MVCC(Multi-Version Concurrent Control),即多版本并发控制协议,广泛使用于数据库系统。本文将介绍HBase中对于MVCC的实现及应用情况。
MVCC基本原理
在介绍MVCC概念之前,我们先来想一下数据库系统里的一个问题:假设有多个用户同时读写数据库里的一行记录,那么怎么保证数据的一致性呢?一个基本的解决方法是对这一行记录加上一把锁,将不同用户对同一行记录的读写操作完全串行化执行,由于同一时刻只有一个用户在操作,因此一致性不存在问题。但是,它存在明显的性能问题:读会阻塞写,写也会阻塞读,整个数据库系统的并发性能将大打折扣。
MVCC(Multi-Version Concurrent Control),即多版本并发控制协议,它的目标是在保证数据一致性的前提下,提供一种高并发的访问性能。在MVCC协议中,每个用户在连接数据库时看到的是一个具有一致性状态的镜像,每个事务在提交到数据库之前对其他用户均是不可见的。当事务需要更新数据时,不会直接覆盖以前的数据,而是生成一个新的版本的数据,因此一条数据会有多个版本存储,但是同一时刻只有最新的版本号是有效的。因此,读的时候就可以保证总是以当前时刻的版本的数据可以被读到,不论这条数据后来是否被修改或删除。
更多关于MVCC基本思想的介绍,参考Wikipedia。
一个MVCC实现类
见org.apache.hadoop.hbase.regionserver.MultiVersionConsistencyControl,用于控制Memstore中读写的一致性,其中维护两个long型的变量:
1)memstoreRead:用于记录当前全局可读的readPoint,同时为了每个客户端读请求能够记录自己发起请求时刻的readPoint,还有一个ThreadLocal的perThreadReadPoint变量,以及相关的set和get方法;
2)memstoreWrite:用于记录当前全局最大的writePoint,根据它为下个事务生成新的writePoint。
MultiVersionConsistencyControl中关键的实现方法如下:
1)WriteEntry beginMemstoreInsert():开始一个更新操作,将memstoreWrite加1,创建writeQueue并插入到writeQueue,并返回WriteEntry对象;
2)void completeMemstoreInsert(WriteEntry e):完成当前更新操作,将WriteEntry对象标记为可读,具体分两步:
boolean advanceMemstore(WriteEntry e):从头开始遍历writeQueue,移除所有已完成的WriteEntry对象,最后将memstoreRead更新为最新已完成的memstoreWrite;
void waitForRead(WriteEntry e):阻塞当前线程,直到memstoreRead等于当前WriteEntry的memstoreWrite,至此表明当前WriteEntry之前的所有更新事务都已经完成。
MVCC使用场景
见org.apache.hadoop.hbase.regionserver.HRegion.java,每个Region包含一个Memstore,维护一个MultiVersionConsistencyControl对象。
写操作
见HRegion.java中的以下写操作的方法:
1)put
2)checkAndPut
3)delete
4)checkAndDelete
5)internalFlushcache
6)mutateRow
7)mutateRowsWithLocks
8)batchMutate
最终会调用到applyFamilyMapToMemstore方法使用MVCC进行写操作:
/*** Atomically apply the given map of family->edits to the memstore.
* This handles the consistency control on its own, but the caller
* should already have locked updatesLock.readLock(). This also does
* not check the families for validity.
*
*@paramfamilyMap Map of kvs per family
*@paramlocalizedWriteEntry The WriteEntry of the MVCC for this transaction.
* If null, then this method internally creates a mvcc transaction.
*@returnthe additional memory usage of the memstore caused by the
* new entries.*/
private long applyFamilyMapToMemstore(Map>familyMap,
MultiVersionConsistencyControl.WriteEntry localizedWriteEntry) {long size = 0;boolean freemvcc = false;try{if (localizedWriteEntry == null) {
localizedWriteEntry=mvcc.beginMemstoreInsert();
freemvcc= true;
}for (Map.Entry>e : familyMap.entrySet()) {byte[] family =e.getKey();
List edits =e.getValue();
Store store=getStore(family);for(KeyValue kv: edits) {
kv.setMemstoreTS(localizedWriteEntry.getWriteNumber());
size+=store.add(kv);
}
}
}finally{if(freemvcc) {
mvcc.completeMemstoreInsert(localizedWriteEntry);
}
}returnsize;
}
View Code
读操作
HRegion.java中通过private ConcurrentHashMap scannerReadPoints;维护各个查询请求的readPoint。
以get或scan请求为例,最终会通过getScanner方法需要构造RegionScannerImpl对象:
org.apache.hadoop.hbase.regionserver.HRegion.RegionScannerImpl:
1)根据Scan对象构造时设置好readPoint,scan.getIsolationLevel()分为READ_UNCOMMITTED和READ_COMMITTED,只有当READ_COMMITTED时根据MultiVersionConsistencyControl.resetThreadReadPoint(mvcc);设置当前scanner线程的readPoint,并插入到scannerReadPoints维护起来。
2)根据scan需要读取的column family,创建StoreScanner(根据bloom filter、time range、ttl筛选需要的MemStoreScanner和StoreFileScanner),添加到scanners中,并最终根据scanners构造出一个KeyValueHeap。
下面看下RegionScannerImpl中的next方法是每次查询时需要调用的函数:
boolean org.apache.hadoop.hbase.regionserver.HRegion.RegionScannerImpl.next(List outResults, int limit) throws IOException
而上述方法会通过KeyValueHeap的next方法读取下一条数据:先定位到当前KeyValueScanner(即之前构造KeyValueHeap时传入的MemStoreScanner或StoreScanner),然后调用next方法。
StoreFileScanner和MemStoreScanner均为KeyValueScanner,通过其中的next()接口方法,分别调用到StoreFileScanner.java的skipKVsNewerThanReadpoint方法、Memstore.java中MemStoreScanner对象的getNext方法。
1)StoreFileScanner.java的skipKVsNewerThanReadpoint方法:
protected boolean skipKVsNewerThanReadpoint() throwsIOException {long readPoint =MultiVersionConsistencyControl.getThreadReadPoint();//We want to ignore all key-values that are newer than our current//readPoint
while(enforceMVCC&& cur != null
&& (cur.getMemstoreTS() >readPoint)) {
hfs.next();
cur=hfs.getKeyValue();
}if (cur == null) {
close();return false;
}//For the optimisation in HBASE-4346, we set the KV's memstoreTS to//0, if it is older than all the scanners' read points. It is possible//that a newer KV's memstoreTS was reset to 0. But, there is an//older KV which was not reset to 0 (because it was//not old enough during flush). Make sure that we set it correctly now,//so that the comparision order does not change.
if (cur.getMemstoreTS() <=readPoint) {
cur.setMemstoreTS(0);
}return true;
}
View Code
2) Memstore.java中MemStoreScanner对象的getNext方法:
protected KeyValue getNext(Iteratorit) {long readPoint =MultiVersionConsistencyControl.getThreadReadPoint();while(it.hasNext()) {
KeyValue v=it.next();if (v.getMemstoreTS() <=readPoint) {returnv;
}
}return null;
}
View Code