一场由TiCDC异常引发的GC不干活导致的Tikv硬盘使用问题

（转载公众号：晓磊聊DB）

当大家看到这个标题时，就已经知道了下面几点：

（1）出了撒子问题？ok，集群所有KV节点存储硬盘使用80%以上，凌晨触发频繁报警，搞DB的兄弟们还能不能愉快的睡觉？

（2）谁搞的？ok，GC不干活。

（3）真正导致问题的原因是啥？OK，TiCDC背锅。

完了，看到这里本篇文章就基本介绍完了，会不会被劝退？但是老铁千万别走，且听我将细节娓娓道来，后面有“干货”，我会说明我到底做了啥子操作降低硬盘使用；GC为啥不干活？GC的原理是啥？还有为啥是TiCDC背锅。

凌晨的报警和处理
晚上11点收到多条tikv集群硬盘空间报警，爬起来打开电脑看看到底出了啥问题，一般TiKV使用磁盘空间一般都有下面几个原因：

•本来硬盘空间就比较紧张，是不是有哪个业务又晚上“偷偷”灌数据了？这种情况一般看下grafana的overview监控就能定位。
•tikv日志暴涨，之前遇到过集群升级后tikv日志由于静默reigon开启导致的日志量过大。这个看下tikv日志产生情况，如果是日志过量，直接rm日志即可，另外tikv支持配置日志保留策略来自动删除日志。
•其他情况
本次就是“其他情况”的类型，这个集群本来使用方就多，写入也没有突增啥的，tikv日志也没有啥异常，因为知道是测试集群，所以就是删除了一些2021年的tikv日志先将KV存储降低到80%以内不报警了。

第二天.......

好好收拾下测试TiDB集群，解决Tikv存储问题，靠2种方法：

（1）找资源扩容KV

（2）看看是不是有表过大或者其他异常。

因为是测试集群，不着急直接扩容，万一是有一些大表可以清理下数据啥的，资源就有了。所以我先通过下面的命令来统计库表资源使用情况(不太准，可以参考)：

如果不涉及分区表用下面的方式查看表的使用情况：

如果不涉及分区表用下面的方式查看表的使用情况：
select TABLE_SCHEMA,TABLE_NAME,TABLE_ROWS,(DATA_LENGTH+INDEX_LENGTH)/1024/1024/1024 as table_size from tables order by table_size desc limit 20;
另外partition表提供了分区表和非分区表的资源使用情况(我们使用分区表较多)：
select TABLE_SCHEMA,TABLE_NAME,PARTITION_NAME,TABLE_ROWS,(DATA_LENGTH+INDEX_LENGTH)/1024/1024/1024 as table_size from information_schema.PARTITIONS order by table_size desc limit 20;

 

另外官方也提供了接口来统计table使用情况(比较准确):

tidb-ctl table disk-usage --database(-d) [database name] --table(-t) [table name]

 

先看下SQL的统计结果，我发现有些表都几亿，并且有些表rows为0，竟然也有size，因为我之前经常遇到统计信息的不准导致的tables系统表信息不准的情况，为此，我竟然回收了下集群所有table的统计信息。 

然后做的操作就是跟业务沟通，删除一个没用的分区，或者turncate一些table数据，然后我就等待着资源回收了。但是我都等了半个多小时，在没有多少新写入的情况下，region数量和磁盘空间丝毫没有动静，难道GC出了问题？

OK，那我就先从mysql.tidb表看下GC的情况，不看不知道，一看吓一跳啊，我的gc-save-point竟然还在2021年的4月份，那集群中到底累积了多少MVCC版本？？我的天。 

分析和解决GC的问题
通过上图已经可以发现是GC的问题，我们需要找出GC出了什么问题。一般通过日志和监控。先查看tidb server的日志：

{"level":"INFO","time":"2022/03/16 22:55:06.268 +08:00","caller":"gc_worker.go:230","message":"[gc worker] quit","uuid":"5fdb15468e00003"}
{"level":"INFO","time":"2022/03/16 22:55:22.603 +08:00","caller":"gc_worker.go:202","message":"[gc worker] start","uuid":"5fe4ce7c7200002"}
{"level":"INFO","time":"2022/03/16 22:58:22.629 +08:00","caller":"gc_worker.go:608","message":"[gc worker] there's another service in the cluster requires an earlier safe point. gc will continue with the earlier one","uuid":"5fe4ce7c7200002","ourSafePoint":431867062548692992,"minSafePoint":424280964185194498}
{"level":"INFO","time":"2022/03/16 22:58:22.637 +08:00","caller":"gc_worker.go:329","message":"[gc worker] starts the whole job","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"concurrency":5}
{"level":"INFO","time":"2022/03/16 22:58:22.639 +08:00","caller":"gc_worker.go:1036","message":"[gc worker] start resolve locks","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"concurrency":5}
{"level":"INFO","time":"2022/03/16 22:58:41.268 +08:00","caller":"gc_worker.go:1057","message":"[gc worker] finish resolve locks","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"regions":81662}
{"level":"INFO","time":"2022/03/16 22:59:22.617 +08:00","caller":"gc_worker.go:300","message":"[gc worker] there's already a gc job running, skipped","leaderTick on":"5fe4ce7c7200002"}
{"level":"INFO","time":"2022/03/16 23:00:21.331 +08:00","caller":"gc_worker.go:701","message":"[gc worker] start delete ranges","uuid":"5fe4ce7c7200002","ranges":0}
{"level":"INFO","time":"2022/03/16 23:00:21.331 +08:00","caller":"gc_worker.go:750","message":"[gc worker] finish delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0,"cost time":"677ns"}
{"level":"INFO","time":"2022/03/16 23:00:21.336 +08:00","caller":"gc_worker.go:773","message":"[gc worker] start redo-delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0}
{"level":"INFO","time":"2022/03/16 23:00:21.336 +08:00","caller":"gc_worker.go:802","message":"[gc worker] finish redo-delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0,"cost time":"4.016µs"}
{"level":"INFO","time":"2022/03/16 23:00:21.339 +08:00","caller":"gc_worker.go:1562","message":"[gc worker] sent safe point to PD","uuid":"5fe4ce7c7200002","safe point":424280964185194498}
{"level":"INFO","time":"2022/03/16 23:08:22.638 +08:00","caller":"gc_worker.go:608","message":"[gc worker] there's another service in the cluster requires an earlier safe point. gc will continue with the earlier one","uuid":"5fe4ce7c7200002","ourSafePoint":431867219835092992,"minSafePoint":424280964185194498}
{"level":"INFO","time":"2022/03/16 23:08:22.647 +08:00","caller":"gc_worker.go:329","message":"[gc worker] starts the whole job","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"concurrency":5}
{"level":"INFO","time":"2022/03/16 23:08:22.648 +08:00","caller":"gc_worker.go:1036","message":"[gc worker] start resolve locks","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"concurrency":5}
{"level":"INFO","time":"2022/03/16 23:08:43.659 +08:00","caller":"gc_worker.go:1057","message":"[gc worker] finish resolve locks","uuid":"5fe4ce7c7200002","safePoint":424280964185194498,"regions":81647}
{"level":"INFO","time":"2022/03/16 23:09:22.616 +08:00","caller":"gc_worker.go:300","message":"[gc worker] there's already a gc job running, skipped","leaderTick on":"5fe4ce7c7200002"}
{"level":"INFO","time":"2022/03/16 23:10:22.616 +08:00","caller":"gc_worker.go:300","message":"[gc worker] there's already a gc job running, skipped","leaderTick on":"5fe4ce7c7200002"}
{"level":"INFO","time":"2022/03/16 23:10:23.727 +08:00","caller":"gc_worker.go:701","message":"[gc worker] start delete ranges","uuid":"5fe4ce7c7200002","ranges":0}
{"level":"INFO","time":"2022/03/16 23:10:23.727 +08:00","caller":"gc_worker.go:750","message":"[gc worker] finish delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0,"cost time":"701ns"}
{"level":"INFO","time":"2022/03/16 23:10:23.729 +08:00","caller":"gc_worker.go:773","message":"[gc worker] start redo-delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0}
{"level":"INFO","time":"2022/03/16 23:10:23.730 +08:00","caller":"gc_worker.go:802","message":"[gc worker] finish redo-delete ranges","uuid":"5fe4ce7c7200002","num of ranges":0,"cost time":"710ns"}
{"level":"INFO","time":"2022/03/16 23:10:23.733 +08:00","caller":"gc_worker.go:1562","message":"[gc worker] sent safe point to PD","uuid":"5fe4ce7c7200002","safe point":424280964185194498}
{"level":"INFO","time":"2022/03/16 23:18:22.647 +08:00","caller":"gc_worker.go:608","message":"[gc worker] there's another service in the cluster requires an earlier safe point. gc will continue with the earlier one","uuid":"5fe4ce7c7200002","ourSafePoint":431867377121492992,"minSafePoint":424280964185194498}

 

通过日志发现GC一直等早期的safe point：424280964185194498

there's another service in the cluster requires an earlier safe point. gc will continue with the earlier one","uuid":"5fe4ce7c7200002","ourSafePoint":431870050977185792,"minSafePoint":424280964185194498

 

查看下该TSO对应的系统时间，发现跟上面mysql.tidb截图的tikv-gc-safe-point对应的时间：2021-04-16 00:17:13一致。

$ tiup ctl:v5.4.0 pd -u http://10.203.178.96:2379 tso 424280964185194498
system:  2021-04-16 00:17:13.934 +0800 CST
logic:   2

 

需要执行 pd-ctl service-gc-safepoint --pd 命令查询当前的 GC safepoint 与 service GC safepoint，发现卡住的 safe point 跟 ticdc 的 safe_point: 424280964185194498 一致。

$ tiup ctl:v5.4.0 pd -u http://10.xxxx.96:2379 service-gc-safepoint
{
  "service_gc_safe_points": [
    {
      "service_id": "gc_worker",
      "expired_at": 9223372036854775807,
      "safe_point": 431880793651412992
    },
    {
      "service_id": "ticdc",
      "expired_at": 1618503433,
      "safe_point": 424280964185194498
    }
  ],
  "gc_safe_point": 424280964185194498
}

 

查看该集群的ticdc任务，发现卡住的safe point跟changefeed id为dxl-xxxx-task这个同步任务的TSO一致： 

到此终于找到GC问题的原因。ticdc同步任务没有及时删除导致的。解决方案：删除ticdc中该同步任务

tiup cdc:v5.4.0 cli changefeed remove --pd='http://10.xxxx.96:2379' --changefeed-id=dxl-replication-task

 

等一段时间观察safe-point的变化，发现已经开始GC: 

不过又卡在了另一个同步任务(ticdc-demo-xxxx-test-shbt-core)上，继续remove

tiup cdc:v5.4.0 cli changefeed remove --pd='http://10.xxxx.96:2379' --changefeed-id=ticdc-demo-xxxx-test-shbt-core

 

删除后，等待较长时间(几个小时)，最终GC safe point恢复正常： 

GC回收完毕后单KV存储又之前的800G降低到200G左右，大家也可以从下图中右边的 cf size的监控看出来，write CF从之前的3.2T降低到790G左右，下降的最多，大家知道对于tikv的列簇来说，write cf是存数据的版本信息(MVCC)的，这个的下降也印证了TiCDC导致的MVCC版本过多的问题，现在问题完全解决： 

GC原理

虽然讲明白了ticdc->gc->硬盘报警的问题处理，但是我们还需要对GC的原理进行下简单的说明。

TiDB server架构中的GC模块

GC作为TiDB Server的模块，主要的作用就是清理不再需要的旧数据。

大家都知道Tikv底层使用是基于LSM的数据架构，这种数据存储的方式在做数据 update 时，不会更新原来的数据，而是重新写一份新的数据，并且给每份数据都给定一个事务时间戳，这种机制也是我们经常提到的 MVCC（多版本并发控制）机制，为了避免同一条记录的版本过多，TiDB设定了一个 tikv_gc_life_time 来控制版本的保留时间(默认10min)，并且每经过 tikv_gc_run_interval 这个时间间隔，就会触发GC的操作，清除历史版本的记录，下面详细聊下GC的流程。

GC的流程
GC都是由GC Leader触发的，一个tidb server的cluster会选举出一个GC leader，具体是哪个节点可以查看mysql.tidb表，gc leader只是一个身份状态，真正run RC的是该Leader节点的gc worker，另外GC的流程其实从文章上面的日志可以看出来。

•（1）Resolve Locks 阶段
从英文翻译过来很直白：该阶段主要是清理锁，通过GC leader日志可以看出，有 start resolve locks；finish resolve locks的操作，并且都跟着具体的safepoint，其实这个safepoint就是一个时间戳，该阶段会扫描所有的region，只要在这个时间戳之前由2PC产生的锁，该提交的提交，该回滚的回滚。

•（2）Delete Ranges阶段
该阶段主要处理由drop/truncate等操作产生的大量连续数据的删除。tidb会将要删除的range首先记录到mysql.gc_delete_range，然后拿这些已经记录的range找对应的sst文件，调用RocksDB的UnsafeDestroyRange接口进行物理删除，删除后的range会记录到mysql.gc_delete_range_done表。从日志中也可以看到start delete ranges；finish delete ranges的操作。

•（3）Do GC阶段
这个阶段主要是清理由DML操作产生的大量历史数据版本，因为默认的保留时间为10min，那么我一张元数据表一年都没有更新过，难道也会被清理？不要紧张，GC会对每个key都保留safe point 前的最后一次写入时的数据(也就是说，就算表数据不变化，GC也不会“误清理”数据的)。另外这部分由DML产生的GC，不会马上释放空间，需要RocksDB的compaction来真正的回收空间，对了还可能涉及“大量空region的merge”(如何回收空region的解决方案)。这个阶段的日志有：start redo-delete ranges；finish delete ranges

•（4）发送safe point给PD，结束本轮的GC。日志为：sent safe point to PD
GC的监控查看
涉及到GC的主要看下面的几个监控即可：

tikv-detail中的GC模块->GC task：跟上文所描述，如果执行了drop/truncate操作就看看下面监控中total-unafe-destroy-range的metric是否有值，另外日常的DML的GC，可以看看total-gc-keys的task情况。 

tikv-detail中的GC模块->GC task duration：

不同gc task的删除耗时。

tikv-detail中的GC模块->GC speed：

按照keys/s的旧MVCC版本的删除速度。