hbase snapshot

 

Snapshot:
(1) take a snapshot
table是在enable状态，那么使用EnabledTableSnapshotHandler
table是在disable状态，使用DisabledTableSnapshotHandler，在HMaster端完成所有操作
(2) getCompletedSnapshots
获得已完成的snapshot，正在进行中的被忽略
(3) deleteSnapshot
直接删除snapshot对应的目录
(4) isSnapshotDone
查看snapshot是否完成，从snapshotHandlers队列中获得信息
(5) restoreSnapshot
table是否存在
存在，必须是disable状态，RestoreSnapshotHandler，restore到相应的状态
不存在，CloneSnapshotHandler，clone一个table。大部分利用CreateTableHandler的代码，把handleCreateHdfsRegions这块替换掉
(6) isRestoreSnapshotDone
查看restore是否完成，从restoreHandlers队列中获得信息

MasterSnapshotVerifier是验证用的
（1） SnapshotDescription is readable
（2） Table info is readable
（3） Regions

• <li>Matching regions in the snapshot as currently in the table</li>
• <li>{@link HRegionInfo} matches the current and stored regions</li>
• <li>All referenced hfiles have valid names</li>
• <li>All the hfiles are present (either in .archive directory in the region)</li>
• <li>All recovered.edits files are present (by name) and have the correct file size</li>

 

在take a snapshot时，table在enable状态下使用 EnabledTableSnapshotHandler

HMaster做为master，执行Procedure
RegionServer做为slave执行Subprocedure，负责snapshot region
通过zk通信，完成任务的协商，slave什么时候开始执行任务（master发起一个任务，slave接收到通知后告诉master可以开始，等所有 slave都ok后就可以开始这次snapshot了），执行什么任务（snapshot那个表的region），slave任务结束后告知master 已完成，等所有slave都完成后就表明本次snapshot完成。

Master端zk处理是在ZKProcedureCoordinatorRpcs.java
Slave端的zk处理是ZKProcedureMemberRpcs.java
通过zk事件来触发相关任务和进度往下执行

 

EnabledTableSnapshotHandler

ZKProcedureCoordinatorRpcs
ZKProcedureMemberRpcs

Master 端启动一个 Procedure
(1) Master Procedure.sendGlobalBarrierAcquire
Notify the members to acquire barrier for the procedure
创建zk节点，类型是acquire znode，比如snapshot名字是t1-s2，那么节点是/hbase/online-snapshot/acquired/t1-s2
等待各个Subprocedure（HRegionServer上面）收到zk事件，并创建相应的zk节点

(2) HRegionServer监听到zk事件，发现是acquiredZnode节点
启动一个Subprocedure，Subprocedure.acquireBarrier
在acquire znode节点下创建一个子节点
等待本次Procedure到达reached状态，Subprocedure.waitForReachedGlobalBarrier

(3) Master收到acquire znode节点下面有子节点创建的消息，进行计数
当收到所有节点创建了子节点后就可以往后执行了，也就是达成一致

(4) Master Procedure.sendGlobalBarrierReached

• Notify members that all members have acquired their parts of the barrier and that they can now execute under the global barrier.
  创建zk节点，类型是reached znode，/hbase/online-snapshot/reached/t1-s2
  等待所有成员完成任务，wait for all members to report barrier release

(5) HRegionServer监听到zk事件，发现是reachedZnode节点，触发Subprocedure的receiveReachedGlobalBarrier
Subprocedure往下执行，Subprocedure.insideBarrier，这里的Subprocedure是FlushSnapshotSubprocedure
在Subprocedure.insideBarrier里面执行HRegion的flush和snapshot
完成任务后，Subprocedure在reachedZnode节点下创建子节点说明他已经完成了任务，并且结束Subprocedure

(6) Master端收到reachedZnode节点下面有子节点创建的消息，进行计数
当收到所有Subprocedure创建子节点的消息后，就可以说明所有任务都执行完成了，往下继续执行

(7) Procedure.sendGlobalBarrierComplete

本次Procedure结束

 

 

Snapshot的目录结构见SnapshotDescriptionUtils

* Snapshots are laid out on disk like this:
 *
 * <pre>
 * /hbase/.snapshots
 *          /.tmp                <---- working directory
 *          /[snapshot name]     <----- completed snapshot
 * </pre>
 *
 * A completed snapshot named 'completed' then looks like (multiple regions, servers, files, etc.
 * signified by '...' on the same directory depth).
 *
 * <pre>
 * /hbase/.snapshots/completed
 *                   .snapshotinfo          <--- Description of the snapshot
 *                   .tableinfo             <--- Copy of the tableinfo
 *                    /.logs
 *                        /[server_name]
 *                            /... [log files]
 *                         ...
 *                   /[region name]           <---- All the region's information
 *                   .regioninfo              <---- Copy of the HRegionInfo
 *                      /[column family name]
 *                          /[hfile name]     <--- name of the hfile in the real region
 *                          ...
 *                      ...
 *                    ...