Major compaction时的scan操作

Major compaction时的scan操作

 

发起major compaction时，通过CompactSplitThread.CompactionRunner.run开始执行

 

-->region.compact(compaction, store)-->store.compact(compaction)-->

 

CompactionContext.compact,发起compact操作

 

CompactionContext的实例通过HStore中的storeEngine.createCompaction()生成，

 

默认值为DefaultStoreEngine,通过hbase.hstore.engine.class配置。

 

默认的CompactionContext实例为DefaultCompactionContext。

 

而DefaultCompactionContext.compact方法最终调用DefaultStoreEngine.compactor来执行

 

compactor的实现通过hbase.hstore.defaultengine.compactor.class配置，默认实现为DefaultCompactor

 

调用DefaultCompactor.compact(request);

 

 

 

1.根据要进行compact的storefile文件,生成对应的StoreFileScanner集合列表。

 

在生成StoreFileScanner实例时，每一个scanner中的ScanQueryMatcher为null

 

 

 

2.创建StoreScanner实例，设置ScanType为ScanType.COMPACT_DROP_DELETES。

 

 

 

privateStoreScanner(Storestore, ScanInfo scanInfo, Scan scan,

 

List<? extendsKeyValueScanner> scanners, ScanTypescanType, longsmallestReadPoint,

 

longearliestPutTs, byte[] dropDeletesFromRow, byte[] dropDeletesToRow) throws IOException {

 

this(store, false, scan, null, scanInfo.getTtl(),

 

scanInfo.getMinVersions());

 

if (dropDeletesFromRow == null) {

 

执行这里,传入的columns为null

 

matcher = newScanQueryMatcher(scan, scanInfo, null, scanType,

 

smallestReadPoint, earliestPutTs, oldestUnexpiredTS);

 

} else {

 

matcher = newScanQueryMatcher(scan, scanInfo, null, smallestReadPoint,

 

earliestPutTs, oldestUnexpiredTS, dropDeletesFromRow, dropDeletesToRow);

 

}

 

 

 

ScanqueryMatcher的构造方法：

 

传入的columns为null

 

publicScanQueryMatcher(Scan scan, ScanInfo scanInfo,

 

NavigableSet<byte[]> columns, ScanTypescanType,

 

longreadPointToUse, longearliestPutTs, longoldestUnexpiredTS) {

 

tr中mintime=0,maxtime=long.maxvalue

 

this.tr = scan.getTimeRange();

 

this.rowComparator = scanInfo.getComparator();

 

此deletes属性中的kv delete信息为到一个新的row时，都会重新进行清空。

 

this.deletes = newScanDeleteTracker();

 

this.stopRow = scan.getStopRow();

 

this.startKey = KeyValue.createFirstDeleteFamilyOnRow(scan.getStartRow(),

 

scanInfo.getFamily());

 

得到filter实例

 

this.filter = scan.getFilter();

 

this.earliestPutTs = earliestPutTs;

 

this.maxReadPointToTrackVersions = readPointToUse;

 

this.timeToPurgeDeletes = scanInfo.getTimeToPurgeDeletes();

 

此处为的值为false

 

/* how to deal with deletes */

 

this.isUserScan = scanType == ScanType.USER_SCAN;

 

此处的值为false,scanInfo.getKeepDeletedCells()的值默认false,

 

可通过table的columnfmaily中配置KEEP_DELETED_CELLS属性

 

scan.isRaw()可通过在scan中setAttribute的_raw_属性，默认为false

 

// keep deleted cells: if compaction or raw scan

 

this.keepDeletedCells = (scanInfo.getKeepDeletedCells() && !isUserScan) || scan.isRaw();

 

此处的值为false,此时是major的compact，不保留delete的数据

 

scan.isRaw()可通过在scan中setAttribute的_raw_属性，默认为false

 

// retain deletes: if minor compaction or raw scan

 

this.retainDeletesInOutput = scanType == ScanType.COMPACT_RETAIN_DELETES || scan.isRaw();

 

此时的值为false

 

// seePastDeleteMarker: user initiated scans

 

this.seePastDeleteMarkers = scanInfo.getKeepDeletedCells() && isUserScan;

 

得到查询的最大版本数,此时的scan.maxversion与scanInfo.maxversion的值是相同的值

 

intmaxVersions =

 

scan.isRaw() ? scan.getMaxVersions() : Math.min(scan.getMaxVersions(),

 

scanInfo.getMaxVersions());

 

 

 

生成columns属性的值为ScanWildcardColumnTracker实例，设置hasNullColumn的值为true

 

// Single branch to deal with two types of reads (columns vs all in family)

 

if (columns == null || columns.size() == 0) {

 

// there is always a null column in the wildcard column query.

 

hasNullColumn = true;

 

columns属性中的index表示当前比对到的column的下标值，每比较一行时，此值会重新清空

 

// use a specialized scan for wildcard column tracker.

 

this.columns = newScanWildcardColumnTracker(

 

scanInfo.getMinVersions(), maxVersions, oldestUnexpiredTS);

 

} else {

 

这个部分在compact时是不会执行的

 

// whether there is null column in the explicit column query

 

hasNullColumn = (columns.first().length == 0);

 

 

 

// We can share the ExplicitColumnTracker, diff is we reset

 

// between rows, not between storefiles.

 

this.columns = newExplicitColumnTracker(columns,

 

scanInfo.getMinVersions(), maxVersions, oldestUnexpiredTS);

 

}

 

}

 

 

 

ScanQueryMatcher.match过滤kv是否包含的方法分析

 

在通过StoreScanner.next(kvlist,limit)得到下一行的kv集合时

 

调用ScanQueryMatcher.match来过滤数据的方法分析

 

其中match方法返回的值具体作用可参见StoreScanner中的如下方法：

 

public boolean next(List<Cell> outResult, int limit).....

 

 

 

public MatchCode match(KeyValue kv) throws IOException {

 

调用filter的filterAllRemaining方法，如果此方法返回true表示此次scan结束

 

if (filter != null && filter.filterAllRemaining()) {

 

returnMatchCode.DONE_SCAN;

 

}

 

得到kv的值

 

byte [] bytes = kv.getBuffer();

 

KV在bytes中的开始位置

 

intoffset = kv.getOffset();

 

得到key的长度

 

keyvalue的组成：

 

4	4	2	~	1	~	~	8	1	~
kenlen	vlen	rowlen	row	cflen	cf	column	timestamp	kvtype	value

 

 

 

intkeyLength = Bytes.toInt(bytes, offset, Bytes.SIZEOF_INT);

 

得到rowkey的长度记录的开始位置（不包含keylen与vlen）

 

offset += KeyValue.ROW_OFFSET;

 

rowkey的长度记录的开始位置

 

intinitialOffset = offset;

 

得到rowkey的长度

 

shortrowLength = Bytes.toShort(bytes, offset, Bytes.SIZEOF_SHORT);

 

得到rowkey的开始位置

 

offset += Bytes.SIZEOF_SHORT;

 

比较当前传入的kv的rowkey部分是否与当前行中第一个kv的rowkey部分相同。也就是是否是同一行的数据

 

intret = this.rowComparator.compareRows(row, this.rowOffset, this.rowLength,

 

bytes, offset, rowLength);

 

如果当前传入的kv中的rowkey大于当前行的kv的rowkey部分，表示现在传入的kv是下一行，

 

结束当前next操作，(不是结束scan，是结束当次的next，表示这个next的一行数据的所有kv都查找完了)

 

if (ret <= -1) {

 

returnMatchCode.DONE;

 

否则表示当前传入的kv是上一行的数据，需要把当前的scanner向下移动一行

 

} elseif (ret >= 1) {

 

// could optimize this, if necessary?

 

// Could also be called SEEK_TO_CURRENT_ROW, but this

 

// should be rare/never happens.

 

returnMatchCode.SEEK_NEXT_ROW;

 

}

 

优化配置，是否需要不执行下面流程，直接把当前的scanner向下移动一行

 

stickyNextRow的值为true的条件:

 

1.ColumnTracker.done返回为true,

 

2.ColumnTracker.checkColumn返回为SEEK_NEXT_ROW.

 

3.filter.filterKeyValue(kv);返回结果为NEXT_ROW。

 

4.ColumnTracker.checkVersions返回为INCLUDE_AND_SEEK_NEXT_ROW。

 

ColumnTracker的实现在scan的columns为null或者是compact scan时为ScanWildcardColumnTracker。

 

否则为ExplicitColumnTracker。

 

 

 

// optimize case.

 

if (this.stickyNextRow)

 

returnMatchCode.SEEK_NEXT_ROW;

 

在ScanWildcardColumnTracker实例中返回值为false,

 

在ExplicitColumnTracker实例中返回值是当前的kv是否大于或等于查找的column列表的总和

 

if (this.columns.done()) {

 

stickyNextRow = true;

 

returnMatchCode.SEEK_NEXT_ROW;

 

}

 

得到familylen的记录位置

 

//Passing rowLength

 

offset += rowLength;

 

得到family的长度

 

//Skipping family

 

bytefamilyLength = bytes [offset];

 

把位置移动到family的名称记录的位置

 

offset += familyLength + 1;

 

得到column的长度

 

intqualLength = keyLength -

 

(offset - initialOffset) - KeyValue.TIMESTAMP_TYPE_SIZE;

 

得到kv中timestamp的值

 

longtimestamp = Bytes.toLong(bytes, initialOffset + keyLength – KeyValue.TIMESTAMP_TYPE_SIZE);

 

检查timestamp是否在指定的范围内，主要检查ttl是否过期

 

// check for early out based on timestamp alone

 

if (columns.isDone(timestamp)) {

 

如果发现kv的ttl过期，在ScanWildcardColumnTracker实例中直接返回SEEK_NEXT_COL。这个在compact中是默认

 

在ExplicitColumnTracker实例中检查是否有下一个column如果有返回SEEK_NEXT_COL。否则返回SEEK_NEXT_ROW。

 

returncolumns.getNextRowOrNextColumn(bytes, offset, qualLength);

 

}

 

 

 

/*

 

* The delete logic is pretty complicated now.

 

* This is corroborated by the following:

 

* 1. The store might be instructed to keep deleted rows around.

 

* 2. A scan can optionally see past a delete marker now.

 

* 3. If deleted rows are kept, we have to find out when we can

 

* remove the delete markers.

 

* 4. Family delete markers are always first (regardless of their TS)

 

* 5. Delete markers should not be counted as version

 

* 6. Delete markers affect puts of the *same* TS

 

* 7. Delete marker need to be version counted together with puts

 

* they affect

 

*/

 

得到kv的类型。

 

bytetype = bytes[initialOffset + keyLength – 1];

 

如果kv是删除的kv

 

if (kv.isDelete()) {

 

在默认情况下，此keepDeletedCells值为false,这里的if检查会进去

 

if (!keepDeletedCells) {

 

// first ignore delete markers if the scanner can do so, and the

 

// range does not include the marker

 

//

 

// during flushes and compactions also ignore delete markers newer

 

// than the readpoint of any open scanner, this prevents deleted

 

// rows that could still be seen by a scanner from being collected

 

此时的值为true,scan中的tr默认为alltime=true

 

booleanincludeDeleteMarker = seePastDeleteMarkers ?

 

tr.withinTimeRange(timestamp) :

 

tr.withinOrAfterTimeRange(timestamp);

 

把删除的kv添加到DeleteTracker中。compact时的实现为ScanDeleteTracker。

 

if (includeDeleteMarker

 

&& kv.getMvccVersion() <= maxReadPointToTrackVersions) {

 

this.deletes.add(bytes, offset, qualLength, timestamp, type);

 

}

 

// Can't early out now, because DelFam come before any other keys

 

}

 

如果非minor compact时，

 

或者在compact的scan时，同时当前时间减去kv的timestamp还不到hbase.hstore.time.to.purge.deletes配置的时间，

 

默认的配置值为0,

 

或者kv的mvcc值大于现在最大的mvcc值时。此if会进行。目前在做major compact的scan,不进去

 

if (retainDeletesInOutput

 

|| (!isUserScan && (EnvironmentEdgeManager.currentTimeMillis() - timestamp) <= timeToPurgeDeletes)

 

|| kv.getMvccVersion() > maxReadPointToTrackVersions) {

 

// always include or it is not time yet to check whether it is OK

 

// to purge deltes or not

 

if (!isUserScan) {

 

// if this is not a user scan (compaction), we can filter this deletemarker right here

 

// otherwise (i.e. a "raw" scan) we fall through to normal version and timerange checking

 

returnMatchCode.INCLUDE;

 

}

 

以下的检查通常情况不会进入

 

} elseif (keepDeletedCells) {

 

if (timestamp < earliestPutTs) {

 

// keeping delete rows, but there are no puts older than

 

// this delete in the store files.

 

returncolumns.getNextRowOrNextColumn(bytes, offset, qualLength);

 

}

 

// else: fall through and do version counting on the

 

// delete markers

 

如果kv是已经delete的kv，添加到DeleteTracker后，直接返回SKIP.

 

} else {

 

returnMatchCode.SKIP;

 

}

 

// note the following next else if...

 

// delete marker are not subject to other delete markers

 

} elseif (!this.deletes.isEmpty()) {

 

如果不是删除的KV时，检查删除的kv中是否包含此kv的版本。

 

a.如果KV是DeleteFamily。同时当前的KV的TIMESTAMP的值小于删除的KV的TIMESTAMP的值，返回FAMILY_DELETED。

 

b.如果KV是DeleteFamilyVersion已经删除掉的版本(删除时指定了timestamp)。返回FAMILY_VERSION_DELETED。

 

c.如果KV的是DeleteColumn，同时deleteTracker中包含的kv中部分qualifier的值

 

与传入的kv中部分qualifier的值相同。同时delete中包含的kv是DeleteColumn返回COLUMN_DELETED。

 

否则deleteTracker中包含的kv中部分qualifier的值与传入的kv中部分qualifier的值相同。

 

同时传入的kv中的timestamp的值是delete中的timestamp值，表示删除指定的版本，返回VERSION_DELETED。

 

d.否则表示没有删除的情况，返回NOT_DELETED。

 

DeleteResultdeleteResult = deletes.isDeleted(bytes, offset, qualLength,

 

timestamp);

 

switch (deleteResult) {

 

caseFAMILY_DELETED:

 

caseCOLUMN_DELETED:

 

returncolumns.getNextRowOrNextColumn(bytes, offset, qualLength);

 

caseVERSION_DELETED:

 

caseFAMILY_VERSION_DELETED:

 

returnMatchCode.SKIP;

 

caseNOT_DELETED:

 

break;

 

default:

 

thrownewRuntimeException("UNEXPECTED");

 

}

 

}

 

检查当前传入的kv的timestamp是否在包含的时间范围内，默认的scan是所有时间都包含

 

inttimestampComparison = tr.compare(timestamp);

 

如果当前kv的时间超过了最大的时间,返回SKIP。

 

if (timestampComparison >= 1) {

 

returnMatchCode.SKIP;

 

} elseif (timestampComparison <= -1) {

 

如果当前kv的时间小于了最小的时间，返回SEEK_NEXT_COL或者SEEK_NEXT_ROW。

 

returncolumns.getNextRowOrNextColumn(bytes, offset, qualLength);

 

}

 

如果时间在正常的范围内,columns.checkColumn如果是compact时的scan 此方法返回INCLUDE。

 

其它情况请参见ExplicitColumnTracker的实现。

 

 

 

// STEP 1: Check if the column is part of the requested columns

 

MatchCodecolChecker = columns.checkColumn(bytes, offset, qualLength, type);

 

此处的IF检查会进入

 

if (colChecker == MatchCode.INCLUDE) {

 

执行filter操作，并根据filter的响应返回相关的值，此处不在说明，比较容易看明白。

 

ReturnCodefilterResponse = ReturnCode.SKIP;

 

// STEP 2: Yes, the column is part of the requested columns. Check if filter is present

 

if (filter != null) {

 

// STEP 3: Filter the key value and return if it filters out

 

filterResponse = filter.filterKeyValue(kv);

 

switch (filterResponse) {

 

caseSKIP:

 

returnMatchCode.SKIP;

 

caseNEXT_COL:

 

returncolumns.getNextRowOrNextColumn(bytes, offset, qualLength);

 

caseNEXT_ROW:

 

stickyNextRow = true;

 

returnMatchCode.SEEK_NEXT_ROW;

 

caseSEEK_NEXT_USING_HINT:

 

returnMatchCode.SEEK_NEXT_USING_HINT;

 

default:

 

//It means it is either include or include and seek next

 

break;

 

}

 

}

 

/*

 

* STEP 4: Reaching this step means the column is part of the requested columns and either

 

* the filter is null or the filter has returned INCLUDE or INCLUDE_AND_NEXT_COL response.

 

* Now check the number of versions needed. This method call returns SKIP, INCLUDE,

 

* INCLUDE_AND_SEEK_NEXT_ROW, INCLUDE_AND_SEEK_NEXT_COL.

 

*

 

* FilterResponse ColumnChecker Desired behavior

 

* INCLUDE SKIP row has already been included, SKIP.

 

* INCLUDE INCLUDE INCLUDE

 

* INCLUDE INCLUDE_AND_SEEK_NEXT_COL INCLUDE_AND_SEEK_NEXT_COL

 

* INCLUDE INCLUDE_AND_SEEK_NEXT_ROW INCLUDE_AND_SEEK_NEXT_ROW

 

* INCLUDE_AND_SEEK_NEXT_COL SKIP row has already been included, SKIP.

 

* INCLUDE_AND_SEEK_NEXT_COL INCLUDE INCLUDE_AND_SEEK_NEXT_COL

 

* INCLUDE_AND_SEEK_NEXT_COL INCLUDE_AND_SEEK_NEXT_COL INCLUDE_AND_SEEK_NEXT_COL

 

* INCLUDE_AND_SEEK_NEXT_COL INCLUDE_AND_SEEK_NEXT_ROW INCLUDE_AND_SEEK_NEXT_ROW

 

*

 

* In all the above scenarios, we return the column checker return value except for

 

* FilterResponse (INCLUDE_AND_SEEK_NEXT_COL) and ColumnChecker(INCLUDE)

 

*/

 

colChecker =

 

columns.checkVersions(bytes, offset, qualLength, timestamp, type,

 

kv.getMvccVersion() > maxReadPointToTrackVersions);

 

//Optimize with stickyNextRow

 

stickyNextRow = colChecker == MatchCode.INCLUDE_AND_SEEK_NEXT_ROW ? true : stickyNextRow;

 

return (filterResponse == ReturnCode.INCLUDE_AND_NEXT_COL &&

 

colChecker == MatchCode.INCLUDE) ? MatchCode.INCLUDE_AND_SEEK_NEXT_COL

 

: colChecker;

 

}

 

stickyNextRow = (colChecker == MatchCode.SEEK_NEXT_ROW) ? true

 

: stickyNextRow;

 

returncolChecker;

 

}

 

 

 

major与minor的compact写入新storefile时的区别

 

如果是major的compact的写入，会在close掉writer时，

 

在meta中写入major==true的值MAJOR_COMPACTION_KEY=true。

 

此值主要用来控制做minor的compact时是否选择这个storefile文件。

 

 

 

if (writer != null) {

 

writer.appendMetadata(fd.maxSeqId, request.isMajor());

 

writer.close();

 

newFiles.add(writer.getPath());

 

}