hbase中HMaster进程自动掉了

转载

mob64ca13f96cda 2024-09-13 12:15:42

文章标签 hbase中HMaster进程自动掉了 3c 5e d3 文章分类 Hbase 数据库

原因可能是hbase集群在split操作时维护人员强行kill hbase造成的

echo 'scan "hbase:meta"'|hbase shell|grep "hbase:namespace"
hbase hbck -repair

最近生产环境hbase集群出现停掉集群之后hmaster无法启动现象，
master日志报异常：Timedout 300000ms waiting for namespace table to be assigned，导致整个集群起不来。

2016-12-12 18:04:12,243 FATAL [adfnn2:16020.activeMasterManager] master.HMaster: Failed to become active master
java.io.IOException: Timedout 300000ms waiting for namespace table to be assigned
        at org.apache.hadoop.hbase.master.TableNamespaceManager.start(TableNamespaceManager.java:98)
        at org.apache.hadoop.hbase.master.HMaster.initNamespace(HMaster.java:868)
        at org.apache.hadoop.hbase.master.HMaster.finishActiveMasterInitialization(HMaster.java:719)
        at org.apache.hadoop.hbase.master.HMaster.access$500(HMaster.java:165)
        at org.apache.hadoop.hbase.master.HMaster$1.run(HMaster.java:1425)

首先怀疑regionserver加载namespace region出现异常，导致namespace表加载失败，但是查看所有regionserver日志，没有异常出现。

那么就分析master日志，更改log4j日志为debug，看看能不能找到关键信息，有没可能是region太多，在五分钟之内没有分配完导致namespace表还没加载，看下面的region分配日志，2016-12-12 17:59:36,571之后再也没有region分配日志，说明所有region已经分配完成，再看看上面异常日志，2016-12-12 18:04:12,243报的异常，17:59:36,571至18:04:12,243期间没有region分配，的确说明region已经全部分配完成。

提取region分配日志，cat hbase-ocdc-master-adfnn2.log|grep “deleted, state:”|awk ‘{print $7}’|awk -F ‘,’ ‘{print $1}’|sort|uniq，看看哪些表已经被分配

SYSTEM.SEQUENCE
THB_GPRS_CHARGE_201610
THB_GPRS_CHARGE_201611
THB_GPRS_CHARGE_201612
THB_GPRS_FLOW_201610
THB_GPRS_FLOW_201611
THB_GPRS_FLOW_201612
THB_GPRS_WAP_201610
THB_GPRS_WAP_201612
THB_USER_FLOW_201610
THB_USER_FLOW_201611
THB_USER_FLOW_201612
THB_USER_INFO_201609
THB_USER_INFO_201610
THB_USER_INFO_201611
THB_USER_INFO_201612
THB_USER_INFO_DAY
aidata:user_phone_info_201508
aidata:user_phone_info_201511
aidata:user_phone_info_201512
hbase:meta
复制代码
hbase:namespace表没有被分配？为什么没有被分配？难道是namespace表有问题导致不能分配？于是让同事把生产环境namespace表目录取下来放到自己的开发环境，namespace表目录比较乱，namespace还嵌namespace，我相信肯定不是维护人员干的，即使namespace嵌套，也不应该影响namepace的region加载。

复制代码

wangkai8@wangkai8 $ hadoop fs -lsr /namespace
 lsr: DEPRECATED: Please use ‘ls -R’ instead.
 -rw-r–r-- 1 wangkai8 supergroup 6148 2016-12-12 15:33 /namespace/.DS_Store
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/.tabledesc
 -rw-r–r-- 1 wangkai8 supergroup 312 2016-12-12 15:33 /namespace/.tabledesc/.tableinfo.0000000001
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/.tmp
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8
 -rw-r–r-- 1 wangkai8 supergroup 6148 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/.DS_Store
 -rw-r–r-- 1 wangkai8 supergroup 42 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/.regioninfo
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/info
 -rw-r–r-- 1 wangkai8 supergroup 4989 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/info/5b15bf56f4ea443ba6b7e212760f8644
 -rw-r–r-- 1 wangkai8 supergroup 4971 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/info/f1c4c097645344269e475ad32cf79913
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/recovered.edits
 -rw-r–r-- 1 wangkai8 supergroup 341 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/recovered.edits/0000000000000000480
 -rw-r–r-- 1 wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/recovered.edits/11.seqid
 -rw-r–r-- 1 wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/recovered.edits/2.seqid
 -rw-r–r-- 1 wangkai8 supergroup 0 2016-12-12 15:33 /namespace/83739a75edfb427d8501a43c553caad8/recovered.edits/479.seqid
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/namespace
 -rw-r–r-- 1 wangkai8 supergroup 6148 2016-12-12 15:33 /namespace/namespace/.DS_Store
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/namespace/.tabledesc
 -rw-r–r-- 1 wangkai8 supergroup 312 2016-12-12 15:33 /namespace/namespace/.tabledesc/.tableinfo.0000000001
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/namespace/.tmp
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 15:33 /namespace/namespace/51c319c701fe6fdf6b77da884541a4dd
 -rw-r–r-- 1 wangkai8 supergroup 42 2016-12-12 15:33 /namespace/namespac

e/51c319c701fe6fdf6b77da884541a4dd/.regioninfo

复制代码
把生产环境/hbase1.0.0/data/hbase/namespace/namespace删掉之后再启动集群，结果还是报一样的异常错误，好吧，那就把/hbase1.0.0/data/hbase/namespace/83739a75edfb427d8501a43c553caad8/recovered*删掉再启动集群，结果还是一样。

为了看namespace表到底有哪些记录，把/hbase1.0.0/data/hbase/namespace/83739a75edfb427d8501a43c553caad8/info下两个HFile考到自己开发环境namespace目录下，然后重启hbase集群，可以看到生产环境namespace有哪些记录。

复制代码

Set<TableName> disabledOrDisablingOrEnabling = tableStateManager.getTablesInStates(
  ZooKeeperProtos.Table.State.DISABLED,
  ZooKeeperProtos.Table.State.DISABLING,
  ZooKeeperProtos.Table.State.ENABLING);

// Region assignment from META
List<Result> results = MetaTableAccessor.fullScanOfMeta(server.getConnection());
// Get any new but slow to checkin region server that joined the cluster
Set<ServerName> onlineServers = serverManager.getOnlineServers().keySet();
// Set of offline servers to be returned
Set<ServerName> offlineServers = new HashSet<ServerName>();
// Iterate regions in META
for (Result result : results) {
  if (result == null && LOG.isDebugEnabled()){
    LOG.debug("null result from meta - ignoring but this is strange.");
    continue;
  }
  RegionLocations rl =  MetaTableAccessor.getRegionLocations(result);
  if (rl == null) continue;
  HRegionLocation[] locations = rl.getRegionLocations();
  if (locations == null) continue;
  for (HRegionLocation hrl : locations) {
    HRegionInfo regionInfo = hrl.getRegionInfo();
    if (regionInfo == null) continue;
    int replicaId = regionInfo.getReplicaId();
    State state = RegionStateStore.getRegionState(result, replicaId);
    ServerName lastHost = hrl.getServerName();　　//由于namespace表info:server为null，导致lastHost也为null
    ServerName regionLocation = RegionStateStore.getRegionServer(result, replicaId);
    regionStates.createRegionState(regionInfo, state, regionLocation, lastHost);  //进而导致namespace region的状态为State.OFFLINE，master不分配此region
    if (!regionStates.isRegionInState(regionInfo, State.OPEN)) {
      // Region is not open (either offline or in transition), skip
      continue;
    }
    TableName tableName = regionInfo.getTable();
    if (!onlineServers.contains(regionLocation)) {
      // Region is located on a server that isn't online
      offlineServers.add(regionLocation);
      if (useZKForAssignment) {
        regionStates.regionOffline(regionInfo);
      }
    } else if (!disabledOrEnablingTables.contains(tableName)) {
      // Region is being served and on an active server
      // add only if region not in disabled or enabling table
      regionStates.regionOnline(regionInfo, regionLocation);
      balancer.regionOnline(regionInfo, regionLocation);
    } else if (useZKForAssignment) {
      regionStates.regionOffline(regionInfo);
    }
    // need to enable the table if not disabled or disabling or enabling
    // this will be used in rolling restarts
    if (!disabledOrDisablingOrEnabling.contains(tableName)
      && !getTableStateManager().isTableState(tableName,
        ZooKeeperProtos.Table.State.ENABLED)) {
      setEnabledTable(tableName);
    }
  }
}
return offlineServers;

Set<TableName> disabledOrDisablingOrEnabling = tableStateManager.getTablesInStates(
  ZooKeeperProtos.Table.State.DISABLED,
  ZooKeeperProtos.Table.State.DISABLING,
  ZooKeeperProtos.Table.State.ENABLING);

// Region assignment from META
List<Result> results = MetaTableAccessor.fullScanOfMeta(server.getConnection());
// Get any new but slow to checkin region server that joined the cluster
Set<ServerName> onlineServers = serverManager.getOnlineServers().keySet();
// Set of offline servers to be returned
Set<ServerName> offlineServers = new HashSet<ServerName>();
// Iterate regions in META
for (Result result : results) {
  if (result == null && LOG.isDebugEnabled()){
    LOG.debug("null result from meta - ignoring but this is strange.");
    continue;
  }
  RegionLocations rl =  MetaTableAccessor.getRegionLocations(result);
  if (rl == null) continue;
  HRegionLocation[] locations = rl.getRegionLocations();
  if (locations == null) continue;
  for (HRegionLocation hrl : locations) {
    HRegionInfo regionInfo = hrl.getRegionInfo();
    if (regionInfo == null) continue;
    int replicaId = regionInfo.getReplicaId();
    State state = RegionStateStore.getRegionState(result, replicaId);
    ServerName lastHost = hrl.getServerName();　　//由于namespace表info:server为null，导致lastHost也为null
    ServerName regionLocation = RegionStateStore.getRegionServer(result, replicaId);
    regionStates.createRegionState(regionInfo, state, regionLocation, lastHost);  //进而导致namespace region的状态为State.OFFLINE，master不分配此region
    if (!regionStates.isRegionInState(regionInfo, State.OPEN)) {
      // Region is not open (either offline or in transition), skip
      continue;
    }
    TableName tableName = regionInfo.getTable();
    if (!onlineServers.contains(regionLocation)) {
      // Region is located on a server that isn't online
      offlineServers.add(regionLocation);
      if (useZKForAssignment) {
        regionStates.regionOffline(regionInfo);
      }
    } else if (!disabledOrEnablingTables.contains(tableName)) {
      // Region is being served and on an active server
      // add only if region not in disabled or enabling table
      regionStates.regionOnline(regionInfo, regionLocation);
      balancer.regionOnline(regionInfo, regionLocation);
    } else if (useZKForAssignment) {
      regionStates.regionOffline(regionInfo);
    }
    // need to enable the table if not disabled or disabling or enabling
    // this will be used in rolling restarts
    if (!disabledOrDisablingOrEnabling.contains(tableName)
      && !getTableStateManager().isTableState(tableName,
        ZooKeeperProtos.Table.State.ENABLED)) {
      setEnabledTable(tableName);
    }
  }
}
return offlineServers;

wangkai8@wangkai8 ~$ hadoop fs -lsr /hbase1.0.0/data/hbase/namespace
 lsr: DEPRECATED: Please use ‘ls -R’ instead.
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tabledesc
 -rw-r–r-- 1 wangkai8 supergroup 312 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tabledesc/.tableinfo.0000000001
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tmp
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 20:30 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79
 -rw-r–r-- 1 wangkai8 supergroup 42 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/.regioninfo
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 20:30 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info
 -rw-r–r-- 3 wangkai8 supergroup 5041 2016-12-12 20:30 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info/89830776b101405e9cb2e13d8a770ca6
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-15 15:58 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/recovered.edits
 -rw-r–r-- 3 wangkai8 supergroup 0 2016-12-15 15:58 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/recovered.edits/22.seqid
 wangkai8@wangkai8 ~$ cd Downloads/
 wangkai8@wangkai8 Downloads$ cd namespace
 wangkai8@wangkai8 namespace$ cd 83739a75edfb427d8501a43c553caad8/
 wangkai8@wangkai8 83739a75edfb427d8501a43c553caad8$ ls
 info recovered.edits
 wangkai8@wangkai8 83739a75edfb427d8501a43c553caad8$ hadoop fs -put info/* /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info/
 wangkai8@wangkai8 83739a75edfb427d8501a43c553caad8$ hadoop fs -lsr /hbase1.0.0/data/hbase/namespace
 lsr: DEPRECATED: Please use ‘ls -R’ instead.
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tabledesc
 -rw-r–r-- 1 wangkai8 supergroup 312 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tabledesc/.tableinfo.0000000001
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/.tmp
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-12 20:30 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79
 -rw-r–r-- 1 wangkai8 supergroup 42 2016-12-12 19:16 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/.regioninfo
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-15 16:01 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info
 -rw-r–r-- 1 wangkai8 supergroup 4989 2016-12-15 16:01 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info/5b15bf56f4ea443ba6b7e212760f8644
 -rw-r–r-- 3 wangkai8 supergroup 5041 2016-12-12 20:30 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info/89830776b101405e9cb2e13d8a770ca6
 -rw-r–r-- 1 wangkai8 supergroup 4971 2016-12-15 16:01 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/info/f1c4c097645344269e475ad32cf79913
 drwxr-xr-x - wangkai8 supergroup 0 2016-12-15 15:58 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/recovered.edits
 -rw-r–r-- 3 wangkai8 supergroup 0 2016-12-15 15:58 /hbase1.0.0/data/hbase/namespace/be6712b3ac638d312c7a4544514a6f79/recovered.edits/22.seqid
 复制代码
 复制代码
 hbase(main):002:0> scan ‘hbase:namespace’
 ROW COLUMN+CELL
 default column=info:d, timestamp=1481541361405, value=\x0A\x07default
 hbase column=info:d, timestamp=1481541361412, value=\x0A\x05hbase
 ocnosql column=info:d, timestamp=1481541506852, value=\x0A\x07ocnosql
 3 row(s) in 0.0980 secondshbase(main):003:0> scan ‘hbase:namespace’
 ROW COLUMN+CELL
 aidata column=info:d, timestamp=1441764317452, value=\x0A\x06aidata
 default column=info:d, timestamp=1481541361405, value=\x0A\x07default
 hbase column=info:d, timestamp=1481541361412, value=\x0A\x05hbase
 ocnosql column=info:d, timestamp=1481541506852, value=\x0A\x07ocnosql
 4 row(s) in 0.0180 seconds
 复制代码
 可以看到namespace多了一条记录，说明namespace hfile也有没问题，那么问题到底在哪？难道hbase:meta表有问题，又让维护同事把meta目录取下来，meta表的HFile就不能直接放到hbase meta目录下了，否则生产环境的region在开发环境肯定找不到导致启动失败。可以通过HFile工具查看HFile里面的数据：wangkai8@wangkai8 info$ hbase org.apache.hadoop.hbase.io.hfile.HFile -e -p -f /meta/1588230740/info/8e498a36f195450d91e07eb8ad4208c1|grep ‘hbase:namespace’
 K: hbase:namespace,1481192140855.83739a75edfb427d8501a43c553caad8./info:regioninfo/1481247939963/Put/vlen=41/seqid=10742 V: PBUF\x08\xB7\xE0\xB9\xEF\x8D+\x12\x12\x0A\x05hbase\x12\x09namespace\x1A\x00"\x00(\x000\x008\x00
 wangkai8@wangkai8 info$ hbase org.apache.hadoop.hbase.io.hfile.HFile -e -p -f /meta/1588230740/info/bc669ae58e5142d69b6fb38979fcd20c|grep ‘hbase:namespace’
 wangkai8@wangkai8 info$
 可以看到只有一条记录，且只有一个列，正常namespace表如下wangkai8@wangkai8 ~$ echo ‘scan “hbase:meta”’|hbase shell|grep “hbase:namespace”
 hbase:namespace,1481791022097.bd5c5e4af06be7540ce00e24c8949399. column=info:regioninfo, timestamp=1481791022263, value={ENCODED => bd5c5e4af06be7540ce00e24c8949399, NAME => ‘hbase:namespace,1481791022097.bd5c5e4af06be7540ce00e24c8949399.’, STARTKEY => ‘’, ENDKEY => ‘’}
 hbase:namespace,1481791022097.bd5c5e4af06be7540ce00e24c8949399. column=info:seqnumDuringOpen, timestamp=1481791022325, value=\x00\x00\x00\x00\x00\x00\x00\x02
 hbase:namespace,1481791022097.bd5c5e4af06be7540ce00e24c8949399. column=info:server, timestamp=1481791022325, value=wangkai8dembp:61020
 hbase:namespace,1481791022097.bd5c5e4af06be7540ce00e24c8949399. column=info:serverstartcode, timestamp=1481791022325, value=1481791009329
 也就是说生产环境namespace表缺少重要的info:server信息，这将导致namespace region在被AssignmentManager分配时跳过。再次查看hbase源码，region分配过程大致如下：org.apache.hadoop.hbase.master.HMaster#finishActiveMasterInitialization->
org.apache.hadoop.hbase.master.AssignmentManager#joinCluster->
org.apache.hadoop.hbase.master.AssignmentManager#rebuildUserRegions->
org.apache.hadoop.hbase.master.AssignmentManager#processDeadServersAndRegionsInTransition->
org.apache.hadoop.hbase.master.AssignmentManager#assignAllUserRegions->
org.apache.hadoop.hbase.master.AssignmentManager#assign(java.util.Map<org.apache.hadoop.hbase.HRegionInfo,org.apache.hadoop.hbase.ServerName>)->
#创建region bulk plan分配计划
org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer#retainAssignment->
#根据region plan开始真正执行region assign
org.apache.hadoop.hbase.master.AssignmentManager#assign(int, int, String, Map<ServerName,List>)->
org.apache.hadoop.hbase.master.AssignmentManager#assign(ServerName, List)->
//在zookeeper /hbase1.0.0/region-in-transition下创建待分配region信息
org.apache.hadoop.hbase.master.AssignmentManager#asyncSetOfflineInZooKeeper->
org.apache.hadoop.hbase.zookeeper.ZKAssign#asyncCreateNodeOffline->
org.apache.hadoop.hbase.zookeeper.ZKUtil#asyncCreate
生产环境这个异常问题在org.apache.hadoop.hbase.master.AssignmentManager#rebuildUserRegions这个方法，先来看看这个方法里面有什么内容
复制代码
 /**• Rebuild the list of user regions and assignment information.
•  
• Returns a set of servers that are not found to be online that hosted
• some regions.
• @return set of servers not online that hosted some regions per meta
• @throws IOException
 */
 Set rebuildUserRegions() throws
 IOException, KeeperException, CoordinatedStateException {
 Set disabledOrEnablingTables = tableStateManager.getTablesInStates(
 ZooKeeperProtos.Table.State.DISABLED, ZooKeeperProtos.Table.State.ENABLING);

}
/**
• Add a region to RegionStates with the specified state.
• If the region is already in RegionStates, this call has
• no effect, and the original state is returned.
• 
• @param hri the region info to create a state for
• @param newState the state to the region in set to
• @param serverName the server the region is transitioning on
• @param lastHost the last server that hosts the region
• @return the current state
 */
 public synchronized RegionState createRegionState(final HRegionInfo hri,
 State newState, ServerName serverName, ServerName lastHost) {
 if (newState == null || (newState == State.OPEN && serverName == null)) {　　//如果region lastHost为null，这个region直接被指定为State.OFFLINE
 newState = State.OFFLINE;
 }
 if (hri.isOffline() && hri.isSplit()) {
 newState = State.SPLIT;
 serverName = null;
 }
 String encodedName = hri.getEncodedName();
 RegionState regionState = regionStates.get(encodedName);
 if (regionState != null) {
 LOG.warn("Tried to create a state for a region already in RegionStates, " 
  • "used existing: " + regionState + ", ignored new: " + newState);
 } else {
 regionState = new RegionState(hri, newState, serverName);
 regionStates.put(encodedName, regionState);
 if (newState == State.OPEN) {
 if (!serverName.equals(lastHost)) {
 LOG.warn("Open region’s last host " + lastHost 
    • " should be the same as the current one " + serverName
• “, ignored the last and used the current one”);
 lastHost = serverName;
 }
 lastAssignments.put(encodedName, lastHost);
 regionAssignments.put(hri, lastHost);
 } else if (!regionState.isUnassignable()) {
 regionsInTransition.put(encodedName, regionState);
 }
 if (lastHost != null && newState != State.SPLIT) {
 addToServerHoldings(lastHost, hri);
 if (newState != State.OPEN) {
 oldAssignments.put(encodedName, lastHost);
 }
 }
 }
 return regionState;
 }

复制代码
到此根本问题已经定位，meta表数据丢失，原因可能是hbase集群在split操作时维护人员强行kill hbase造成，不管是什么原因导致这个问题，肯定跟hbase自身健壮性有关，不管维护人员在何种状态下stop或kill hbase，也不应该出现meta有问题。meta表本身可以通过hbase hbck -repair来修复，很尴尬的是在集群没启动的情况下hbase hbck用不了。

既然找到了问题，那就想办法解决。麻烦的是集群起不来，meta表没法修复，解决思想有两种：

方案1）修改hbase源码，如果region所属namespace，且server信息为空，人为指定一台regionserver给这个region，使其能被master分配出去，集群能起来，集群起来之后就好办，可以hbck repair meta，也可以进到hbase shell操作meta表。meta表修复之后stop集群，把hbase原来的jar包换回来，启动集群就ok了。

方案2）在开发环境meta表中put一条namespace info:server记录，指定一个regionserver地址，强制执行flush ‘hbase:meta’，生成包含namespace server信息的HFile，然后把HFile放到生产环境meta目录下，重启之后就可以恢复namespace server信息。

这次采用的解决方案是修改hbase源码，方案二没去验证，根据理论，应该是没问题。

/**
• org.apache.hadoop.hbase.MetaTableAccessor#getServerName
• 修改此方法，给namespace region指定一个regionserver地址
• Returns a {@link ServerName} from catalog table {@link Result}.
• @param r Result to pull from
• @return A ServerName instance or null if necessary fields not found or empty.
 */
 private static ServerName getServerName(final Result r, final int replicaId) {
 byte[] serverColumn = getServerColumn(replicaId);
 Cell cell = r.getColumnLatestCell(getFamily(), serverColumn);
//---------------------------------------------------------------------------
 HRegionInfo regionInfo = getHRegionInfo(r, getRegionInfoColumn());
 String tableName = regionInfo.getTable().getNameAsString();
 if((cell == null || cell.getValueLength() == 0) && tableName.equals(“hbase:namespace”)) {
 return ServerName.valueOf(“adfdn14:16020”, 1481534251178L);
 }
 //---------------------------------------------------------------------------

if (cell == null || cell.getValueLength() == 0) return null;
String hostAndPort = Bytes.toString(
  cell.getValueArray(), cell.getValueOffset(), cell.getValueLength());
byte[] startcodeColumn = getStartCodeColumn(replicaId);
cell = r.getColumnLatestCell(getFamily(), startcodeColumn);
if (cell == null || cell.getValueLength() == 0) return null;
return ServerName.valueOf(hostAndPort,
  Bytes.toLong(cell.getValueArray(), cell.getValueOffset(), cell.getValueLength()));

}

至此集群已经能正常启动，花了一天时间解决故障。

再来谈谈hbase的设计，master启动时首先分配meta表，然后再分配其它表，但是hbase:namespace系统表和其它用户表分配时同等对待，并没有先分配系统表再分配用户表，如果一个集群region非常多，默认5分钟可能还分配不到namespace表，还得修改hbase.master.namespace.init.timeout超时时间，总感觉有点不合理。

hbase hbck工具是很好用，但是在集群启动不了的情况下，meta表数据没法通过工具来修改，是否可以提供一个离线工具来方便生成HFile，或者提供一个miniCluster在集群瘫痪下能启动起来管理meta、namespace表。

本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。