fix(kafka): preserve partially fetched message batches

server/kafka_runtime.go

@@ -301,6 +301,7 @@ type kafkaMessageCommitter func(context.Context, kafka.Message) error
 type kafkaRetryWaiter func(context.Context, time.Duration) error
 type kafkaBatchHandler func(context.Context, []kafka.Message) error
 type kafkaBatchCommitter func(context.Context, ...kafka.Message) error
+type kafkaMessageFetcher func(context.Context) (kafka.Message, error)
 
 func startKafkaConsumer(topicName string, handler kafkaMessageHandler) {
 	startKafkaConsumerWithGroup(topicName, "", handler)
@@ -389,19 +390,23 @@ func startKafkaBatchConsumer(topicName string, handler kafkaBatchHandler) {
 		defer kafkaBus.wg.Done()
 		defer reader.Close()
 		for {
-			messages, err := fetchKafkaMessageBatch(kafkaBus.ctx, reader, kafkaBus.config.MessageBatchSize,
+			messages, fetchErr := fetchKafkaMessageBatch(kafkaBus.ctx, reader, kafkaBus.config.MessageBatchSize,
 				time.Duration(kafkaBus.config.MessageBatchWait)*time.Millisecond)
-			if err != nil {
-				if errors.Is(err, context.Canceled) || errors.Is(kafkaBus.ctx.Err(), context.Canceled) {
+			if fetchErr != nil && len(messages) == 0 {
+				if errors.Is(fetchErr, context.Canceled) || errors.Is(kafkaBus.ctx.Err(), context.Canceled) {
 					return
 				}
-				logs.Err.Println("Kafka batch fetch failed:", err)
+				logs.Err.Println("Kafka batch fetch failed:", fetchErr)
 				continue
 			}
+			if fetchErr != nil {
+				logs.Warn.Printf("Kafka batch fetch interrupted; processing retained messages: topic=%s size=%d err=%v",
+					topic, len(messages), fetchErr)
+			}
 			if len(messages) > 0 {
 				readerMetrics.observe(messages[0].Time)
 			}
-			if err = consumeKafkaMessageBatch(kafkaBus.ctx, messages, handler, reader.CommitMessages, waitKafkaRetry); err != nil {
+			if err := consumeKafkaMessageBatch(kafkaBus.ctx, messages, handler, reader.CommitMessages, waitKafkaRetry); err != nil {
 				if errors.Is(err, context.Canceled) || errors.Is(kafkaBus.ctx.Err(), context.Canceled) {
 					return
 				}
@@ -468,7 +473,11 @@ func (m *kafkaReaderMetrics) update() {
 }
 
 func fetchKafkaMessageBatch(ctx context.Context, reader *kafka.Reader, maxSize int, maxWait time.Duration) ([]kafka.Message, error) {
-	first, err := reader.FetchMessage(ctx)
+	return fetchKafkaMessageBatchWith(ctx, maxSize, maxWait, reader.FetchMessage)
+}
+
+func fetchKafkaMessageBatchWith(ctx context.Context, maxSize int, maxWait time.Duration, fetch kafkaMessageFetcher) ([]kafka.Message, error) {
+	first, err := fetch(ctx)
 	if err != nil {
 		return nil, err
 	}
@@ -480,13 +489,16 @@ func fetchKafkaMessageBatch(ctx context.Context, reader *kafka.Reader, maxSize i
 			break
 		}
 		fetchCtx, cancel := context.WithTimeout(ctx, remaining)
-		message, fetchErr := reader.FetchMessage(fetchCtx)
+		message, fetchErr := fetch(fetchCtx)
 		cancel()
 		if errors.Is(fetchErr, context.DeadlineExceeded) && ctx.Err() == nil {
 			break
 		}
 		if fetchErr != nil {
-			return nil, fetchErr
+			// FetchMessage advances the reader's internal position before offsets are
+			// committed. Preserve a partial batch so a later higher-offset commit cannot
+			// skip messages already fetched into this batch.
+			return messages, fetchErr
 		}
 		messages = append(messages, message)
 	}