Redis client performance issue

[glassfox]

version 4.1.2 The performance of Redis client in Load is very pure. Lot of errors:

1. java.io.IOException: Broken pipe
2. io.netty.channel.StacklessClosedChannelException: null
3. java.net.SocketException: Connection reset

Following test:

@Test
void test1(Vertx vertx, VertxTestContext context) {
    int iterations = 100;
    int instances = 100;
    AtomicInteger countOfErrors = new AtomicInteger();
    vertx.deployVerticle(() -> new AbstractVerticle() {
        @Override
        public void start() throws Exception {
            var redis = RedisAPI.api(Redis.createClient(this.vertx, new RedisOptions().setConnectionString("redis://127.0.0.1:6379")));
            this.vertx.eventBus().consumer("test.redis.load").handler(m -> {
                redis.set(List.of("foo","bar")).onComplete(res ->{
                    
                    if(res.failed()) {
                        countOfErrors.incrementAndGet();
                    }
                });
            });
        }
    }, new DeploymentOptions().setInstances(instances))
    .onComplete(res -> {
        for (int i = 0; i < iterations; i++) {
            vertx.eventBus().publish("test.redis.load", true);
        }
    });
    
    try {
        context.awaitCompletion(5, TimeUnit.SECONDS);
    } catch (InterruptedException e) {
    }
    
    if(countOfErrors.get() > 0) {
        context.failNow("count of errors:" + countOfErrors.get());
    } else {
        context.completeNow();
    }
}

[vietj]

@pmlopes ?

[pmlopes]

I've been trying to reproduce this issue in master, I've been increasing the values until it crashes, but on my side it seems to happen at REDIS side:

  @Test
//  @Ignore
  public void perfRegression(TestContext should) {
    final Async test = should.async();

    int iterations = 500;
    int instances = 100;
    AtomicInteger count = new AtomicInteger();
    AtomicBoolean done = new AtomicBoolean();
    rule.vertx()
      .deployVerticle(() -> new AbstractVerticle() {
        @Override
        public void start(Promise<Void> onStart) {
          Redis redisClient = Redis.createClient(rule.vertx(), new RedisOptions()
            .setConnectionString("redis://" + redis.getContainerIpAddress() + ":" + redis.getFirstMappedPort() + "/0")
            .setMaxWaitingHandlers(iterations + 10)
            .setMaxPoolSize(100)
            .setMaxPoolWaiting(500));

          rule.vertx().eventBus()
            .consumer("test.redis.load")
            .handler(m -> {
              redisClient
                .send(cmd(SET).arg("foo").arg("bar"))
                .onSuccess(res -> {
                  if (count.incrementAndGet() == iterations * instances) {
                    if (done.compareAndSet(false, true)) {
                      rule.vertx()
                        .setTimer(2000L, v -> test.complete());
                    }
                  }
                })
                .onFailure(err -> {
                  should.fail(err);
                });
            });

          onStart.complete();
        }
      }, new DeploymentOptions().setInstances(instances).setWorker(true))
      .onComplete(res -> {
        for (int i = 0; i < iterations; i++) {
          rule.vertx().eventBus().publish("test.redis.load", i);
        }
      });

  }

The first failure happens with:

Oct 25, 2021 10:18:19 AM io.vertx.redis.client.impl.RedisStandaloneConnection
WARNING: No handler waiting for message: ERR max number of clients reached
CONNECTION_CLOSED

So we've reached the maximum allowed client connections from redis.

If I reduce to 100, 100 like in the OP and avoid thread pools by not making the verticles workers, the test passes on my side.

[glassfox]

The test is failed in my side in versions 4.2.0-SNAPSHOT and 4.1.2 (mac and centos 7 x64) @pmlopes , in what version did you check ?