watermarkRejected.Add(1) }
func (m *Metrics) ObserveVerificationAccepted() { m.verificationAccepted.Add(1) }
func (m *Metrics) ObserveHeartbeat(outcome string) {
if outcome == "accepted" {
m.heartbeatAccepted.Add(1)
return
}
m.heartbeatFailed.Add(1)
}
func (m *Metrics) ObserveIntrospection(outcome string) {
switch outcome {
case "active":
m.introspectionActive.Add(1)
case "inactive":
m.introspectionInactive.Add(1)
default:
m.introspectionFailed.Add(1)
}
}
func (m *Metrics) ObserveJWKSRefreshFailure(source string) {
if source == "ssf" {
m.jwksSSFFailed.Add(1)
return
}
m.jwksOIDCFailed.Add(1)
}
func (m *Metrics) ObserveBillingEvent(event string) {
switch event {
case "settlement_completed":
m.billingSettlementCompleted.Add(1)
case "settlement_retry":
m.billingSettlementRetry.Add(1)
case "manual_review":
m.billingManualReview.Add(1)
case "estimate_failed":
m.billingEstimateFailed.Add(1)
case "idempotent_replay":
m.billingIdempotentReplay.Add(1)
case "pricing_unavailable":
m.billingPricingUnavailable.Add(1)
}
}
func (m *Metrics) SetAsyncWorkerCapacity(current, desired, hardLimit int, capped bool) {
m.asyncWorkerCapacity.Store(int64(current))
m.asyncWorkerDesiredCapacity.Store(int64(desired))
m.asyncWorkerHardLimit.Store(int64(hardLimit))
cappedValue := int64(0)
if capped {
cappedValue = 1
}
m.asyncWorkerCapacityCapped.Store(cappedValue)
}
func (m *Metrics) SetDistributedWorkerCapacity(activeInstances, globalCapacity, allocatedCapacity int) {
m.asyncWorkerActiveInstances.Store(int64(activeInstances))
m.asyncWorkerGlobalCapacity.Store(int64(globalCapacity))
m.asyncWorkerAllocated.Store(int64(allocatedCapacity))
}
func (m *Metrics) SetWorkerLoad(activeLimit, heavyLimit, active, preparing, waiting, finalizing int, pressure string) {
m.workerSafeCapacity.Store(int64(activeLimit))
m.workerHeavyCapacity.Store(int64(heavyLimit))
m.workerActiveTasks.Store(int64(active))
m.workerPreparingTasks.Store(int64(preparing))
m.workerWaitingTasks.Store(int64(waiting))
m.workerFinalizingTasks.Store(int64(finalizing))
state := int64(-1)
switch pressure {
case "normal":
state = 0
case "busy":
state = 1
case "critical":
state = 2
}
m.workerPressureState.Store(state)
}
func (m *Metrics) ObserveProviderQuotaWait(metric string) {
switch metric {
case "rpm":
m.providerQuotaWaitRPM.Add(1)
case "tpm", "tpm_total", "tpm_input", "tpm_output":
m.providerQuotaWaitTPM.Add(1)
case "concurrent":
m.providerQuotaWaitConcurrent.Add(1)
default:
m.providerQuotaWaitOther.Add(1)
}
}
func (m *Metrics) ObserveTaskAdmission(event string) {
switch event {
case "admitted":
m.taskAdmissionAdmitted.Add(1)
case "queue_full":
m.taskAdmissionQueueFull.Add(1)
case "timeout":
m.taskAdmissionTimeout.Add(1)
case "cancelled":
m.taskAdmissionCancelled.Add(1)
case "expired":
m.taskAdmissionExpired.Add(1)
case "candidate_migrated":
m.taskAdmissionMigrated.Add(1)
}
}
func (m *Metrics) ObserveTaskAdmissionWait(wait time.Duration) {
if wait < 0 {
wait = 0
}
for index, bound := range taskAdmissionWaitBounds {
if wait <= bound {
m.taskAdmissionWaitBuckets[index].Add(1)
}
}
m.taskAdmissionWaitBuckets[len(m.taskAdmissionWaitBuckets)-1].Add(1)
m.taskAdmissionWaitMicros.Add(uint64(wait / time.Microsecond))
}
func (m *Metrics) ObserveAsyncWorkerResize(outcome string) {
switch outcome {
case "success":
m.asyncWorkerResizeSuccess.Add(1)
case "refresh_failed":
m.asyncWorkerRefreshFailed.Add(1)
case "create_failed":
m.asyncWorkerCreateFailed.Add(1)
case "start_failed":
m.asyncWorkerStartFailed.Add(1)
}
}
func (m *Metrics) ObserveConcurrencyLeaseRenewal(outcome string) {
switch outcome {
case "success":
m.leaseRenewalSuccess.Add(1)
case "lost":
m.leaseRenewalLost.Add(1)
default:
m.leaseRenewalFailure.Add(1)
}
}
func (m *Metrics) ObserveTaskEventSkip(reason string) {
switch reason {
case "duplicate":
m.taskEventDuplicate.Add(1)
case "budget_exceeded":
m.taskEventBudgetExceeded.Add(1)
default:
m.taskEventUnknownType.Add(1)
}
}
func (m *Metrics) Handler(provider MetricsSnapshotProvider, issuer, audience string, enabled bool) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
mode := "disabled"
var lastVerificationAt *time.Time
if enabled {
ctx, cancel := context.WithTimeout(r.Context(), 2*time.Second)
defer cancel()
var err error
mode, lastVerificationAt, err = provider.SecurityEventMetrics(ctx, issuer, audience)
if err != nil {
http.Error(w, "metrics unavailable", http.StatusServiceUnavailable)
return
}
}
billing := store.BillingMetricsSnapshot{}
if billingProvider, ok := provider.(interface {
BillingMetrics(context.Context) (store.BillingMetricsSnapshot, error)
}); ok {
var err error
billing, err = billingProvider.BillingMetrics(r.Context())
if err != nil {
http.Error(w, "metrics unavailable", http.StatusServiceUnavailable)
return
}
}
admission := store.TaskAdmissionMetricsSnapshot{}
if admissionProvider, ok := provider.(interface {
TaskAdmissionMetrics(context.Context) (store.TaskAdmissionMetricsSnapshot, error)
}); ok {
var err error
admission, err = admissionProvider.TaskAdmissionMetrics(r.Context())
if err != nil {
http.Error(w, "metrics unavailable", http.StatusServiceUnavailable)
return
}
}
postgresPool := store.PostgresPoolMetricsSnapshot{}
if poolProvider, ok := provider.(interface {
PostgresPoolMetrics() store.PostgresPoolMetricsSnapshot
}); ok {
postgresPool = poolProvider.PostgresPoolMetrics()
}
criticalPostgresPool := store.PostgresPoolMetricsSnapshot{}
if poolProvider, ok := provider.(interface {
CriticalPostgresPoolMetrics() store.PostgresPoolMetricsSnapshot
}); ok {
criticalPostgresPool = poolProvider.CriticalPostgresPoolMetrics()
}
riverPostgresPool := store.PostgresPoolMetricsSnapshot{}
if poolProvider, ok := provider.(interface {
RiverPostgresPoolMetrics() store.PostgresPoolMetricsSnapshot
}); ok {
riverPostgresPool = poolProvider.RiverPostgresPoolMetrics()
}
modelRateLimits := []store.ModelRateLimitStatus{}
if rateLimitProvider, ok := provider.(interface {
ListModelRateLimitStatuses(context.Context) ([]store.ModelRateLimitStatus, error)
}); ok {
var err error
modelRateLimits, err = rateLimitProvider.ListModelRateLimitStatuses(r.Context())
if err != nil {
http.Error(w, "metrics unavailable", http.StatusServiceUnavailable)
return
}
}
w.Header().Set("Content-Type", "text/plain; version=0.0.4; charset=utf-8")
outcomeCounters(w, "easyai_gateway_ssf_receipts_total", "Received SETs by bounded outcome.", []outcomeValue{
{"accepted", m.accepted.Load()}, {"rejected", m.rejected.Load()}, {"duplicate", m.duplicate.Load()},
})
plainCounter(w, "easyai_gateway_ssf_sessions_deleted_total", "OIDC browser sessions deleted by accepted session-revoked events.", m.sessionsDeleted.Load())
plainCounter(w, "easyai_gateway_ssf_watermark_rejections_total", "OIDC access tokens rejected by a local revocation watermark.", m.watermarkRejected.Load())
plainCounter(w, "easyai_gateway_ssf_verifications_total", "Valid verification SETs accepted by the receiver.", m.verificationAccepted.Load())
outcomeCounters(w, "easyai_gateway_ssf_heartbeat_requests_total", "Verification requests by bounded outcome.", []outcomeValue{
{"accepted", m.heartbeatAccepted.Load()}, {"failed", m.heartbeatFailed.Load()},
})
outcomeCounters(w, "easyai_gateway_oidc_introspection_total", "OIDC introspection requests by bounded outcome.", []outcomeValue{
{"active", m.introspectionActive.Load()}, {"inactive", m.introspectionInactive.Load()}, {"failed", m.introspectionFailed.Load()},
})
outcomeCounters(w, "easyai_gateway_jwks_refresh_failures_total", "JWKS refresh failures by bounded source.", []outcomeValue{
{"ssf", m.jwksSSFFailed.Load()}, {"oidc", m.jwksOIDCFailed.Load()},
})
fmt.Fprintln(w, "# HELP easyai_gateway_ssf_processing_duration_seconds SSF receiver transaction processing time.")
fmt.Fprintln(w, "# TYPE easyai_gateway_ssf_processing_duration_seconds histogram")
for index, bound := range processingDurationBounds {
fmt.Fprintf(w, "easyai_gateway_ssf_processing_duration_seconds_bucket{le=\"%.2f\"} %d\n", bound.Seconds(), m.processingBuckets[index].Load())
}
fmt.Fprintf(w, "easyai_gateway_ssf_processing_duration_seconds_bucket{le=\"+Inf\"} %d\n", m.processingCount.Load())
fmt.Fprintf(w, "easyai_gateway_ssf_processing_duration_seconds_sum %.6f\n", float64(m.processingNanos.Load())/float64(time.Second))
fmt.Fprintf(w, "easyai_gateway_ssf_processing_duration_seconds_count %d\n", m.processingCount.Load())
verificationAge := float64(0)
if lastVerificationAt != nil {
verificationAge = time.Since(*lastVerificationAt).Seconds()
if verificationAge < 0 {
verificationAge = 0
}
}
fmt.Fprintln(w, "# HELP easyai_gateway_ssf_verification_age_seconds Seconds since the last matched verification SET.")
fmt.Fprintln(w, "# TYPE easyai_gateway_ssf_verification_age_seconds gauge")
fmt.Fprintf(w, "easyai_gateway_ssf_verification_age_seconds %.6f\n", verificationAge)
fmt.Fprintln(w, "# HELP easyai_gateway_ssf_mode Current SSF receiver mode as a one-hot gauge.")
fmt.Fprintln(w, "# TYPE easyai_gateway_ssf_mode gauge")
for _, modeName := range []string{"disabled", "bootstrap", "push_healthy", "introspection_fallback"} {
value := 0
if mode == modeName {
value = 1
}
fmt.Fprintf(w, "easyai_gateway_ssf_mode{mode=\"%s\"} %d\n", modeName, value)
}
fmt.Fprintln(w, "# HELP easyai_gateway_billing_settlement_backlog Current unsettled Outbox records.")
fmt.Fprintln(w, "# TYPE easyai_gateway_billing_settlement_backlog gauge")
fmt.Fprintf(w, "easyai_gateway_billing_settlement_backlog %d\n", billing.SettlementBacklog)
fmt.Fprintln(w, "# HELP easyai_gateway_billing_settlement_delay_seconds Age of the oldest unsettled Outbox record.")
fmt.Fprintln(w, "# TYPE easyai_gateway_billing_settlement_delay_seconds gauge")
fmt.Fprintf(w, "easyai_gateway_billing_settlement_delay_seconds %.6f\n", billing.SettlementDelaySecs)
fmt.Fprintln(w, "# HELP easyai_gateway_billing_manual_review Current billing records requiring manual review.")
fmt.Fprintln(w, "# TYPE easyai_gateway_billing_manual_review gauge")
fmt.Fprintf(w, "easyai_gateway_billing_manual_review %d\n", billing.ManualReview)
fmt.Fprintln(w, "# HELP easyai_gateway_billing_orphan_frozen Wallets whose frozen amount differs from active reservations.")
fmt.Fprintln(w, "# TYPE easyai_gateway_billing_orphan_frozen gauge")
fmt.Fprintf(w, "easyai_gateway_billing_orphan_frozen %d\n", billing.OrphanFrozen)
fmt.Fprintln(w, "# HELP easyai_gateway_billing_pricing_unavailable_tasks Current tasks rejected because no effective price was available.")
fmt.Fprintln(w, "# TYPE easyai_gateway_billing_pricing_unavailable_tasks gauge")
fmt.Fprintf(w, "easyai_gateway_billing_pricing_unavailable_tasks %d\n", billing.PricingUnavailable)
plainCounter(w, "easyai_gateway_billing_settlements_completed_total", "Billing settlements completed by this process.", m.billingSettlementCompleted.Load())
plainCounter(w, "easyai_gateway_billing_settlement_retries_total", "Billing settlement retries scheduled by this process.", m.billingSettlementRetry.Load())
plainCounter(w, "easyai_gateway_billing_manual_review_transitions_total", "Billing records transitioned to manual review by this process.", m.billingManualReview.Load())
plainCounter(w, "easyai_gateway_billing_estimate_failures_total", "Pricing estimate requests that failed.", m.billingEstimateFailed.Load())
plainCounter(w, "easyai_gateway_billing_idempotent_replays_total", "Generation requests replayed idempotently.", m.billingIdempotentReplay.Load())
plainCounter(w, "easyai_gateway_billing_pricing_unavailable_total", "Pricing requests rejected because no effective price was available.", m.billingPricingUnavailable.Load())
plainGauge(w, "easyai_gateway_async_worker_capacity", "Current River asynchronous worker capacity.", m.asyncWorkerCapacity.Load())
plainGauge(w, "easyai_gateway_async_worker_desired_capacity", "Uncapped asynchronous worker capacity derived from model and user-group policies.", m.asyncWorkerDesiredCapacity.Load())
plainGauge(w, "easyai_gateway_async_worker_hard_limit", "Per-process asynchronous worker safety limit.", m.asyncWorkerHardLimit.Load())
plainGauge(w, "easyai_gateway_async_worker_capacity_capped", "Whether desired asynchronous worker capacity is capped by the hard limit.", m.asyncWorkerCapacityCapped.Load())
plainGauge(w, "easyai_gateway_worker_active_instances", "Active distributed worker instances with a fresh heartbeat.", m.asyncWorkerActiveInstances.Load())
plainGauge(w, "easyai_gateway_worker_global_capacity", "Global asynchronous execution capacity before instance allocation.", m.asyncWorkerGlobalCapacity.Load())
plainGauge(w, "easyai_gateway_worker_allocated_capacity", "Asynchronous execution capacity allocated to this worker instance.", m.asyncWorkerAllocated.Load())
plainGauge(w, "easyai_gateway_worker_safe_capacity", "Locally resource-safe asynchronous task capacity.", m.workerSafeCapacity.Load())
plainGauge(w, "easyai_gateway_worker_heavy_capacity", "Locally resource-safe preparing and finalizing capacity.", m.workerHeavyCapacity.Load())
plainGauge(w, "easyai_gateway_worker_active_tasks", "Tasks currently owned by this Worker process.", m.workerActiveTasks.Load())
plainGauge(w, "easyai_gateway_worker_preparing_tasks", "Tasks in the local preparing phase.", m.workerPreparingTasks.Load())
plainGauge(w, "easyai_gateway_worker_waiting_upstream_tasks", "Tasks waiting for an upstream result.", m.workerWaitingTasks.Load())
plainGauge(w, "easyai_gateway_worker_finalizing_tasks", "Tasks in the local finalizing phase.", m.workerFinalizingTasks.Load())
plainGauge(w, "easyai_gateway_worker_pressure_state", "Local Worker pressure state: -1 unknown, 0 normal, 1 busy, 2 critical.", m.workerPressureState.Load())
outcomeCounters(w, "easyai_gateway_provider_quota_waits_total", "Tasks delayed before an upstream call by a cluster-wide provider quota.", []outcomeValue{
{"rpm", m.providerQuotaWaitRPM.Load()},
{"tpm", m.providerQuotaWaitTPM.Load()},
{"concurrent", m.providerQuotaWaitConcurrent.Load()},
{"other", m.providerQuotaWaitOther.Load()},
})
platformModelRateLimitUtilizationGauges(w, modelRateLimits)
plainGauge(w, "easyai_gateway_postgres_pool_max_connections", "Maximum PostgreSQL connections in this process pool.", int64(postgresPool.MaxConnections))
plainGauge(w, "easyai_gateway_postgres_pool_total_connections", "Current PostgreSQL connections in this process pool.", int64(postgresPool.TotalConnections))
plainGauge(w, "easyai_gateway_postgres_pool_acquired_connections", "Currently acquired PostgreSQL connections in this process pool.", int64(postgresPool.AcquiredConnections))
plainGauge(w, "easyai_gateway_postgres_pool_idle_connections", "Currently idle PostgreSQL connections in this process pool.", int64(postgresPool.IdleConnections))
plainCounter(w, "easyai_gateway_postgres_pool_empty_acquire_total", "Pool acquires that had to wait for a PostgreSQL connection.", uint64(max(postgresPool.EmptyAcquireCount, 0)))
plainCounter(w, "easyai_gateway_postgres_pool_canceled_acquire_total", "Canceled PostgreSQL pool acquires.", uint64(max(postgresPool.CanceledAcquireCount, 0)))
plainGauge(w, "easyai_gateway_postgres_critical_pool_max_connections", "Maximum PostgreSQL connections reserved for leases and coordination.", int64(criticalPostgresPool.MaxConnections))
plainGauge(w, "easyai_gateway_postgres_critical_pool_total_connections", "Current PostgreSQL connections in the leases and coordination pool.", int64(criticalPostgresPool.TotalConnections))
plainGauge(w, "easyai_gateway_postgres_critical_pool_acquired_connections", "Currently acquired PostgreSQL connections in the leases and coordination pool.", int64(criticalPostgresPool.AcquiredConnections))
plainGauge(w, "easyai_gateway_postgres_critical_pool_idle_connections", "Currently idle PostgreSQL connections in the leases and coordination pool.", int64(criticalPostgresPool.IdleConnections))
plainCounter(w, "easyai_gateway_postgres_critical_pool_empty_acquire_total", "Coordination pool acquires that had to wait for a PostgreSQL connection.", uint64(max(criticalPostgresPool.EmptyAcquireCount, 0)))
plainCounter(w, "easyai_gateway_postgres_critical_pool_canceled_acquire_total", "Canceled PostgreSQL coordination pool acquires.", uint64(max(criticalPostgresPool.CanceledAcquireCount, 0)))
plainGauge(w, "easyai_gateway_postgres_river_pool_max_connections", "Maximum PostgreSQL connections reserved for River queue coordination.", int64(riverPostgresPool.MaxConnections))
plainGauge(w, "easyai_gateway_postgres_river_pool_total_connections", "Current PostgreSQL connections in the River queue pool.", int64(riverPostgresPool.TotalConnections))
plainGauge(w, "easyai_gateway_postgres_river_pool_acquired_connections", "Currently acquired PostgreSQL connections in the River queue pool.", int64(riverPostgresPool.AcquiredConnections))
plainGauge(w, "easyai_gateway_postgres_river_pool_idle_connections", "Currently idle PostgreSQL connections in the River queue pool.", int64(riverPostgresPool.IdleConnections))
plainCounter(w, "easyai_gateway_postgres_river_pool_empty_acquire_total", "River queue pool acquires that had to wait for a PostgreSQL connection.", uint64(max(riverPostgresPool.EmptyAcquireCount, 0)))
plainCounter(w, "easyai_gateway_postgres_river_pool_canceled_acquire_total", "Canceled PostgreSQL River queue pool acquires.", uint64(max(riverPostgresPool.CanceledAcquireCount, 0)))
outcomeCounters(w, "easyai_gateway_async_worker_resizes_total", "Asynchronous worker resize attempts by bounded outcome.", []outcomeValue{
{"success", m.asyncWorkerResizeSuccess.Load()},
{"refresh_failed", m.asyncWorkerRefreshFailed.Load()},
{"create_failed", m.asyncWorkerCreateFailed.Load()},
{"start_failed", m.asyncWorkerStartFailed.Load()},
})
outcomeCounters(w, "easyai_gateway_concurrency_lease_renewals_total", "Concurrency lease renewals by bounded outcome.", []outcomeValue{
{"success", m.leaseRenewalSuccess.Load()},
{"failure", m.leaseRenewalFailure.Load()},
{"lost", m.leaseRenewalLost.Load()},
})
outcomeCounters(w, "easyai_gateway_task_events_skipped_total", "Task events skipped before persistence by bounded reason.", []outcomeValue{
{"duplicate", m.taskEventDuplicate.Load()},
{"unknown_type", m.taskEventUnknownType.Load()},
{"budget_exceeded", m.taskEventBudgetExceeded.Load()},
})
plainGauge(w, "easyai_gateway_task_admission_queue_depth", "Current persistent non-text task admission queue depth.", int64(admission.QueueDepth))
plainGauge(w, "easyai_gateway_task_admission_waiting_sync", "Current synchronous admission waiters.", int64(admission.WaitingSync))
plainGauge(w, "easyai_gateway_task_admission_waiting_async", "Current asynchronous admission waiters.", int64(admission.WaitingAsync))
plainFloatGauge(w, "easyai_gateway_task_admission_oldest_wait_seconds", "Age of the oldest persistent admission waiter.", admission.OldestWaitSeconds)
plainGauge(w, "easyai_gateway_task_admission_expired_waiter_backlog", "Expired synchronous waiter leases awaiting reclamation.", int64(admission.ExpiredWaiterBacklog))
plainGauge(w, "easyai_gateway_task_admission_expired_deadline_backlog", "Expired queue deadlines awaiting reclamation.", int64(admission.ExpiredDeadlineBacklog))
outcomeCounters(w, "easyai_gateway_task_admissions_total", "Task admission transitions by bounded outcome.", []outcomeValue{
{"admitted", m.taskAdmissionAdmitted.Load()},
{"queue_full", m.taskAdmissionQueueFull.Load()},
{"timeout", m.taskAdmissionTimeout.Load()},
{"cancelled", m.taskAdmissionCancelled.Load()},
{"expired", m.taskAdmissionExpired.Load()},
{"candidate_migrated", m.taskAdmissionMigrated.Load()},
})
taskAdmissionWaitHistogram(w, m)
})
}
func (m *Metrics) DynamicHandler(provider DynamicMetricsSnapshotProvider) http.Handler {
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
connection, err := provider.SecurityEventConnection(r.Context())
if errors.Is(err, store.ErrSecurityEventConnectionNotFound) {
m.Handler(provider, "", "", false).ServeHTTP(w, r)
return
}
if err != nil {
http.Error(w, "metrics unavailable", http.StatusServiceUnavailable)
return
}
if connection.Audience == nil {
m.Handler(provider, "", "", false).ServeHTTP(w, r)
return
}
m.Handler(provider, connection.TransmitterIssuer, *connection.Audience, true).ServeHTTP(w, r)
})
}
type outcomeValue struct {
name string
value uint64
}
func outcomeCounters(w http.ResponseWriter, name, help string, values []outcomeValue) {
fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s counter\n", name, help, name)
for _, value := range values {
fmt.Fprintf(w, "%s{outcome=\"%s\"} %d\n", name, value.name, value.value)
}
}
func plainCounter(w http.ResponseWriter, name, help string, value uint64) {
fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s counter\n%s %d\n", name, help, name, name, value)
}
func plainGauge(w http.ResponseWriter, name, help string, value int64) {
fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s gauge\n%s %d\n", name, help, name, name, value)
}
func plainFloatGauge(w http.ResponseWriter, name, help string, value float64) {
fmt.Fprintf(w, "# HELP %s %s\n# TYPE %s gauge\n%s %.6f\n", name, help, name, name, value)
}
func platformModelRateLimitUtilizationGauges(w http.ResponseWriter, statuses []store.ModelRateLimitStatus) {
const name = "easyai_gateway_platform_model_rate_limit_utilization"
fmt.Fprintf(w, "# HELP %s Cluster-wide platform model quota utilization.\n# TYPE %s gauge\n", name, name)
for _, status := range statuses {
platformModelID := escapePrometheusLabel(status.PlatformModelID)
for _, metric := range []struct {
name string
ratio float64
}{
{name: "rpm", ratio: status.RPM.Ratio},
{name: "tpm", ratio: status.TPM.Ratio},
{name: "concurrent", ratio: status.Concurrent.Ratio},
} {
fmt.Fprintf(w, "%s{platform_model_id=\"%s\",metric=\"%s\"} %.6f\n", name, platformModelID, metric.name, metric.ratio)
}
}
}
func escapePrometheusLabel(value string) string {
return strings.NewReplacer("\\", "\\\\", "\n", "\\n", "\"", "\\\"").Replace(value)
}
func taskAdmissionWaitHistogram(w http.ResponseWriter, metrics *Metrics) {
const name = "easyai_gateway_task_admission_wait_seconds"
fmt.Fprintf(w, "# HELP %s Time spent waiting for persistent task admission.\n# TYPE %s histogram\n", name, name)
for index, bound := range taskAdmissionWaitBounds {
fmt.Fprintf(
w,
"%s_bucket{le=\"%g\"} %d\n",
name,
bound.Seconds(),
metrics.taskAdmissionWaitBuckets[index].Load(),
)
}
count := metrics.taskAdmissionWaitBuckets[len(metrics.taskAdmissionWaitBuckets)-1].Load()
fmt.Fprintf(w, "%s_bucket{le=\"+Inf\"} %d\n", name, count)
fmt.Fprintf(w, "%s_sum %.6f\n", name, float64(metrics.taskAdmissionWaitMicros.Load())/1_000_000)
fmt.Fprintf(w, "%s_count %d\n", name, count)
}