在HDFS1中NameNode启动顺序是这样的:
1. 读取Fsimage文件
2. 读取edit logs文件,逐行执行里面的操作
3. 写checkpoint,生成新的Fsimage(老的Fsimage + editlogs)
4. 进入safe mode,等待datanodes的block reports,直到达到最小的replication数的block百分比才退出
在安全模式期间,client是不能修改namespace信息,也不允许复制blocks,client基本上是被block住的
而且有些问题导致从namenode启动到client能请求request会耗费很长时间
1. 如果editlogs变得很大(比如由于secondary namenode服务挂了,没有及时merge一个比较新的fsimage),导致读入很大的editlogs,执行操作会比较慢
2. 一般fsimage和editlogs的文件都会做raid 1镜像, 在写新的fsimage checkpoint的时候会写多份,这就要求多份都写成功后这个操作才算成功,所以任何一块盘有性能瓶颈,都会导致延迟
另外一个问题是Namenode的Web UI Server是在写checkpoint之后才会启动的,这就导致了如果长时间在startup期间,管理员是无法直观通过WwebUI来看到整个启动进度,只能通过namenode.hadoop.log来看。
不过在2.1.0beta中已经加上了一个feature,能在WebUI上查看NM startup status( https://issues.apache.org/jira/browse/HDFS-4249),它的做法是将Web UI Server启动放到NM启动顺序的很前面,让用户可以尽早看到。而且在UI上,增加了不同stage的详细信息,包括加载的fsimage在NM节点上的绝对路径,它的文件大小,加载的inode的个数等,在safe mode的时候,也有显示已经收到的block数和block总数的占比,用户能大致估算出退出safe mode的时间。HDFS不仅仅是内部变得更健壮和稳定,在外围的用户体验也在变得越来越棒啊.
NM Startup progress:
除了WebUI,用户还可以wget http://namenode-address:50070/startupProgress,获取JSON格式StartUp信息
{
"elapsedTime": 35866,
"percentComplete": 1,
"phases": [
{
"name": "LoadingFsImage",
"status": "COMPLETE",
"percentComplete": 1,
"elapsedTime": 165,
"file": "/data/yarn/name/current/fsimage_0000000000000002434",
"size": 22763,
"steps": [
{
"name": "Inodes",
"count": 215,
"total": 215,
"percentComplete": 1,
"elapsedTime": 25
},
{
"name": "DelegationKeys",
"count": 0,
"total": 0,
"percentComplete": 1,
"elapsedTime": 0
},
{
"name": "DelegationTokens",
"count": 0,
"total": 0,
"percentComplete": 1,
"elapsedTime": 0
}
]
},
{
"name": "LoadingEdits",
"status": "COMPLETE",
"percentComplete": 1,
"elapsedTime": 171,
"steps": [
{
"count": 1,
"file": "/data/yarn/name/current/edits_0000000000000002435-0000000000000002435",
"size": 1048576,
"total": 1,
"percentComplete": 1,
"elapsedTime": 15
},
{
"count": 1044,
"file": "/data/yarn/name/current/edits_0000000000000002436-0000000000000003479",
"size": 1048576,
"total": 1044,
"percentComplete": 1,
"elapsedTime": 155
}
]
},
{
"name": "SavingCheckpoint",
"status": "COMPLETE",
"percentComplete": 1,
"elapsedTime": 77,
"steps": [
{
"name": "Inodes",
"count": 299,
"file": "/data/yarn/name",
"total": 299,
"percentComplete": 1,
"elapsedTime": 14
},
{
"name": "DelegationKeys",
"count": 0,
"file": "/data/yarn/name",
"total": 0,
"percentComplete": 1,
"elapsedTime": 0
},
{
"name": "DelegationTokens",
"count": 0,
"file": "/data/yarn/name",
"total": 0,
"percentComplete": 1,
"elapsedTime": 0
}
]
},
{
"name": "SafeMode",
"status": "COMPLETE",
"percentComplete": 1,
"elapsedTime": 35118,
"steps": [
{
"name": "AwaitingReportedBlocks",
"count": 218,
"total": 218,
"percentComplete": 1,
"elapsedTime": 0
}
]
}
]
}
参考jira:
https://issues.apache.org/jira/browse/HDFS-4249
本文链接http://blog.csdn.net/lalaguozhe/article/details/10586555,转载请注明