HBase依靠ZooKeeper来感知集群成员及其存活性。假设一个server暂停了非常长时间,它将无法给ZooKeeper quorum发送心跳信息,其他server会觉得这台server已死亡。这将导致master为其启动恢复进程。当该server脱离停顿时,它会发现它的全部租约都已失效(hbase
client端每次和regionserver交互的时候,都会在服务器端生成一个租约(Lease)。租约的有效期由參数hbase.regionserver.lease.period确定)。然后自杀。

HBase开发团队亲切地称这个场景为Juliet Pause。

在HBase中使用默认GC,能够看到全部线程中的长时间停顿,包含Juliet Pause, 又名"GC of
Death"。能够在JAVA虚拟机中开启GC日志,来帮助调试这样的问题。

hbase-env.sh中,取消下面任一凝视就可以开启GC日志:

# This enables basic gc logging to the .out file.
# export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps" # This enables basic gc logging to its own file.
# export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH>" # This enables basic GC logging to its own file with automatic log rolling. Only applies to jdk 1.6.0_34+ and 1.7.0_2+.
# export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH> -XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles=1 -XX:GCLogFileSize=512M" # If <FILE-PATH> is not replaced, the log file(.gc) would be generated in the HBASE_LOG_DIR.

HBase GC日志-LMLPHP

点击"Local logs",查看gc-xxx.log日志,格式例如以下:

15578637.583: [GC [1 CMS-initial-mark: 12076002K(12582912K)] 18788432K(29360128K), 1.0113310 secs] [Times: user=1.01 sys=0.00, real=1.01 secs]
15578638.595: [CMS-concurrent-mark-start]
15578639.125: [CMS-concurrent-mark: 0.530/0.530 secs] [Times: user=1.61 sys=0.01, real=0.53 secs]
15578639.125: [CMS-concurrent-preclean-start]
15578639.160: [CMS-concurrent-preclean: 0.035/0.035 secs] [Times: user=0.03 sys=0.00, real=0.03 secs]
15578639.160: [CMS-concurrent-abortable-preclean-start]
CMS: abort preclean due to time 15578644.606: [CMS-concurrent-abortable-preclean: 5.446/5.446 secs] [Times: user=5.82 sys=0.07, real=5.45 secs]
15578644.607: [GC[YG occupancy: 6769031 K (16777216 K)]15578644.607: [Rescan (parallel) , 0.9756010 secs]15578645.583: [weak refs processing, 0.0000280 secs]15578645.583: [scrub string table, 0.0006370 secs] [1 CMS-remark: 12076002K(12582912K)] 18845034K(29360128K), 0.9764070 secs] [Times: user=9.52 sys=0.03, real=0.97 secs]
15578645.584: [CMS-concurrent-sweep-start]
15578645.829: [CMS-concurrent-sweep: 0.245/0.245 secs] [Times: user=0.30 sys=0.00, real=0.24 secs]
15578645.829: [CMS-concurrent-reset-start]
15578645.855: [CMS-concurrent-reset: 0.026/0.026 secs] [Times: user=0.05 sys=0.01, real=0.03 secs] 15578647.856: [GC [1 CMS-initial-mark: 12075999K(12582912K)] 20737398K(29360128K), 1.0872730 secs] [Times: user=1.09 sys=0.00, real=1.08 secs]
15578648.943: [CMS-concurrent-mark-start]
15578649.472: [CMS-concurrent-mark: 0.528/0.528 secs] [Times: user=2.06 sys=0.02, real=0.53 secs]
15578649.472: [CMS-concurrent-preclean-start]
15578649.507: [CMS-concurrent-preclean: 0.035/0.035 secs] [Times: user=0.04 sys=0.00, real=0.04 secs]
15578649.507: [CMS-concurrent-abortable-preclean-start]
CMS: abort preclean due to time 15578654.961: [CMS-concurrent-abortable-preclean: 5.454/5.454 secs] [Times: user=8.15 sys=0.17, real=5.45 secs]
15578654.962: [GC[YG occupancy: 11315403 K (16777216 K)]15578654.963: [Rescan (parallel) , 1.1274320 secs]15578656.090: [weak refs processing, 0.0000240 secs]15578656.090: [scrub string table, 0.0005890 secs] [1 CMS-remark: 12075999K(12582912K)] 23391403K(29360128K), 1.1281890 secs] [Times: user=11.00 sys=0.03, real=1.13 secs]
15578656.091: [CMS-concurrent-sweep-start]
15578656.327: [CMS-concurrent-sweep: 0.236/0.236 secs] [Times: user=0.52 sys=0.01, real=0.24 secs]
15578656.327: [CMS-concurrent-reset-start]
15578656.353: [CMS-concurrent-reset: 0.026/0.026 secs] [Times: user=0.05 sys=0.00, real=0.02 secs] </span></span>
CMS垃圾回收的运行过程:

1、CMS-initial-mark :初始标记,从root对象開始标记存活的对象。它花费的时间是real=1.01 secs,因为这一步骤是STW的,所以整个application被暂停了1.01秒。而且,user time和real time相差不大,所以确实是仅仅有一个thread运行这一步;

2、CMS-concurrent-mark-start:并发标记。它运行时间是real=0.53 secs,但由于这一步是能够并发运行的。所以系统在这段时间内并没有暂停;

3、CMS-concurrent-preclean-start :concurrent precleaning。相同的。这一步也是并发运行;
4、CMS-remark:又一次标记。暂停应用程序。同一时候启动一定数目的垃圾回收线程。并行地标记第2阶段遗漏的对象(在并发标记阶段结束后对象状态的更新导致)。

它的运行时间是real=0.97
secs。

由于这是STW的步骤,而且它的pause time通常是最长的,所以这一步的运行时间会直接决定这次Full GC对系统的影响。

5、CMS-concurrent-sweep-start:并发清除。
6、CMS-concurrent-reset-start:并发重设状态等待下一次CMS的触发。
仅仅有在第1、4阶段须要暂停全部的应用程序。其他阶段的垃圾回收线程与应用程序线程并发运行。

05-28 08:12