RocketMQ设计之同步刷盘的示例分析

这篇文章主要介绍RocketMQ设计之同步刷盘的示例分析，文中介绍的非常详细，具有一定的参考价值，感兴趣的小伙伴们一定要看完！

在同步刷盘模式下，当消息写到内存后，会等待数据写到磁盘的CommitLog文件。

CommitLog的handleDiskFlush方法：

public void handleDiskFlush(AppendMessageResult result, PutMessageResult putMessageResult, MessageExt messageExt) {
    // Synchronization flush
    if (FlushDiskType.SYNC_FLUSH == this.defaultMessageStore.getMessageStoreConfig().getFlushDiskType()) {
        final GroupCommitService service = (GroupCommitService) this.flushCommitLogService;
        if (messageExt.isWaitStoreMsgOK()) {
            GroupCommitRequest request = new GroupCommitRequest(result.getWroteOffset() + result.getWroteBytes());
            service.putRequest(request);
            boolean flushOK = request.waitForFlush(this.defaultMessageStore.getMessageStoreConfig().getSyncFlushTimeout());
            if (!flushOK) {
                log.error("do groupcommit, wait for flush failed, topic: " + messageExt.getTopic() + " tags: " + messageExt.getTags()
                    + " client address: " + messageExt.getBornHostString());
                putMessageResult.setPutMessageStatus(PutMessageStatus.FLUSH_DISK_TIMEOUT);
            }
        } else {
            service.wakeup();
        }
    }
    // Asynchronous flush
    else {
        if (!this.defaultMessageStore.getMessageStoreConfig().isTransientStorePoolEnable()) {
            flushCommitLogService.wakeup();
        } else {
            commitLogService.wakeup();
        }
    }
}


class GroupCommitService extends FlushCommitLogService {
        private volatile List<GroupCommitRequest> requestsWrite = new ArrayList<GroupCommitRequest>();
        private volatile List<GroupCommitRequest> requestsRead = new ArrayList<GroupCommitRequest>();

        //提交刷盘任务到任务列表
        public synchronized void putRequest(final GroupCommitRequest request) {
            synchronized (this.requestsWrite) {
                this.requestsWrite.add(request);
            }
            if (hasNotified.compareAndSet(false, true)) {
                waitPoint.countDown(); // notify
            }
        }

        private void swapRequests() {
            List<GroupCommitRequest> tmp = this.requestsWrite;
            this.requestsWrite = this.requestsRead;
            this.requestsRead = tmp;
        }

        private void doCommit() {
            synchronized (this.requestsRead) {
                if (!this.requestsRead.isEmpty()) {
                    for (GroupCommitRequest req : this.requestsRead) {
                        // There may be a message in the next file, so a maximum of
                        // two times the flush
                        boolean flushOK = false;
                        for (int i = 0; i < 2 && !flushOK; i++) {
                            flushOK = CommitLog.this.mappedFileQueue.getFlushedWhere() >= req.getNextOffset();

                            if (!flushOK) {
                                CommitLog.this.mappedFileQueue.flush(0);
                            }
                        }

                        req.wakeupCustomer(flushOK);
                    }

                    long storeTimestamp = CommitLog.this.mappedFileQueue.getStoreTimestamp();
                    if (storeTimestamp > 0) {
                        CommitLog.this.defaultMessageStore.getStoreCheckpoint().setPhysicMsgTimestamp(storeTimestamp);
                    }

                    this.requestsRead.clear();
                } else {
                    // Because of individual messages is set to not sync flush, it
                    // will come to this process
                    CommitLog.this.mappedFileQueue.flush(0);
                }
            }
        }

        public void run() {
            CommitLog.log.info(this.getServiceName() + " service started");

            while (!this.isStopped()) {
                try {
                    this.waitForRunning(10);
                    this.doCommit();
                } catch (Exception e) {
                    CommitLog.log.warn(this.getServiceName() + " service has exception. ", e);
                }
            }

            // Under normal circumstances shutdown, wait for the arrival of the
            // request, and then flush
            try {
                Thread.sleep(10);
            } catch (InterruptedException e) {
                CommitLog.log.warn("GroupCommitService Exception, ", e);
            }

            synchronized (this) {
                this.swapRequests();
            }

            this.doCommit();

            CommitLog.log.info(this.getServiceName() + " service end");
        }

        @Override
        protected void onWaitEnd() {
            this.swapRequests();
        }

        @Override
        public String getServiceName() {
            return GroupCommitService.class.getSimpleName();
        }

        @Override
        public long getJointime() {
            return 1000 * 60 * 5;
        }
    }

GroupCommitRequest是刷盘任务，提交刷盘任务后，会在刷盘队列中等待刷盘，而刷盘线程

GroupCommitService每隔10毫秒写一批数据到磁盘。之所以不直接写是磁盘io压力大，写入性能低，每隔10毫秒写一次可以提升磁盘io效率和写入性能。

• putRequest(request) 提交刷盘任务到任务列表

• request.waitForFlush同步等待GroupCommitService将任务列表中的任务刷盘完成。

两个队列读写分离，requestsWrite是写队列，用户保存添加进来的刷盘任务，requestsRead是读队列，在刷盘之前会把写队列的数据放入读队列。

CommitLog的doCommit方法：

private void doCommit() {
            synchronized (this.requestsRead) {
                if (!this.requestsRead.isEmpty()) {
                    for (GroupCommitRequest req : this.requestsRead) {
                        // There may be a message in the next file, so a maximum of
                        // two times the flush
                        boolean flushOK = false;
                        for (int i = 0; i < 2 && !flushOK; i++) {
                            //根据offset确定是否已经刷盘
                            flushOK = CommitLog.this.mappedFileQueue.getFlushedWhere() >= req.getNextOffset();

                            if (!flushOK) {
                                CommitLog.this.mappedFileQueue.flush(0);
                            }
                        }

                        req.wakeupCustomer(flushOK);
                    }

                    long storeTimestamp = CommitLog.this.mappedFileQueue.getStoreTimestamp();
                    if (storeTimestamp > 0) {
                        CommitLog.this.defaultMessageStore.getStoreCheckpoint().setPhysicMsgTimestamp(storeTimestamp);
                    }
                    //清空已刷盘的列表
                    this.requestsRead.clear();
                } else {
                    // Because of individual messages is set to not sync flush, it
                    // will come to this process
                    CommitLog.this.mappedFileQueue.flush(0);
                }
            }
        }

• 刷盘的时候依次读取requestsRead中的数据写入磁盘，

• 写入完成后清空requestsRead。

读写分离设计的目的是在刷盘时不影响任务提交到列表。

CommitLog.this.mappedFileQueue.flush(0);是刷盘操作：

public boolean flush(final int flushLeastPages) {
    boolean result = true;
    MappedFile mappedFile = this.findMappedFileByOffset(this.flushedWhere, this.flushedWhere == 0);
    if (mappedFile != null) {
        long tmpTimeStamp = mappedFile.getStoreTimestamp();
        int offset = mappedFile.flush(flushLeastPages);
        long where = mappedFile.getFileFromOffset() + offset;
        result = where == this.flushedWhere;
        this.flushedWhere = where;
        if (0 == flushLeastPages) {
            this.storeTimestamp = tmpTimeStamp;
        }
    }

    return result;
}

通过MappedFile映射的CommitLog文件写入磁盘

这就是RocketMQ高可用设计之同步刷盘的基本情况了，大体思路就是一个读写分离的队列来刷盘，同步刷盘任务提交后会在刷盘队列中等待刷盘完成后再返回，而GroupCommitService每隔10毫秒写一批数据到磁盘。

以上是“RocketMQ设计之同步刷盘的示例分析”这篇文章的所有内容，感谢各位的阅读！希望分享的内容对大家有帮助，更多相关知识，欢迎关注亿速云行业资讯频道！