WebFlux+Lettuce读取Redis性能问题：单线程反序列化与优化

测试WebFlux读取Redis数据时发现，WebFlux会使用Lettuce的单线程（例如lettuce-ioEventLoop-5-1）完成读取及反序列化操作，耗时主要集中在反序列化环节，无法利用多核CPU的优势。根据Lettuce官方文档，可通过配置ClientOptions优化该问题，以下是具体配置过程及性能对比：

原代码实现 #

Controller代码 #

@GetMapping(value = "/{testName}", produces = MediaType.APPLICATION_JSON_VALUE)
public Mono<Test> getTest(@PathVariable String testName) {
    return testService.getCachedTest(testName)
            .log();
}

Redis配置类（原） #

@Configuration
public class RedisConfig {

    @Bean
    public ReactiveRedisOperations<String, Test> redisOperations(ReactiveRedisConnectionFactory reactiveRedisConnectionFactory) {
        Jackson2JsonRedisSerializer<Test> serializer = new Jackson2JsonRedisSerializer<>(Test.class);

        RedisSerializationContext.RedisSerializationContextBuilder<String, Test> builder =
                RedisSerializationContext.newSerializationContext(new Jackson2JsonRedisSerializer<>(Test.class));

        RedisSerializationContext<String, Test> context =
                builder
                        .key(new StringRedisSerializer())
                        .value(serializer)
                        .build();

        return new ReactiveRedisTemplate<>(reactiveRedisConnectionFactory, context);
    }
}

优化后的Redis配置类 #

通过自定义LettuceConnectionFactory并配置ClientOptions，开启publishOnScheduler参数，让反序列化操作脱离Lettuce的IO线程，交由调度器处理以利用多核优势：

@Configuration
public class RedisConfig {

    @Bean
    public ReactiveRedisOperations<String, Pokemon> redisOperations(ReactiveRedisConnectionFactory reactiveRedisConnectionFactory) {
        Jackson2JsonRedisSerializer<Pokemon> serializer = new Jackson2JsonRedisSerializer<>(Pokemon.class);

        RedisSerializationContext.RedisSerializationContextBuilder<String, Pokemon> builder =
                RedisSerializationContext.newSerializationContext(new Jackson2JsonRedisSerializer<>(Pokemon.class));

        RedisSerializationContext<String, Pokemon> context =
                builder
                        .key(new StringRedisSerializer())
                        .value(serializer)
                        .build();

        return new ReactiveRedisTemplate<>(reactiveRedisConnectionFactory, context);
    }

    @Bean
    public LettuceConnectionFactory lettuceConnectionFactory() {
        RedisStandaloneConfiguration redisStandaloneConfiguration = new RedisStandaloneConfiguration("localhost", 6379);
        ClientOptions clientOptions = ClientOptions.builder()
                .publishOnScheduler(true)
                .build();
        LettuceClientConfiguration clientConfiguration = LettuceClientConfiguration.builder()
                .clientOptions(clientOptions)
                .build();
        return new LettuceConnectionFactory(redisStandaloneConfiguration, clientConfiguration);
    }
}

性能对比数据 #

配置前响应时间统计 #

"http.response_time": {
    "min": 2,
    "max": 33,
    "count": 8330,
    "mean": 8.4,
    "p50": 6,
    "median": 6,
    "p75": 12.1,
    "p90": 16.9,
    "p95": 19.1,
    "p99": 22.9,
    "p999": 27.9
  }

配置后响应时间统计 #

"http.response_time": {
    "min": 3,
    "max": 47,
    "count": 8330,
    "mean": 4.7,
    "p50": 4,
    "median": 4,
    "p75": 5,
    "p90": 6,
    "p95": 7,
    "p99": 8.9,
    "p999": 19.9
  }

优化后在多核机器上性能大幅提升，平均响应时间从8.4ms降至4.7ms，各分位响应时间也明显优化。

内容的提问来源于stack exchange，提问作者Jitesh