HBase Scan流程分析
HBase Scan流程分析
HBase的读流程目前看来比较复杂,主要由于:
- HBase的表数据分为多个层次,HRegion->HStore->[HFile,HFile,...,MemStore]
- RegionServer的LSM-Like存储引擎,不断flush产生新的HFile,同时产生新的MemStore用于后续数据写入,并且为了防止由于HFile过多而导致Scan时需要扫描的文件过多而导致的性能下降,后台线程会适时的进行Compaction,Compaction的过程会产生新的HFile,并且会删除Compact完成的HFile
- 具体实现中的各种优化,比如lazy seek优化,导致代码比较复杂
读流程中充斥着各种Scanner,如下图:
+--------------+
| |
+-----------+ RegionScanner+----------+
| +------+-------+ |
| | |
| | |
+-----v+-------+ +------v-------+ +------v+------+
| | | | | |
| StoreScanner | | StoreScanner | | StoreScanner |
| | | | | |
+--------------+ +--+---+-----+-+ +--------------+
| | |
+-----------------------+ | +----------+
| | |
| | |
+-------v---------+ +-------------v----+ +---------v------+
| | | | | |
|StoreFileScanner | |StoreFileScanner | | MemStoreScanner|
| | | | | |
+-------+---------+ +--------+---------+ +-------+--------+
| | |
| | |
| | |
| | |
+-------v---------+ +--------v---------+ +-------v--------+
| | | | | |
| HFileScanner | | HFileScanner | | HFileScanner |
| | | | | |
+-----------------+ +------------------+ +----------------+
在HBase中,一张表可以有多个Column Family,在一次Scan的流程中,每个Column Family(后续叫Store)的数据读取由一个StoreScanner对象负责。每个Store的数据由一个内存中的MemStore和磁盘上的HFile文件组成,相对应的,StoreScanner对象雇佣一个MemStoreScanner和N个StoreFileScanner来进行实际的数据读取。
从逻辑上看,读取一行的数据需要
- 按照顺序读取出每个Store
- 对于每个Store,合并Store下面的相关的HFile和内存中的MemStore
实现上,这两步都是通过堆完成。RegionScanner的读取通过下面的多个StoreScanner组成的堆
完成,使用RegionScanner的成员变量KeyValueHeap storeHeap表示
组成StoreScanner的多个Scanner在RegionScannerImpl构造函数中获得:
for (Map.Entry<byte[], NavigableSet<byte[]>> entry :
scan.getFamilyMap().entrySet()) {
Store store = stores.get(entry.getKey());
// 实际是StoreScanner类型
KeyValueScanner scanner = store.getScanner(scan, entry.getValue(), this.readPt);
if (this.filter == null || !scan.doLoadColumnFamiliesOnDemand()
|| this.filter.isFamilyEssential(entry.getKey())) {
scanners.add(scanner);
} else {
joinedScanners.add(scanner);
}
}
store.getScanner(scan, entry.getValue(), this.readPt)内部就是new 一个StoreScanner,逻辑都在StoreScanner的构造函数中
构造函数内部其实就是找到相关的HFile和MemStore,然后建堆,注意,这个堆是StoreScanner级别的,一个StoreScanner一个堆,堆中的元素就是底下包含的HFile和MemStore对应的StoreFileScanner和MemStoreScanner
得到相关的HFile和MemStore逻辑在StoreScanner::getScannersNoCompaction()中,内部会根据请求指定的TimeRange,KeyRange过滤掉不需要的HFile,同时也会利用bloom filter过滤掉不需要的HFIle.接着,调用
seekScanners(scanners, matcher.getStartKey(), explicitColumnQuery && lazySeekEnabledGlobally,
isParallelSeekEnabled);
对这些StoreFileScanner和MemStoreScanner分别进行seek,seekKey是matcher.getStartKey(),
如下构造
return new KeyValue(row, family, null, HConstants.LATEST_TIMESTAMP,
Type.DeleteFamily);
Seek语义
seek是针对KeyValue的,seek的语义是seek到指定KeyValue,如果指定KeyValue不存在,则seek到指定KeyValue的下一
个。举例来说,假设名为X的column family里有两列a和b,文件中有两行rowkey分别为aaa和
bbb,如下表所示.
Column Family X | ||
rowkey | column a | column b |
aaa | 1 | abc |
bbb | 2 | def |
HBase客户端设置scan请求的start key为aaa,那么matcher.getStartKey()会被初始化为(rowkey, family, qualifier,timestamp,type)=(aaa,X,null,LATEST_TIMESTAMP,Type.DeleteFamily),根据KeyValue的比较原则,这个KeyValue比aaa行的第一个列a更
小(因为没有qualifier),所以对这个StoreFileScanner seek时,会seek到aaa这行的第一列a
实际上
seekScanners(scanners, matcher.getStartKey(), explicitColumnQuery && lazySeekEnabledGlobally,
isParallelSeekEnabled);
有可能不会对StoreFileScanner进行实际的seek,而是进行lazy seek,seek的工作放到不得不做的时候。后续会专门说lazy seek
上面得到了请求scan涉及到的所有的column family对应的StoreScanner,随后调用如下函数进行建堆:
protected void initializeKVHeap(List<KeyValueScanner> scanners,
List<KeyValueScanner> joinedScanners, HRegion region)
throws IOException {
this.storeHeap = new KeyValueHeap(scanners, region.comparator);
if (!joinedScanners.isEmpty()) {
this.joinedHeap = new KeyValueHeap(joinedScanners, region.comparator);
}
}
KeyValueScanner是一个接口,表示一个可以向外迭代出KeyValue
的Scanner,StoreFileScanner,MemStoreScanner和StoreScanner都实现了该接口。这里的comparator类型为KVScannerComparator,用于比较两个KeyValueScanner,实际上内部使用了KVComparator,它是用来比较两个KeyValue的。从后面可以看出,实际上,这个由KeyValueScanner组成的堆,堆顶KeyValueScanner满足的特征是: 它的堆顶(KeyValue)最小
堆用类KeyValueHeap表示,看KeyValueHeap构造函数做了什么
KeyValueHeap(List<? extends KeyValueScanner> scanners,
KVScannerComparator comparator) throws IOException {
this.comparator = comparator;
if (!scanners.isEmpty()) {
// 根据传入的KeyValueScanner构造出一个优先级队列(内部实现就是堆)
this.heap = new PriorityQueue<KeyValueScanner>(scanners.size(),
this.comparator);
for (KeyValueScanner scanner : scanners) {
if (scanner.peek() != null) {
this.heap.add(scanner);
} else {
scanner.close();
}
}
//以上将元素加入堆中
// 从堆顶pop出一个KeyValueScanner放入成员变量current,那么这个堆的堆顶
// 就是current这个KeyValueScanner的堆顶,KeyValueHeap的peek()取堆顶
// 操作直接返回current.peek()
this.current = pollRealKV();
}
}
在看pollRealKV()怎么做的之前需要先看看HBase 0.94引入的Lazy Seek
Lazy Seek优化
在这个优化之前,读取一个column family(Store),需要seek其下的所有HFile和MemStore到指定的查询KeyValue(seek的语义为如果KeyValue存在则seek到对应位置,如果不存在,则seek到这个KeyValue的后一个KeyValue,假设Store下有3个HFile和一个MemStore,按照时序递增记为[HFile1, HFile2, HFile3, MemStore],在lazy seek优化之前,需要对所有的HFile和MemStore进行seek,对HFile文件的seek比较慢,往往需要将HFile相应的block加载到内存,然后定位。在有了lazy seek优化之后,如果需要的KeyValue在HFile3中就存在,那么HFIle1和HFile2都不需要进行seek,大大提高速度。大体来说,思路是请求seek某个KeyValue时实际上没有对StoreFileScanner进行真正的seek,而是对于每个StoreFileScanner,设置它的peek为(rowkey,family,qualifier,lastTimestampInStoreFile)
KeyValueHeap有两个重要的接口,peek()和next(),他们都是返回堆顶,区别在于next()会将堆顶出堆,然后重新调整堆,对外来说就是迭代器向前移动,而peek()不会将堆顶出堆,堆顶不变。实现中,
peek()操作非常简单,只需要调用堆的成员变量current的peek()方法操作即可.拿StoreScanner堆举例,current要么是StoreFileScanner类型要么是MemStore,那么到底current是如何选择出来的以及Lazy Seek是如何实现的?
下面举个例子说明。
前提:
HBase开启了Lazy Seek优化(实际上默认开启)
假设:
Store下有三个HFile和MemStore,按照时间顺序记作[HFile1,HFile2,HFile3,MemStore],seek KeyValue为(rowkey,family,qualifier,timestamp),记作seekKV.
并且它只在HFile3中存在,不在其他HFile和MemStore中存在
Lazy Seek过程
seekScanner()的逻辑,如果是lazy seek,则对于每个Scanner都调
用requestSeek(seekKV)方法,方法内部首先进行rowcol类型的bloom filter过滤
- 如果结果判定seekKV在StoreFile中肯定不存在,则直接设置StoreFileScanner的peek(实际上StoreFileScanner不是一个
堆只是为了统一代码)为 kv.createLastOnRowCol(),并且将realSeekDone设置true,表示实际的seek完成.
public KeyValue createLastOnRowCol() {
return new KeyValue(
bytes, getRowOffset(), getRowLength(),
bytes, getFamilyOffset(), getFamilyLength(),
bytes, getQualifierOffset(), getQualifierLength(),
HConstants.OLDEST_TIMESTAMP, Type.Minimum, null, 0, 0);
}
可以看出ts设置为最小,说明这个KeyValue排在所有的同rowkey同column family同qualifier的KeyValue最后。显然,当上层StoreScanner取堆顶时,
如果其它StoreFileScanner/MemStoreScanner中存在同rowkey同column family同qualifier的真实的KeyValue则会优先弹出。
- 如果seekKV在StoreFile中,那么会执行如下逻辑:
realSeekDone = false;
long maxTimestampInFile = reader.getMaxTimestamp();
long seekTimestamp = kv.getTimestamp();
if (seekTimestamp > maxTimestampInFile) {
// Create a fake key that is not greater than the real next key.
// (Lower timestamps correspond to higher KVs.)
// To understand this better, consider that we are asked to seek
// to
// a higher timestamp than the max timestamp in this file. We
// know that
// the next point when we have to consider this file again is
// when we
// pass the max timestamp of this file (with the same
// row/column).
cur = kv.createFirstOnRowColTS(maxTimestampInFile);
} else {
enforceSeek();
}
显然,当kv的ts比HFile中最大的ts都更大时,那么这个HFile中显然不存在seekKV,但是可能存在
相同rowkey,family,qualifier的不同ts的KeyValue,那么这里设置堆顶时要注意,不能把堆顶设置为比当前HFile文件中的可能真实存在的相同rowkey,family,qualifier的KeyValue大,如下:
public KeyValue createFirstOnRowColTS(long ts) {
return new KeyValue(
bytes, getRowOffset(), getRowLength(),
bytes, getFamilyOffset(), getFamilyLength(),
bytes, getQualifierOffset(), getQualifierLength(),
ts, Type.Maximum, bytes, getValueOffset(), getValueLength());
}
Type的比较中,Type.Maximum最小,这样产生的KeyValue保证了不会大于当前HFile文件中的可能存在的相同rowkey,family,qualifier的KeyValue,同时将seekKV保存到StoreFileScanner成员变量delayedSeekKV中,以便后续真正seek的时候获取.
考虑一下如果seekKV的ts比当前HFile中的maxTimestamp更小怎么办?可以设置一个ts为latest_timestamp
的KeyValue么?如果设置了,它会比其它HFile中存在实际的KeyValue先弹出,这样顺序就乱了,所以这种情况下,只能进行实际的seek,enforceSeek()函数中进行实际的seek后,将realSeekDone设置为
true.
取StoreScanner堆顶逻辑
因为HFile3的latestTimestampInStoreFile最大,所以会首先取到HFile3对应的StoreFileScanner的pee
k(KeyValue的比较原则是timestamp大的KeyValue更小),
这个时候会检查这个KeyValueScanner是否进行了实际的seek(对于StoreFileScanner来说,通过布尔变量realSeekDone进行标记,对于MemStoreScanner来说,始终返回true),在这里,没有进行real seek
,接着进行实际的seek操作,seek到HFile3中存在的seekKV,接着拿着seekKV去和HFile2的peek进行比较,显然seekKV比HFile2的peek小(由于timestamp > lastTimestampInStoreFile2),故
StoreScanner的peek操作返回seekKV。
实现中,KeyValueHeap有两个重要的接口,peek()和next(),他们都是返回堆顶,区别在于next()会将堆顶出堆,然后重新调整堆,对外来说就是迭代器向前移动,而peek()不会将堆顶出堆,堆顶不变。实现中,
peek()操作非常简单,只需要调用堆的成员变量current的peek()方法操作即可.拿StoreScanner堆举例,current要么是StoreFileScanner类型要么是MemStore,而current的选择则是pollRealKV()
完成的,这个函数之所以内部有while循环就是因为考虑了Lazy Seek优化,实际上,pollRealKV()代码的逻辑就是例子中"取StoreScanner堆顶逻辑"。pollRealKV()的返回值会赋给current
protected KeyValueScanner pollRealKV() throws IOException {
KeyValueScanner kvScanner = heap.poll();
if (kvScanner == null) {
return null;
}
while (kvScanner != null && !kvScanner.realSeekDone()) {
if (kvScanner.peek() != null) {
kvScanner.enforceSeek();
KeyValue curKV = kvScanner.peek();
if (curKV != null) {
KeyValueScanner nextEarliestScanner = heap.peek();
if (nextEarliestScanner == null) {
// The heap is empty. Return the only possible scanner.
return kvScanner;
}
// Compare the current scanner to the next scanner. We try to avoid
// putting the current one back into the heap if possible.
KeyValue nextKV = nextEarliestScanner.peek();
if (nextKV == null || comparator.compare(curKV, nextKV) < 0) {
// We already have the scanner with the earliest KV, so return it.
return kvScanner;
}
// Otherwise, put the scanner back into the heap and let it compete
// against all other scanners (both those that have done a "real
// seek" and a "lazy seek").
heap.add(kvScanner);
} else {
// Close the scanner because we did a real seek and found out there
// are no more KVs.
kvScanner.close();
}
} else {
// Close the scanner because it has already run out of KVs even before
// we had to do a real seek on it.
kvScanner.close();
}
kvScanner = heap.poll();
}
return kvScanner;
}
Store下HFile集合发生变化如何处理
内存中的Memstore被flush到文件系统或者compaction完成都会改变Store的HFile文件集合。
在每次做完一批mutate操作后,会通过HRegion::isFlushSize(newSize)检查是否需要对当前HRegion内的memstore进行flush
其实就是判断HRegion内的所有的memstore大小和是否大于hbase.hregion.memstore.flush.size,默认128MB,如果需要flush,会将请求放入后台flush线程(MemStoreFlusher)的队列中,由后台flush线程处理,调用路径HRegion::flushcache()->internalFlushcache(...)->StoreFlushContext.flushCache(...)->StoreFlushContext.commit(...)=>HStore::updateStorefiles(),这块逻辑在HBase Snapshot原理和实现中有讲到,这里不赘述。只说一下最后一步的updateStorefiles()操作,该函数主要工作是拿住HStore级别的写锁,然后将新产生的HFile文件插入到StoreEngine中,解写锁,然后释放snapshot,最后调用
notifyChangedReadersObservers(),如下:
this.lock.writeLock().lock();
try {
this.storeEngine.getStoreFileManager().insertNewFiles(sfs);
this.memstore.clearSnapshot(set);
} finally {
// We need the lock, as long as we are updating the storeFiles
// or changing the memstore. Let us release it before calling
// notifyChangeReadersObservers. See HBASE-4485 for a possible
// deadlock scenario that could have happened if continue to hold
// the lock.
this.lock.writeLock().unlock();
}
// Tell listeners of the change in readers.
notifyChangedReadersObservers();
重点在于notifyChangedReadersObservers(),看看代码:
private void notifyChangedReadersObservers() throws IOException {
for (ChangedReadersObserver o: this.changedReaderObservers) {
o.updateReaders();
}
}
实际上,每个observer类型都是StoreScanner,每次新开一个StoreScanner都会注册在Store内部的这个observer集合中,当Store下面的HFile集合变化时,通知这些注册上来的StoreScanner即可。
具体的通知方式就是首先拿住StoreScanner的锁,将这个时候的堆顶保存在成员变量lastTop中,
然后将StoreScanner内部的堆置为null(this.heap=null)最后解锁,而StoreScanner那边next/seek/reseek时,都会首先通过函数checkReseek()函数来检查是否this.heap为null,为null
,为null说明当前Store下的HFile集合改变了,那么调用resetScannerStack(lastTop),将当前
Store下的所有StoreFileScanner/MemStoreScanner都seek到lastTop,然后重新建StoreScanner对应的堆。checkReseek()代码如下:
protected boolean checkReseek() throws IOException {
if (this.heap == null && this.lastTop != null) {
resetScannerStack(this.lastTop);
if (this.heap.peek() == null
|| store.getComparator().compareRows(this.lastTop, this.heap.peek()) != 0) {
LOG.debug("Storescanner.peek() is changed where before = " + this.lastTop.toString()
+ ",and after = " + this.heap.peek());
this.lastTop = null;
return true;
}
this.lastTop = null; // gone!
}
// else dont need to reseek
return false;
}