xiner/Edu
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(push-gateway,msg): redis pubsub backplane for real-time notifications

Browse Source 
M7: ADR-040 Redis Pub/Sub as state routing backplane

- push-gateway: remove Kafka consumer, add SSE endpoint

- SSE: subscribe to Redis user:{userId}:notify on connect

- msg: publish notifications to Redis Pub/Sub instead of HTTP push

- docker-compose: remove Kafka env from push-gateway
This commit is contained in:
SpecialX
2026-07-15 01:28:55 +08:00
parent a75527be80
commit 6af1aa0d82
14 changed files with 402 additions and 484 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
services/push-gateway/internal/health/readyz.go
View File

@@ -6,18 +6,19 @@
//
// /readyz reports readiness based on downstream dependency health. Per
// ARB-015 §17.4 / ISSUE-058 the readiness probe uses SOFT FAILURE semantics:
// when Redis or Kafka is unavailable the endpoint still returns HTTP 200 but
// carries `degraded: true` plus a `dependencies` block describing which
// component failed. This prevents Kubernetes from evicting the pod when a
// transient dependency blip occurs, at the cost of accepting some degraded
// behavior (no cross-instance fanout, no Kafka consumption) during the blip.
// Only the Hub being in a shutting-down state returns a non-200 (503).
// when Redis is unavailable the endpoint still returns HTTP 200 but carries
// `degraded: true` plus a `dependencies` block describing which component
// failed. This prevents Kubernetes from evicting the pod when a transient
// dependency blip occurs, at the cost of accepting some degraded behavior
// (no Redis Pub/Sub fanout, no SSE delivery) during the blip. Only the Hub
// being in a shutting-down state returns a non-200 (503).
//
// M7 (ADR-040): Kafka probe removed — push-gateway no longer mounts Kafka.
//
// The hard-failure case is limited to:
// - Hub.closing == true (process is shutting down) -> 503
// - Internal misconfiguration (both Redis and Kafka missing in non-DevMode)
// -> still 200 + degraded, since the pod can still serve local WebSocket
// traffic.
// - Redis missing in non-DevMode -> still 200 + degraded, since the pod can
// still serve local WebSocket traffic.
package health
import (
@@ -26,7 +27,6 @@ import (
"time"
"github.com/edu-cloud/push-gateway/internal/hub"
"github.com/edu-cloud/push-gateway/internal/kafkaconsumer"
"github.com/edu-cloud/push-gateway/internal/redisclient"
"github.com/gin-gonic/gin"
)
@@ -47,12 +47,12 @@ type dependencyStatus struct {
// readyzResponse is the ActionState-shaped envelope returned by /readyz.
// `degraded` is true when at least one non-critical dependency is unhealthy.
type readyzResponse struct {
Status string `json:"status"`
Service string `json:"service"`
InstanceID string `json:"instance_id"`
Degraded bool `json:"degraded"`
Connections int `json:"connections"`
Users int `json:"users"`
Status string `json:"status"`
Service string `json:"service"`
InstanceID string `json:"instance_id"`
Degraded bool `json:"degraded"`
Connections int `json:"connections"`
Users int `json:"users"`
Dependencies map[string]*dependencyStatus `json:"dependencies"`
}
@@ -68,28 +68,26 @@ func Healthz(service string) gin.HandlerFunc {
}
}
// Readyzer builds the /readyz handler. It probes Redis (PING) and Kafka
// (reader lag / connectivity) and reports degraded state per ARB-015 §17.4.
// Readyzer builds the /readyz handler. It probes Redis (PING) and reports
// degraded state per ARB-015 §17.4.
//
// The Hub is used to report local connection counts and to detect the
// shutting-down state (which triggers a hard 503). The redisClient may be nil
// in DevMode; the kafkaConsumer may be nil when KAFKA_BROKERS is unset. Both
// nil cases are reported as degraded rather than failing the probe.
// in DevMode; the nil case is reported as degraded rather than failing the
// probe.
type Readyzer struct {
hub *hub.Hub
redis *redisclient.Client
kafka *kafkaconsumer.Consumer
instance string
service string
}
// NewReadyzer constructs a Readyzer. redis and kafka may be nil; the resulting
// probe will mark the missing dependency as degraded.
func NewReadyzer(h *hub.Hub, r *redisclient.Client, k *kafkaconsumer.Consumer, service, instanceID string) *Readyzer {
// NewReadyzer constructs a Readyzer. redis may be nil; the resulting probe
// will mark the missing dependency as degraded.
func NewReadyzer(h *hub.Hub, r *redisclient.Client, service, instanceID string) *Readyzer {
return &Readyzer{
hub: h,
redis: r,
kafka: k,
instance: instanceID,
service: service,
}
@@ -98,8 +96,6 @@ func NewReadyzer(h *hub.Hub, r *redisclient.Client, k *kafkaconsumer.Consumer, s
// Handler is the gin.HandlerFunc for GET /readyz.
func (rz *Readyzer) Handler(c *gin.Context) {
// Hard failure: Hub is shutting down — reject new traffic.
// (Hub.CloseAll sets closing=true; we treat this as 503 so the load
// balancer stops sending WebSocket upgrades during drain.)
if rz.hub.IsClosing() {
c.JSON(http.StatusServiceUnavailable, readyzResponse{
Status: "shutting_down",
@@ -115,7 +111,7 @@ func (rz *Readyzer) Handler(c *gin.Context) {
return
}
deps := make(map[string]*dependencyStatus, 2)
deps := make(map[string]*dependencyStatus, 1)
degraded := false
// Redis probe (soft failure).
@@ -129,17 +125,6 @@ func (rz *Readyzer) Handler(c *gin.Context) {
}
}
// Kafka probe (soft failure).
if rz.kafka == nil {
deps["kafka"] = &dependencyStatus{Ok: false, Error: "not configured"}
degraded = true
} else {
deps["kafka"] = probeKafka(rz.kafka)
if !deps["kafka"].Ok {
degraded = true
}
}
// Always 200 (unless shutting down) per ISSUE-058/006 soft-failure rule.
c.JSON(http.StatusOK, readyzResponse{
Status: statusText(degraded),
@@ -171,16 +156,3 @@ func probeRedis(r *redisclient.Client) *dependencyStatus {
}
return &dependencyStatus{Ok: true, Latency: time.Since(start).Milliseconds()}
}
// probeKafka checks that the consumer reader is still reachable. We use a
// lightweight Lag() call (segmentio/kafka-go Client API); on failure the
// consumer is marked degraded. Note: a degraded Kafka does NOT block WebSocket
// traffic — it only pauses notification consumption until recovery.
func probeKafka(k *kafkaconsumer.Consumer) *dependencyStatus {
ctx, cancel := context.WithTimeout(context.Background(), probeTimeout)
defer cancel()
if err := k.HealthCheck(ctx); err != nil {
return &dependencyStatus{Ok: false, Error: err.Error()}
}
return &dependencyStatus{Ok: true}
}