seldon core 通过 kafka 实现数据回流

一、官方方案

https://docs.seldon.io/projects/seldon-core/en/latest/analytics/logging.html?highlight=kafka
https://docs.seldon.io/projects/seldon-core/en/latest/streaming/kafka.html?highlight=kafka

示例:

apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: cifar10
  namespace: seldon
spec:
  name: resnet32
  predictors:
  - graph:
      implementation: TRITON_SERVER
      logger:
        mode: all
      modelUri: gs://seldon-models/triton/tf_cifar10
      name: cifar10
    name: default
    svcOrchSpec:
      env:
      - name: LOGGER_KAFKA_BROKER
        value: seldon-kafka-plain-0.kafka:9092
      - name: LOGGER_KAFKA_TOPIC
        value: seldon
    replicas: 1
  protocol: v2

The two required environment variables are:

  • LOGGER_KAFKA_BROKER : The Kafka Broker service endpoint.
  • LOGGER_KAFKA_TOPIC : The kafka Topic to log the requests.

实践:
seldon-log-demo.yaml

[root@k8s-master seldon]# cat seldon-log-demo.yaml 
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: iris-modelkafka-log
  namespace: seldon
spec:
  name: iris
  predictors:
  - graph:
      implementation: SKLEARN_SERVER
      logger:
        mode: all
      modelUri: gs://seldon-models/v1.12.0-dev/sklearn/iris
      name: classifier
    svcOrchSpec:
      env:
      - name: LOGGER_KAFKA_BROKER
        value: 192.168.1.191:9092
      - name: LOGGER_KAFKA_TOPIC
        value: test-kafka
    name: default
    replicas: 1

重要的两个参数:

LOGGER_KAFKA_BROKER:192.168.1.191:9092

# topic 需要提前创建
LOGGER_KAFKA_TOPIC: test-kafka   

部署模型:

[root@k8s-master seldon]# kubectl apply -f seldon-log-demo.yaml 

[root@k8s-master seldon]# kubectl get pods -n seldon
NAME                                                        READY   STATUS        RESTARTS   AGE
iris-modelkafka-log-default-0-classifier-7c985485b5-5gs8f   3/3     Running       0          7h37m

查看部署的svc:

[root@k8s-master ~]# kubectl get svc -n seldon
NAME                                     TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)             AGE
iris-modelkafka-log-default              ClusterIP   10.233.50.157   <none>        8000/TCP,5001/TCP   7h29m
iris-modelkafka-log-default-classifier   ClusterIP   10.233.63.196   <none>        9000/TCP,9500/TCP   7h29m

在k8s集群测试服务:

curl -X POST http://10.233.50.157:8000/api/v1.0/predictions \
    -H 'Content-Type: application/json' \
    -d '{ "data": { "ndarray": [[1,2,3,4]] } }'

测试打印:

[root@k8s-master ~]# curl -X POST http://10.233.50.157:8000/api/v1.0/predictions \
>     -H 'Content-Type: application/json' \
>     -d '{ "data": { "ndarray": [[1,2,3,4]] } }'
{"data":{"names":["t:0","t:1","t:2"],"ndarray":[[0.0006985194531162835,0.00366803903943666,0.995633441507447]]},"meta":{"requestPath":{"classifier":"seldonio/sklearnserver:1.14.1"}}}

[root@k8s-master ~]# curl -X POST http://10.233.50.157:8000/api/v1.0/predictions     -H 'Content-Type: application/json'     -d '{ "data": { "ndarray": [[1,2,3,4]] } }'
{"data":{"names":["t:0","t:1","t:2"],"ndarray":[[0.0006985194531162835,0.00366803903943666,0.995633441507447]]},"meta":{"requestPath":{"classifier":"seldonio/sklearnserver:1.14.1"}}}

监听kafka消息,获取回流日志:
file

offset = 7, value = { "data": { "ndarray": [[1,2,3,4]] } }
offset = 8, value = {"data":{"names":["t:0","t:1","t:2"],"ndarray":[[0.0006985194531162835,0.00366803903943666,0.995633441507447]]},"meta":{"requestPath":{"classifier":"seldonio/sklearnserver:1.14.1"}}}

offset = 9, value = { "data": { "ndarray": [[1,2,3,4]] } }
offset = 10, value = {"data":{"names":["t:0","t:1","t:2"],"ndarray":[[0.0006985194531162835,0.00366803903943666,0.995633441507447]]},"meta":{"requestPath":{"classifier":"seldonio/sklearnserver:1.14.1"}}}

二、kafka生产者和消费者

KafkaProducerDemo.java


import org.apache.kafka.clients.producer.KafkaProducer;
import org.apache.kafka.clients.producer.ProducerConfig;
import org.apache.kafka.clients.producer.ProducerRecord;
import org.apache.kafka.common.serialization.StringSerializer;

import java.util.Properties;

/**
 * Kafka生产者Demo
 */
public class KafkaProducerDemo {

    public static void main(String[] args) {
        Properties props = new Properties();

        // 服务器ip:端口号,集群用逗号分隔
        props.put(ProducerConfig.BOOTSTRAP_SERVERS_CONFIG, "192.168.1.191:9092");
        // key序列化指定类
        props.put(ProducerConfig.KEY_SERIALIZER_CLASS_CONFIG, StringSerializer.class.getName());
        // value序列化指定类
        props.put(ProducerConfig.VALUE_SERIALIZER_CLASS_CONFIG, StringSerializer.class.getName());

        // 生产者对象
        KafkaProducer<String, String> producer = new KafkaProducer<>(props);

        // 向test_topic发送hello, kafka
        producer.send(new ProducerRecord<String, String>("test-kafka", "哇哈哈0000哈999"));
        producer.close();
    }

}

消费者:
KafkaConsumerDemo.java

import org.apache.kafka.clients.consumer.ConsumerConfig;
import org.apache.kafka.clients.consumer.ConsumerRecord;
import org.apache.kafka.clients.consumer.ConsumerRecords;
import org.apache.kafka.clients.consumer.KafkaConsumer;
import org.apache.kafka.clients.producer.KafkaProducer;
import org.apache.kafka.clients.producer.ProducerConfig;
import org.apache.kafka.clients.producer.ProducerRecord;
import org.apache.kafka.common.serialization.StringDeserializer;
import org.apache.kafka.common.serialization.StringSerializer;

import java.util.Arrays;
import java.util.Properties;

/**
 * Kafka消费者Demo
 */
public class KafkaConsumerDemo {

    public static void main(String[] args) {
        Properties props = new Properties();

        // 服务器ip:端口号,集群用逗号分隔
        props.put(ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG, "192.168.1.191:9092");
        // 消费者指定组,名称可以随意,注意相同消费组中的消费者只能对同一个分区消费一次
        props.put(ConsumerConfig.GROUP_ID_CONFIG, "test");
        // 是否启用自动提交,默认true
        props.put(ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG, true);
        // 自动提交间隔时间1s
        props.put(ConsumerConfig.AUTO_COMMIT_INTERVAL_MS_CONFIG, 1000);
        // key反序列化指定类
        props.put(ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG, StringDeserializer.class.getName());
        // value反序列化指定类,注意生产者与消费者要保持一致,否则解析出问题
        props.put(ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG, StringDeserializer.class.getName());

        // 消费者对象
        KafkaConsumer<String, String> kafkaConsumer = new KafkaConsumer<>(props);

        kafkaConsumer.subscribe(Arrays.asList("test-kafka"));
        while (true) {
            ConsumerRecords<String, String> records = kafkaConsumer.poll(100);
            for (ConsumerRecord<String, String> record : records) {
                System.out.printf("offset = %d, value = %s", record.offset(), record.value());
                System.out.println();
            }
        }

    }

}

pom.xml引入的依赖:

<?xml version="1.0" encoding="UTF-8"?>
<project xmlns="http://maven.apache.org/POM/4.0.0"
         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
    <modelVersion>4.0.0</modelVersion>

    <groupId>org.example</groupId>
    <artifactId>kafka-test</artifactId>
    <version>1.0-SNAPSHOT</version>

    <properties>
        <maven.compiler.source>8</maven.compiler.source>
        <maven.compiler.target>8</maven.compiler.target>
    </properties>

    <dependencies>
        <dependency>
            <groupId>org.apache.kafka</groupId>
            <artifactId>kafka-clients</artifactId>
            <version>3.2.0</version>
        </dependency>

    </dependencies>

</project>

实践

https://towardsdatascience.com/real-time-stream-processing-for-machine-learning-at-scale-with-spacy-kafka-seldon-core-6360f2fedbe

istio

我们选择在 seldon 命名空间下部署模型,所以这里先给 seldon 打个 Istio 自动注入的标签。这使得在 seldon 命名空间下创建的 k8s object 都会被 Istio 自动注入,纳入到 Istio 的管理。这里是为了使用 Istio 的路由管理功能:

 kubectl label namespace seldon istio-injection=enabled

为者常成,行者常至