Add support for disconnect on timeout to recover early from no `RST` packet failures

[yangbodong22011]

Bug Report

I'm one of the Jedis Reviewers and our customers are experiencing unrecoverable issues with Lettuce in production.

Lettuce connects to a Redis host and reads and writes normally. However, if the host fails (the hardware problem directly causes the shutdown, and there is no RST reply to the client at this time), the client will continue to time out until the tcp retransmission ends, and it can be recovered. At this time, it takes about 925.6 s in Linux ( Refer to tcp_retries2 ).

          set k v
client ------------------> redis

          redis server down, no rst
          
          set k v (retran)  1
tcp ------------------> redis (no reply)

      	  set k v (retran)  2
tcp ------------------> redis (no reply)     

    ... after 925.6s

           RST 
tcp ------------------> redis 

      reconnect

Why KeepAlive doesn't fix this

https://github.com/lettuce-io/lettuce-core/issues/1437 (Lettuce supports the option to set KEEPALIVE since version 6.1.0 )

Because the priority of the retransmission packet is higher than that of keepalive, before reaching the keepalive stage, it will continue to retransmit until it is reconnected.

In what scenario is this question sent?

• In most cases, when the operating system is shut down and the process exits, RST can be returned to the client, but RST will not be returned when power is cut off or some machine hardware fails.
• In cloud environments, SLB is usually used. When the backend host fails, if the SLB does not support connection draining, there will be problems.

How to reproduce this issue

1. Start a Redis on a certain port, let's say 6379, and use the following code to connect to Redis.

        RedisClient client = RedisClient.create(RedisURI.Builder.redis(host, port).withPassword(args[1])
            .withTimeout(Duration.ofSeconds(timeout)).build());

        client.setOptions(ClientOptions.builder()
            .socketOptions(socketOptions)
            .autoReconnect(autoReconnect)
            .disconnectedBehavior(disconnectedBehavior)
            .build());

        RedisCommands<String, String> sync = client.connect().sync();

        for (int i = 0; i < times; i++) {
            Thread.sleep(1000);

            try {
                LOGGER.info("{}:{}", i, sync.set("" + i, "" + i));
            } catch (Exception e) {
                LOGGER.error("Set Exception: {}", e.getMessage());
            }
        }

2. Use iptables to disable port 6379 packets on the Redis machine.

iptables -A INPUT -p tcp --dport 6379 -j DROP
iptables -A OUTPUT -p tcp --sport 6379 -j DROP

3. Observe that the client starts timing out and cannot recover until after 925.6 s (related to tcp_retries2)

4. After the test, clear the iptables rules

iptables -F INPUT
iptables -F OUTPUT

How to fix this

We should provide the activation mechanism of the application layer, that is, on the underlying Netty link, periodically insert the activation data packet, if the activation data packet times out, the client will initiate a reconnection to recover quickly.

How Jedis avoids this problem

Jedis is a connection pool mode. When an API times out, Jedis will destroy the link and obtain it again from the connection pool, which can avoid the above problems.

Environment

• Lettuce version(s): main branch
• Redis version: unstable branch

[wpstan]

I have the same problem.
I thought this way may solve it: Schedule Task ping redis if error occur, then call: LettuceConnectionFactory.validateConnection

[yangbodong22011]

@wpstan You can't Schedule Task ping redis in another thread or a new connection, you can only choose an existing connection in netty, otherwise you won't be able to recover from connection draining scenarios.

[wpstan]

@yangbodong22011 That's what I mean. Must use the same Connection. Also ,I found the similar issue #1428 . It recommends setting keepalive parameters for the underlying TCP connection. If the JDK version used is less than 11，Netty Epoll dependencies need to import. I thought call LettuceConnectionFactory.validateConnection() better.

[yangbodong22011]

It recommends setting keepalive parameters for the underlying TCP connection. If the JDK version used is less than 11，Netty Epoll dependencies need to import.

@wpstan keepalive is only valid for sub links in pubsub (without actively sending data, just receiving data from server). see Why KeepAlive doesn't fix this section in my top comments.

[mp911de]

Thanks a lot for the detailed analysis and write up. I wasn't aware of the retransmit vs keepalive priority relationship.

With a blocking I/O client like Jedis is makes a lot of sense to discard connections that have timed out as the structural read is initiated by the method that is being called. Otherwise, a late data reception (caused by a slow server response) leads to protocol desynchronization.

Lettuce has a command timeout mechanism, too. However, the internal architecture allows us to continue operations even in the case a command is timing out (mostly to protected the caller) as the command response parsing is event-driven.

Disconnecting the connection upon command timeout is a pretty drastic approach that can work in certain scenarios in which you cannot customize retransmit parameters. With the current configuration means (new connection events, customization of the Netty bootstrap, command listeners) you have all the necessary bits to implement such a behavior within your application.

If we learn that a larger part of our community is interested in such a feature out of the box, then we will consider such an enhancement.

[yangbodong22011]

@mp911de

Disconnecting the connection upon command timeout is a pretty drastic approach that can work in certain scenarios in which you cannot customize retransmit parameters.

yes, adjusting parameters such as tcp_retries2 is unreachable for many services in docker and k8s.

If we learn that a larger part of our community is interested in such a feature out of the box, then we will consider such an enhancement.

In many of our customers, because of this issue, we need documentation explaining why and do not recommend Lettuce unless the user can accept a long recovery time.

spring-data-redis uses the Lettuce driver by default. After this problem occurs, in order to avoid the impact, users will also try to switch the driver to Jedis, which brings extra burden to them.

I personally also really like the advanced Lettuce client, and I think the community needs to fix this, just like the keepAlive issue.

Can you guide me to enhance this if you don't have time to spare, thanks.

[mp911de]

Indeed, tcp_retries2 can be impossible to customize in a container or serverless environment. I renamed this ticket to reflect its intent. I think it makes sense to host such a feature (disconnect on timeout) within TimeoutOptions. We have CommandExpiryWriter that expires commands and there we could trigger a netty channel disconnect. I'm not fully sure about the design, we currently only have access to a RedisChannelWriter but handing out a netty Channel doesn't make sense. Maybe we could introduce a disconnect() method or so.

[yangbodong22011]

I think it makes sense to host such a feature (disconnect on timeout) within TimeoutOptions.

Agreed, but I think a better strategy is to reconnect after X (1 by default) consecutive timeouts. The reasons are as follows:

1. Some users configure the timout to be very small, and the timeout is frequent for them, but the continuous timeout of X times may be an abnormal situation.
2. Lettuce is a non-connection pool mode, and there is an overhead for new connections, which may not be acceptable for users in point 1.

[mp911de]

I agree that it might not make sense to disconnect after the first timeout. The tricky part is to detect the right state. Let's assume a disconnect after 2 timeouts:

One might have a command sequence of TIMEOUT - SUCCESS - TIMEOUT and without further knowledge, we might disconnect on the second timeout. Or a case where there is no activity between two commands: TIMEOUT (because of RST) - long time without activity, where the remote host recovers - TIMEOUT (because of something else).

We have in other places a delay where we debounce events and delay activity, such as adaptive topology updates in Redis cluster where not every MOVED redirect causes an immediate topology refresh.

In any case, such a feature requires a bit more thought.

[yangbodong22011]

I think a better strategy is to reconnect after X (1 by default) consecutive timeouts

I mean from X consecutive timeouts, if x = 2, the sequence should be: TIMEOUT - TIMEOUT, then, we can reconnect. for TIMEOUT - SUCCESS - TIMEOUT will set X = 0 again after SUCCESS and will not reconnect.

[yangbodong22011]

I think a better strategy is to reconnect after X (1 by default) consecutive timeouts

I mean from X consecutive timeouts, if x = 2, the sequence should be: TIMEOUT - TIMEOUT, then, we can reconnect. for TIMEOUT - SUCCESS - TIMEOUT will set X = 0 again after SUCCESS and will not reconnect.

@mp911de What do you think of this strategy? If agreed, I will prepare a PR.

[chuckz1321]

@yangbodong22011 How's the PR going :) We need this mechanism badly.

[yangbodong22011]

@yangbodong22011 How's the PR going :) We need this mechanism badly.

Waiting for @mp911de to have time to process it, we don't have a firm strategy yet.

[shelltea]

Any update?

[shelltea]

@yangbodong22011 Hi, The PR #2253 is closed. So can you provide a custom Spring Boot starter if you have time. We really need this connection validation mechanism. Compare to switch driver to Jedis, a Spring Boot starter is more easy to use for the users of Lettuce.

[yunbozhang-msft]

Hi, now it seems that keep-alive is a better measure for Lettuce. Can refer to this comments: https://github.com/lettuce-io/lettuce-core/pull/2253#issuecomment-1318351065

[yangbodong22011]

Hi, now it seems that keep-alive is a better measure for Lettuce. Can refer to this comments: #2253 (comment)

But keep-alive does not solve the problem raised by this issue, which is why we have kept it open. If https://github.com/lettuce-io/lettuce-core/pull/2253 is not what we want, I think we need to continue to communicate and discuss to completely solve this problem.

[vongosling]

@mp911de I think you should consider the resolution seriously. many people are deeply troubled by this problem. I also found we had many discussions about this problem, but no further solution was packed by our lettuce team. We really have no choice but to let our customers to chose Jedis for a better network shrink problem. @yangbodong22011

[richieyan]

Refer #1428 , we resolve this problem by add keep-alive and tcp_user_timeout options in our client configuration.

Keep-alive wouldn't work in a situation that client send request to server continuously，but server never ack. So we need tcp_user_timeout to check whether the connection is active or not. You can refer https://blog.cloudflare.com/when-tcp-sockets-refuse-to-die/ to know how tcp_user_timeout work.

You need add netty-transport-native-epoll to your dependencies

Gradle deps: check your lettuce version and use proper netty version

api 'io.netty:netty-transport-native-epoll:4.1.65.Final:linux-x86_64'
api 'io.netty:netty-transport-native-epoll:4.1.65.Final:linux-aarch_64'

Java code example:

       // customize your netty
        ClientResources clientResources = ClientResources.builder()
                .nettyCustomizer(new NettyCustomizer() {
                    @Override
                    public void afterBootstrapInitialized(Bootstrap bootstrap) {
                        if (EpollProvider.isAvailable()) {
                            // TCP_USER_TIMEOUT >= TCP_KEEPIDLE + TCP_KEEPINTVL * TCP_KEEPCNT
                            // https://blog.cloudflare.com/when-tcp-sockets-refuse-to-die/
                            bootstrap.option(EpollChannelOption.TCP_USER_TIMEOUT, tcpUserTimeout);
                        }
                    }
                })
                .build();


       // create your socket options
       SocketOptions socketOptions = SocketOptions.builder()
                .connectTimeout(connectTimeout)
                .keepAlive(SocketOptions.KeepAliveOptions.builder()
                        .enable()
                        .idle(Duration.ofSeconds(5))
                        .interval(Duration.ofSeconds(1))
                        .count(3)
                        .build())
                .build();
         

[yangbodong22011]

@richieyan Thanks for your comments and code, I did some tests and here are the results:

The following tests have the following prerequisites:

• Configure TCP_USER_TIMEOUT
• Limit Redis networking using iptables.

	KeepAlive(on)	KeepAlive(off)
TCP Retran(Yes)	Tcp Retran has higher priority than KeepAlive, so KeepAlive will not start, but after TCP_USER_TIMEOUT arrives, the connection will be closed.	same as left
TCP Retran(No)	KEEPALIVE_TIME = TCP_KEEPIDLE + TCP_KEEPINTVL * TCP_KEEPCNT, If TCP_USER_TIMEOUT is less than KEEPALIVE_TIME, the KeepAlive process will be interrupted and the connection will be closed; if TCP_USER_TIMEOUT is greater than or equal to KEEPALIVE_TIME, KeepAlive will reconnect first.	has no effect (because TCP_USER_TIMEOUT needs unacknowledged packets to trigger)

To reproduce this test, need to pay attention to:

1. The maven configuration of netty-transport-native-epoll needs to add classifier

<dependency>
    <groupId>io.netty</groupId>
    <artifactId>netty-transport-native-epoll</artifactId>
    <version>4.1.65.Final</version>
    <classifier>linux-x86_64</classifier>
</dependency>

2. Open the Debug log of Lettuce and ensure that the following logs appear to ensure that Epoll is used normally:

2023-01-31 10:24:30 [main] DEBUG i.n.u.internal.NativeLibraryLoader - Successfully loaded the library /tmp/libnetty_transport_native_epoll_x86_642162049332005825051.so
2023-01-31 10:24:30 [main] DEBUG i.l.core.resource.EpollProvider - Starting with epoll library

summary

1, TCP_USER_TIMEOUT can indeed solve the problem of this issue on the Linux platform.

2，Not yet verified on MacOS and Windows.

[chuckz1321]

Hi @yangbodong22011 @richieyan Hope you are doing all well! Have you changed the TCP_USER_TIMEOUT? The reconnect happens after 10s even I set the value to 15 or 5. Is there any other parameters controlling the reconnection machenism?

My test step:

1. Connect to Redis (one primary and one replica)
2. Send 'GET' continously
3. Restart primary (NO 'RST' sent from primary)

After primary restarts, Lettuce can reconnect to replica. But the reconnect time, I think it should be controlled by TCP_USER_TIMEOUT. I don't know why the reconnection time is 10s constantly.

[yangbodong22011]

@chuckz1321 Your reconnection is not due to no RST return. Obviously, your Server sent a FIN to the Client, then client begin to reconnect.

If you want to simulate this problem, you need to use the iptables command mentioned in the top comment.

[chuckz1321]

@yangbodong22011

Thanks for your feedback. For this one, the 'GET' did retrans for two times. But the WatchDog start the reconnection immediately. This time the server didn't send RST or FIN. The process seems to be unexpected...

Once the connection is broken, it seems netty awares of this firstly. And following Redis command will request another connection. ` 2023-04-12 08:40:05.467 INFO 44928 --- [nio-9888-exec-8] com.azure.redis4jedis.TestController : start to query from redis1681288805467 2023-04-12 08:40:05.917 ERROR 44928 --- [nio-9888-exec-8] o.a.c.c.C.[.[.[/].[dispatcherServlet] : Servlet.service() for servlet [dispatcherServlet] in context with path [] threw exception [Request processing failed; nested exception is org.springframework.data.redis.RedisSystemException: Redis exception; nested exception is io.lettuce.core.RedisException: io.netty.channel.unix.Errors$NativeIoException: readAddress(..) failed: Connection timed out] with root cause

io.netty.channel.unix.Errors$NativeIoException: readAddress(..) failed: Connection timed out

2023-04-12 08:40:05.927 INFO 44928 --- [nio-9888-exec-9] com.azure.redis4jedis.TestController : start to query from redis1681288805927 2023-04-12 08:40:05.980 INFO 44928 --- [xecutorLoop-1-1] i.l.core.protocol.ConnectionWatchdog : Reconnecting, last destination was cn3redis.redis.cache.chinacloudapi.cn/163.228.235.133:6379 2023-04-12 08:40:08.931 ERROR 44928 --- [nio-9888-exec-9] o.a.c.c.C.[.[.[/].[dispatcherServlet] : Servlet.service() for servlet [dispatcherServlet] in context with path [] threw exception [Request processing failed; nested exception is org.springframework.dao.QueryTimeoutException: Redis command timed out; nested exception is io.lettuce.core.RedisCommandTimeoutException: Command timed out after 3 second(s)] with root cause `

Any insights?

[yangbodong22011]

@chuckz1321 This issue does not apply if the client machine is on macOS. Because the parameter like tcp_retries2 on macOS is very short (I haven't found documentation about this parameter, maybe here, but I'm not sure)

If the client machine is already on Linux, please confirm the size of the tcp_retries2 parameter.

#sysctl -a | grep net.ipv4.tcp_retries2
net.ipv4.tcp_retries2 = 15

[chuckz1321]

@yangbodong22011 Checked before. I am running the application on CentOS. tcp_retries2 is 15.

If I leave all the configuration default, I can reproduce 15 times retrans

[yangbodong22011]

@chuckz1321 hey, here, It's millseconds not seconds.

[chuckz1321]

@yangbodong22011 Sorry, that's a dumb question... Great appreciate for your time and patience.

[huaxne]

@yangbodong22011 Brother, any update with your solution?

[yangbodong22011]

@yangbodong22011 Brother, any update with your solution?

@huaxne Set TCP_USER_TIMEOUT，see https://github.com/lettuce-io/lettuce-core/issues/2082#issuecomment-1407609439

@mp911de Would you consider adding a TCP_USER_TIMEOUT config to Lettuce to fix this, I can contribute a PR.

[mp911de]

@yangbodong22011 Adding TCP_USER_TIMEOUT to SocketOptions (tcpUserTimeout) would be a great addition, happy to include that one.

[yangbodong22011]

@yangbodong22011 Adding TCP_USER_TIMEOUT to SocketOptions (tcpUserTimeout) would be a great addition, happy to include that one.

okay, I will prepare a PR.

[mp911de]

Snapshots are deployed at https://oss.sonatype.org/content/repositories/snapshots/io/lettuce/lettuce-core/6.3.0.BUILD-SNAPSHOT/lettuce-core-6.3.0.BUILD-20230901.094627-49.jar, also available for 6.2.7-BUILD-SNAPSHOT