此外，这一段 scala code 可能有点难理解的地方就是在一个function里面定义了很多其他的function，比如“ sendTxnMarkersCallback”，但是这段callback并不是立刻被执行的，而只是在此定义而已。它真正被执行的地方是在 txnManager.appendTransactionToLog 之后，也就是当这个record被成功append并且备份之后作为callback才会被执行。其他还有好几个地方都有类似的design pattern，可能刚开始看的时候会误解。
国璋 On Tue, Aug 9, 2022 at 10:28 AM Guozhang Wang <wangg...@gmail.com> wrote: > commit / abort 请求收到后，coordinator除了检查状态后还是要append 一个 prepare-commit/abort > record到transaction log，以及要等到这个append被完全备份后才会回复的。然后再广播markers。具体的code在这里： > > > https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/coordinator/transaction/TransactionCoordinator.scala#L595 > > 您指出的code只是事先准备好这个append成功后的callback，并不是直接广播markers。 > > 也就是说，commit/abort的synchronization > barrier是这个prepare-commit/abort被成功append并且备份的时候，而不是广播markers的时候，因为只要append这个record成功，即使之后coordinator宕机了，恢复之后看到这个record还是会继续之前没有完成的广播markers。 > > 希望能够对您理解这段code有帮助。 > > > 国璋 > > On Tue, Aug 9, 2022 at 10:19 AM Guozhang Wang <wangg...@gmail.com> wrote: > >> Hello Zheyan, >> >> Sorry for the late reply.. the email was buried in my inbox for a while. >> >> On Thu, Jul 21, 2022 at 6:08 AM 城哲闫 <18846837...@163.com> wrote: >> >>> Hello, I am looking at the transaction implementation of kafka recently. >>> I have some questions about the transaction implementation and would like >>> to consult with you: >>> >>> class : TransactionCoordinator >>> >>> >>> >>> When the client sends a commit request, the server will directly reply >>> to the client that the commit is successful after checking the status. >>> Next, the server will broadcast Markers, and then write the log to the >>> transaction topic. If the server is down after replying to the commit If >>> the machine is up, is there a state inconsistency problem? I read the code >>> comment that the log was spliced successfully before replying to the >>> commit, but I didn't find the relevant logic, just saw a debug log printed. >>> Is there a mistake in my understanding? ?Will there be a situation where >>> the client thinks it has been submitted, but the server does not submit it? >>> (中文 : 在客户端发送 commit 请求时候 , 服务端在检验状态没问题之后会直接回复客户端 commit 成功 , 接下来服务端再去广播 >>> Markers ,之后去向事务 topic 写入日志, 假如说在回复 commit >>> 之后服务端宕机了,那么是不是会有状态不一致问题,我看代码注释说在回复 commit 之前日志拼接成功,但是我并没有找到相关逻辑,只是看到打印了一个 >>> debug 日志.请问是我的理解有错误么?是否会出现客户端认为已提交,但是服务端没提交的情况那. ) >>> >>> >>> >>> >>> >>> >> >> -- >> -- Guozhang >> > > > -- > -- Guozhang > -- -- Guozhang
