我在使用KStreams作业时遇到了问题，该作业消耗了大量的磁盘IO。这是一台最大IOPS为500的Azure VM，正在消耗约500tps（从RHEL中的iostat开始）每秒处理400条记录。作业从Kafka主题（12个分区）中提取记录，解构JSON字符串并进行一些处理（字符串搜索，替换和数据丰富），然后将其传递到Kafka接收器（6个分区）。就是说，据我所知，这项工作应该是无状态的，并且不需要RocksDB或其他状态存储。
我有commit.interval.ms=60000（1分钟），在Properties中设置为StreamsConfig.COMMIT_INTERVAL_MS_CONFIG=60000，并且我尝试了其他设置，但这似乎并没有影响磁盘IO或我们请参阅写入checkpoint.tmp。我尝试使用StreamsConfig.NUM_STREAM_THREADS_CONFIG在1到12个线程中运行。此作业可以处理的每秒IO和记录数直接受到线程数的影响，更多线程意味着更高的IO和吞吐量。我已经关闭了所有日志记录（DEBUG，INFO等），因此那里应该没有IO。据我所知，这项工作一直在致力于checkpoint.tmp，而且还经常达到imjournal.state.tmp。通过使用echo 1 > /proc/sys/vm/block_dump打开日志记录，然后使用dmesg -wH查看正在发生的事情，可以得到此信息。以下是当我查看dmesg时看到的内容的一个小示例。
有人对我可以尝试的设置有提示吗，或者我可以在其他地方找出导致大量IO的原因？
谢谢。
[ +0.001097] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000002] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.005463] java(73280): dirtied inode 17178982 (.checkpoint.tmp) on sda2
[ +0.000003] java(73280): dirtied inode 17178982 (.checkpoint.tmp) on sda2
[ +0.000022] java(73280): WRITE block 16922536 on sda2 (8 sectors)
[ +0.000134] java(73280): WRITE block 33057066 on sda2 (14 sectors)
[ +0.000855] java(73284): dirtied inode 17178985 (.checkpoint.tmp) on sda2
[ +0.000001] java(73281): dirtied inode 51158861 (.checkpoint.tmp) on sda2
[ +0.000002] java(73281): dirtied inode 51158861 (.checkpoint.tmp) on sda2
[ +0.000002] java(73284): dirtied inode 17178985 (.checkpoint.tmp) on sda2
[ +0.000017] java(73284): WRITE block 16922544 on sda2 (8 sectors)
[ +0.000000] java(73281): WRITE block 50389296 on sda2 (8 sectors)
[ +0.000184] java(73282): dirtied inode 892020 (.checkpoint.tmp) on sda2
[ +0.000000] java(73283): dirtied inode 51158863 (.checkpoint.tmp) on sda2
[ +0.000002] java(73283): dirtied inode 51158863 (.checkpoint.tmp) on sda2
[ +0.000002] java(73282): dirtied inode 892020 (.checkpoint.tmp) on sda2
[ +0.000015] java(73282): WRITE block 1140808 on sda2 (8 sectors)
[ +0.000001] java(73283): WRITE block 50389304 on sda2 (8 sectors)
[ +0.000576] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000002] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000671] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000002] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.001975] java(73285): dirtied inode 35443635 (.checkpoint.tmp) on sda2
[ +0.000002] java(73285): dirtied inode 35443635 (.checkpoint.tmp) on sda2
[ +0.000015] java(73285): WRITE block 34930256 on sda2 (8 sectors)
[ +0.001344] java(73282): WRITE block 33057080 on sda2 (23 sectors)
[ +0.006529] java(73280): dirtied inode 17178984 (.checkpoint.tmp) on sda2
[ +0.000003] java(73280): dirtied inode 17178984 (.checkpoint.tmp) on sda2
[ +0.000016] java(73280): WRITE block 16922280 on sda2 (8 sectors)
[ +0.002142] java(73280): WRITE block 33057103 on sda2 (7 sectors)
[ +0.001010] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000003] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.003449] java(73281): dirtied inode 51158862 (.checkpoint.tmp) on sda2
[ +0.000000] java(73284): dirtied inode 17178983 (.checkpoint.tmp) on sda2
[ +0.000002] java(73284): dirtied inode 17178983 (.checkpoint.tmp) on sda2
[ +0.000003] java(73281): dirtied inode 51158862 (.checkpoint.tmp) on sda2
[ +0.000017] java(73281): WRITE block 50389280 on sda2 (8 sectors)
[ +0.000001] java(73284): WRITE block 16924528 on sda2 (8 sectors)
[ +0.000224] java(73283): dirtied inode 51158864 (.checkpoint.tmp) on sda2
[ +0.000003] java(73283): dirtied inode 51158864 (.checkpoint.tmp) on sda2
[ +0.000001] java(73285): dirtied inode 35443622 (.checkpoint.tmp) on sda2
[ +0.000001] java(73285): dirtied inode 35443622 (.checkpoint.tmp) on sda2
[ +0.000016] java(73283): WRITE block 50389376 on sda2 (8 sectors)
[ +0.000001] java(73285): WRITE block 34930264 on sda2 (8 sectors)
[ +0.000171] java(73282): dirtied inode 892021 (.checkpoint.tmp) on sda2
[ +0.000003] java(73282): dirtied inode 892021 (.checkpoint.tmp) on sda2
[ +0.000018] java(73282): WRITE block 1140824 on sda2 (8 sectors)
[ +0.000331] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000003] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000351] java(73284): WRITE block 33057110 on sda2 (25 sectors)
[ +0.000357] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000002] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.004964] java(73280): dirtied inode 17178982 (.checkpoint.tmp) on sda2
[ +0.000002] java(73280): dirtied inode 17178982 (.checkpoint.tmp) on sda2
[ +0.000018] java(73280): WRITE block 16922728 on sda2 (8 sectors)
[ +0.000996] java(73285): WRITE block 33057135 on sda2 (26 sectors)
[ +0.004757] java(73284): dirtied inode 17178985 (.checkpoint.tmp) on sda2
[ +0.000002] java(73284): dirtied inode 17178985 (.checkpoint.tmp) on sda2
[ +0.000016] java(73284): WRITE block 16922720 on sda2 (8 sectors)
[ +0.000310] java(73281): dirtied inode 51158861 (.checkpoint.tmp) on sda2
[ +0.000002] java(73281): dirtied inode 51158861 (.checkpoint.tmp) on sda2
[ +0.000015] java(73281): WRITE block 50389288 on sda2 (8 sectors)
[ +0.001163] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000003] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000082] java(73284): WRITE block 33057161 on sda2 (17 sectors)
[ +0.004419] java(73285): dirtied inode 35443635 (.checkpoint.tmp) on sda2
[ +0.000003] java(73285): dirtied inode 35443635 (.checkpoint.tmp) on sda2
[ +0.000018] java(73285): WRITE block 34930280 on sda2 (8 sectors)
[ +0.000357] java(73282): dirtied inode 892020 (.checkpoint.tmp) on sda2
[ +0.000003] java(73282): dirtied inode 892020 (.checkpoint.tmp) on sda2
[ +0.000017] java(73283): dirtied inode 51158863 (.checkpoint.tmp) on sda2
[ +0.000002] java(73283): dirtied inode 51158863 (.checkpoint.tmp) on sda2
[ +0.000002] java(73282): WRITE block 1140816 on sda2 (8 sectors)
[ +0.000012] java(73283): WRITE block 50389272 on sda2 (8 sectors)
[ +0.000237] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000003] in:imjournal(1190): dirtied inode 50351527 (imjournal.state.tmp) on sda2
[ +0.000530] java(73280): dirtied inode 17178984 (.checkpoint.tmp) on sda2
[ +0.000002] java(73280): dirtied inode 17178984 (.checkpoint.tmp) on sda2
[ +0.000027] java(73280): WRITE block 16922536 on sda2 (8 sectors)
[ +0.000465] java(73285): WRITE block 33057178 on sda2 (38 sectors)
[ +0.000637] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
[ +0.000003] in:imjournal(1190): dirtied inode 50351525 (imjournal.state.tmp) on sda2
答案 0 :(得分：1)
在Slack与Confluent社区交谈之后，我意识到我正在运行org.apache.kafka-> kafka-streams-> 1.0.2。这个框架的最新版本是2.3.0，所以我将我的应用升级到2.2.1，然后重试。该版本的应用程序几乎没有明显的tps或iowait运行，只有1个应用程序，产生了12个线程（线程数与分区数相同），每秒能够处理多达1,800条记录。我将应用程序升级到2.3.0并获得了相同的结果，因此我将继续使用该版本。
我不确定100％地确定为什么修复了该问题，但是我将假定它与检查点对开发团队在较新版本中所做的更改有关。