Files
Edu/services/push-gateway/internal/kafkaconsumer/consumer.go

254 lines
8.2 KiB
Go

// Package kafkaconsumer subscribes to the edu.notify.notification.sent topic
// (ARB-013 canonical topic name) on behalf of push-gateway. Each consumed
// NotificationRequested event is dispatched to the Hub for delivery to online
// clients. At-least-once delivery is enforced via manual commit; idempotency
// is provided by Redis SETNX on event_id (TTL 24h).
//
// Topic naming follows ARB-013 (edu.<domain>.<aggregate>.<action>); the
// previous name edu.notification.requested is deprecated (v2 alignment).
package kafkaconsumer
import (
"context"
"encoding/json"
"errors"
"fmt"
"io"
"time"
"github.com/edu-cloud/push-gateway/internal/hub"
"github.com/edu-cloud/push-gateway/internal/observability"
"github.com/edu-cloud/push-gateway/internal/redisclient"
"github.com/segmentio/kafka-go"
)
// MaxRetries is the per-message retry count before sending to the dead-letter
// topic. After MaxRetries failures the message is committed (skipped) and
// logged; a future P6 enhancement can republish to a DLQ topic.
const MaxRetries = 3
// NotificationRequested is the message shape produced by msg (ai10). Field
// names mirror the JSON wire format produced by msg's Outbox publisher.
type NotificationRequested struct {
EventID string `json:"event_id"`
UserID string `json:"user_id"`
EventType string `json:"event_type"`
Channel string `json:"channel"`
Title string `json:"title"`
Content string `json:"content"`
Data json.RawMessage `json:"data"`
Broadcast bool `json:"broadcast"`
OccurredAt int64 `json:"occurred_at"`
}
// Consumer wraps a kafka.Reader for the notification topic.
type Consumer struct {
reader *kafka.Reader
hub *hub.Hub
redis *redisclient.Client
metrics *observability.Metrics
topic string
groupID string
dlqTopic string
}
// Config configures the Kafka consumer.
type Config struct {
Brokers []string
Topic string
GroupID string
DLQTopic string
}
// New creates a Consumer for the configured topic and group.
func New(cfg Config, h *hub.Hub, r *redisclient.Client, m *observability.Metrics) *Consumer {
if cfg.DLQTopic == "" {
cfg.DLQTopic = cfg.Topic + ".dlq"
}
reader := kafka.NewReader(kafka.ReaderConfig{
Brokers: cfg.Brokers,
Topic: cfg.Topic,
GroupID: cfg.GroupID,
MinBytes: 1,
MaxBytes: 10 * 1024 * 1024,
CommitInterval: 1 * time.Second, // periodic background commit
StartOffset: kafka.LastOffset, // skip historical backlog on first start
})
return &Consumer{
reader: reader,
hub: h,
redis: r,
metrics: m,
topic: cfg.Topic,
groupID: cfg.GroupID,
dlqTopic: cfg.DLQTopic,
}
}
// Run blocks until ctx is canceled, consuming messages and dispatching them.
// Errors are logged but do not stop the consumer unless ctx is canceled.
func (c *Consumer) Run(ctx context.Context) error {
observability.Logger().Info("kafka consumer starting",
"topic", c.topic, "group", c.groupID, "brokers", c.reader.Config().Brokers)
for {
select {
case <-ctx.Done():
return ctx.Err()
default:
}
msg, err := c.reader.ReadMessage(ctx)
if err != nil {
if errors.Is(err, io.EOF) || errors.Is(err, context.Canceled) {
return nil
}
observability.Logger().Warn("kafka read failed", "err", err)
select {
case <-ctx.Done():
return nil
case <-time.After(2 * time.Second):
}
continue
}
c.metrics.IncKafkaConsumed(c.topic, fmt.Sprintf("%d", msg.Partition))
c.processMessage(ctx, msg)
}
}
// Close releases the underlying reader. Should be called after Run returns.
func (c *Consumer) Close() error {
return c.reader.Close()
}
// HealthCheck verifies that the Kafka broker is reachable. Used by /readyz as
// a soft-failure probe: a failing check marks the consumer as degraded but
// does not return 503 (per ARB-015 §17.4 / ISSUE-058).
//
// The check issues a metadata request for the consumer topic; on success the
// broker is considered reachable. This is cheaper than a Lag() call which
// requires partition assignment to have completed.
func (c *Consumer) HealthCheck(ctx context.Context) error {
conn, err := kafka.DialContext(ctx, "tcp", c.reader.Config().Brokers[0])
if err != nil {
return fmt.Errorf("kafka: dial broker: %w", err)
}
defer conn.Close()
partitions, err := conn.ReadPartitions(c.topic)
if err != nil {
return fmt.Errorf("kafka: read partitions: %w", err)
}
if len(partitions) == 0 {
return fmt.Errorf("kafka: topic %q has no partitions", c.topic)
}
return nil
}
// Topic returns the configured topic name. Exposed for diagnostics.
func (c *Consumer) Topic() string {
return c.topic
}
// GroupID returns the configured consumer group name.
func (c *Consumer) GroupID() string {
return c.groupID
}
// processMessage handles a single Kafka message with retry and idempotency.
func (c *Consumer) processMessage(ctx context.Context, msg kafka.Message) {
var event NotificationRequested
if err := json.Unmarshal(msg.Value, &event); err != nil {
observability.Logger().Warn("kafka: invalid message payload",
"topic", c.topic, "partition", msg.Partition, "offset", msg.Offset, "err", err)
c.metrics.IncPushed("notification.invalid", "invalid")
// Commit to skip the poison message.
_ = c.reader.CommitMessages(ctx, msg)
return
}
// eventLabel is the metrics label for IncPushed. Prefer event_type (the
// actual event name like "ExamExtended"); fall back to a generic label.
eventLabel := event.EventType
if eventLabel == "" {
eventLabel = "notification.sent"
}
// Idempotency: skip already-processed events by event_id.
fresh, err := c.redis.DedupEventId(ctx, event.EventID)
if err != nil {
observability.Logger().Warn("kafka: dedup check failed, processing anyway",
"event_id", event.EventID, "err", err)
}
if !fresh {
observability.Logger().Debug("kafka: duplicate event skipped",
"event_id", event.EventID)
c.metrics.IncPushed(eventLabel, "duplicate")
_ = c.reader.CommitMessages(ctx, msg)
return
}
// Retry up to MaxRetries on dispatch failure.
for attempt := 1; attempt <= MaxRetries; attempt++ {
if err := c.dispatch(ctx, event); err == nil {
c.metrics.IncPushed(eventLabel, "delivered")
_ = c.reader.CommitMessages(ctx, msg)
return
} else if attempt < MaxRetries {
observability.Logger().Warn("kafka: dispatch retry",
"event_id", event.EventID, "attempt", attempt, "err", err)
time.Sleep(time.Duration(attempt) * time.Second)
}
}
// MaxRetries exhausted: send to DLQ (best-effort) and commit.
observability.Logger().Error("kafka: dispatch failed after retries, sending to DLQ",
"event_id", event.EventID, "topic", c.topic)
c.sendToDLQ(ctx, msg.Value)
c.metrics.IncPushed(eventLabel, "dlq")
_ = c.reader.CommitMessages(ctx, msg)
}
// dispatch delivers the event to local Hub connections and, when the user is
// not local, asks Redis to fan out via Pub/Sub.
func (c *Consumer) dispatch(ctx context.Context, event NotificationRequested) error {
payload, err := json.Marshal(map[string]any{
"type": "message",
"event": event.EventType,
"data": event.Data,
"timestamp": time.Now().UTC().Format(time.RFC3339),
})
if err != nil {
return fmt.Errorf("marshal payload: %w", err)
}
if event.Broadcast {
// Broadcast: deliver locally + publish to other instances.
c.hub.Broadcast(payload)
return c.redis.PublishBroadcast(ctx, redisclient.CrossInstanceMessage{
Event: event.EventType,
Data: event.Data,
})
}
// Directed: try local first, fall back to Redis Pub/Sub.
if c.hub.HasUser(event.UserID) {
if c.hub.SendToUser(event.UserID, payload) > 0 {
return nil
}
}
return c.redis.PublishUser(ctx, event.UserID, redisclient.CrossInstanceMessage{
Event: event.EventType,
Data: event.Data,
})
}
// sendToDLQ publishes a raw message to the dead-letter topic. Best-effort;
// failure is logged but does not block commit.
func (c *Consumer) sendToDLQ(ctx context.Context, value []byte) {
conn, err := kafka.DialLeader(ctx, "tcp", c.reader.Config().Brokers[0], c.dlqTopic, 0)
if err != nil {
observability.Logger().Error("kafka: DLQ dial failed", "topic", c.dlqTopic, "err", err)
return
}
defer conn.Close()
_ = conn.SetWriteDeadline(time.Now().Add(5 * time.Second))
_, _ = conn.WriteMessages(kafka.Message{Value: value})
}