[Bug] Producer should be retry put message if OS_PAGE_CACHE_BUSY

[biningo]

Before Creating the Bug Report

• I found a bug, not just asking a question, which should be created in GitHub Discussions.

• I have searched the GitHub Issues and GitHub Discussions of this repository and believe that this is not a duplicate.

• I have confirmed that this bug belongs to the current repository, not other repositories of RocketMQ.

Runtime platform environment

Ubuntu 20.04

RocketMQ version

develop

JDK Version

JDK 8

Describe the Bug

If the broker is overloaded or disk IO is busy, it will clean up some expired requests and return with SYSTEM_BUSY, at which point the producer will not retry to put the message into the broker. So I think PutMessageStatus is OS_PAGE_CACHE_BUSY should also return SYSTEM_BUSY.

reference: https://github.com/apache/rocketmq/blob/develop/broker/src/main/java/org/apache/rocketmq/broker/processor/SendMessageProcessor.java#L432

Steps to Reproduce

1. mock PutMessageStatus=OS_PAGE_CACHE_BUSY
2. put the message into broker
3. observe the number of produce retries

What Did You Expect to See?

The producer does not retry putting the message when the PutMessageStatus is OS_PAGE_CACHE_BUSY.

What Did You See Instead?

The producer retries putting the message when the PutMessageStatus is OS_PAGE_CACHE_BUSY.

Additional Context

No response

[frinda]

Why can't you try again?
When an MQ is busy, you can retry to send messages to other MQs.

[biningo]

Why can't you try again? When an MQ is busy, you can retry to send messages to other MQs.

Refer to the discussion below:
#5838 #1196 #2713

SYSTEM_BUSY should not retry.

[biningo]

The community is very divided on whether SYSTEM_BUSY should be retried. But there is no retry in the latest version.

[RongtongJin]

From another perspective, semantically returning busy is indeed more reasonable than error.

[humkum]

Generally speaking, page cache busy will only cause process write operations to be suspended for a few seconds. Retrying here will not increase the pressure on page cache synchronization reclaim, and there is no problem in retrying. But semantically speaking, it should indeed belong to SYSTEM BUSY. But this comes back to the original question. Should SYSTEM BUSY be retried? This change will cause messages that were successfully sent before because the PC was busy to be unsuccessful now.

[yuz10]

Maybe we can retry SYSTEM_BUSY by default. currently need to call producer.addRetryResponseCode(ResponseCode.SYSTEM_BUSY); to add to retry list

[biningo]

In my opinion:

1. If the MQ is a single broker node：retry is unreasonable.
2. If the MQ has multiple broker nodes：retry is reasonable.

And it's not possible to have only one broker in a production environment.

[humkum]

In my opinion:

1. If the MQ is a single broker node：retry is unreasonable.
2. If the MQ has multiple broker nodes：retry is reasonable.

And it's not possible to have only one broker in a production environment.

agree

[biningo]

@RongtongJin So what do you think? I think the producer can retry SYSTEM_BUSY by default. If the community agrees, I can start a new PR to fix this issue.

[cserwen]

In my opinion:

1. If the MQ is a single broker node：retry is unreasonable.
2. If the MQ has multiple broker nodes：retry is reasonable.

And it's not possible to have only one broker in a production environment.

We have discussed this issue countless times (like: #5838 #2726 #1196... ), and I think it is time to come to a definite conclusion. There is no doubt that we should support retries in system_busy.

Maybe we can start an email to vote for it. @RongtongJin @yuz10 @biningo

[wz2cool]

In my opinion:

1. If the MQ is a single broker node：retry is unreasonable.
2. If the MQ has multiple broker nodes：retry is reasonable.

And it's not possible to have only one broker in a production environment.

We have discussed this issue countless times (like: #5838 #2726 #1196... ), and I think it is time to come to a definite conclusion. There is no doubt that we should support retries in system_busy.

Maybe we can start an email to vote for it. @RongtongJin @yuz10 @biningo

agree.

[wz2cool]

Maybe we can retry SYSTEM_BUSY by default. currently need to call producer.addRetryResponseCode(ResponseCode.SYSTEM_BUSY); to add to retry list

@biningo Currently, you can try this to solve your problem.

[biningo]

#5845 already merged. So I changed the issue title.