kafka本地存储4-LogCleaner

LogCleanerManager

管理每个TopicAndPartition的清理状态

inProgress = mutable.HashMap[TopicAndPartition, LogCleaningState]()

清理状态如下三种

LogCleaningInProgress

LogCleaningAborted

LogCleaningPaused

会有多个数据根目录存储log信息

数据根目录目录下有多个{Topic}-{Partition}目录下记录的是这个log的信息

每个数据根目录下,有一个文件名为 cleaner-offset-checkpoint的文件

记录每个TopicAndPartition清理的offset偏移量

文件格式

version

expectedSize topicandpartition的个数

topic partition offset

topic partition offset

topic partition offset

grabFilthiestLog方法

得到当前需要清理的log,该方法返回的对象类型为 LogToClean(topicPartition: TopicAndPartition, log: Log, firstDirtyOffset: Long)

遍历LogCleanerManager.logs,这个LogCleanerManager.logs是所有数据根目录下读取到的topicAndPartition的log信息

1.Log. config. compact 为true

2.TopicAndPartition的这个log不包含在inProgress中

3. LogToClean.totalBytes > 0

创建TopicAndPartition对应的LogToClean

TopicAndPartition的LogToClean.firstDirtyOffset参数有可能存储在cleaner-offset-checkpoint文件中,如果不在文件中,就使用 log.logSegments.head.baseOffset来设置给LogToClean.firstDirtyOffset,创建LogToClean

4. LogToClean.cleanableRatio > log.config.minCleanableRatio

返回的是LogToClean集合中 cleanableRatio最大的一个

LogToClean

LogToClean(topicPartition: TopicAndPartition, log: Log, firstDirtyOffset: Long)

cleanBytes, log中各segment记录的消息偏移量在-1到firstDirtyOffset之前的段的字节总和

dirtyBytes,log中各segment记录的消息偏移量在firstDirtyOffset 到log.activeSegment.baseOffset的段的字节总和

cleanableRatio, dirtyBytes / totalBytes

totalBytes, cleanBytes + dirtyBytes

LogCleaner

LogCleaner(val config: CleanerConfig,
                 val logDirs: Array[File],
                 val logs: Pool[TopicAndPartition, Log],

                 time: Time = SystemTime)

创建 config.numThreads个线程,线程对象为 CleanerThread(threadId: Int)

CleanerThread

初始化时创建Cleaner

new Cleaner(id = threadId,

                              offsetMap = new SkimpyOffsetMap(memory =

math.min(config.dedupeBufferSize/config.numThreads, Int.MaxValue).toInt,

                             hashAlgorithm = config.hashAlgorithm),

                              ioBufferSize = config.ioBufferSize / config.numThreads / 2,
                              maxIoBufferSize = config.maxMessageSize,
                              dupBufferLoadFactor = config.dedupeBufferLoadFactor,
                              throttler = throttler,
                              time = time,

                              checkDone = checkDone)

线程执行时

调用 cleanerManager.grabFilthiestLog()返回的最该清理的topicAndPartition的LogToClean对象

如果该LogToClean对象唯恐,表示现在暂时没有需要符合清理条件的LogToClean,就调用 backOffWaitLatch.await(config.backOffMs, TimeUnit.MILLISECONDS)

LogToClean不为空,调用 endOffset = cleaner.clean(cleanable),

endOffset为

    //1.得到这个log中offset从cleanable.firstDirtyOffset到log.activeSegment.baseOffset的segment的列表
    //2.计算minStopOffset = (cleanable.firstDirtyOffset + map.slots * this.dupBufferLoadFactor).toLong
    //this.dupBufferLoadFactor取值为config.dedupeBufferLoadFactor
    //3.开始逐个把segment的里的消息放进到OffsetMap中
    //直到segment.baseOffset <= minStopOffset || map.utilization < this.dupBufferLoadFactor停止

    //endOffset是LogSegment的最后一个消息放进OffsetMap时的offset

如果是LogCleaningInProgress

更新topicAndPartition需要清理的偏移值offset

把topicAndPartition和offset写入到参数目录的cleaner-offset-checkpoint文件中

如果是LogCleaningPaused状态,就给pausedCleaningCond发信号

cleanerManager.doneCleaning(cleanable.topicPartition, cleanable.log.dir.getParentFile, endOffset)

Cleaner

class Cleaner(val id: Int,
                           val offsetMap: OffsetMap,
                           ioBufferSize: Int,//config.ioBufferSize / config.numThreads / 2
                           maxIoBufferSize: Int,//config.maxMessageSize
                           dupBufferLoadFactor: Double,//config.dedupeBufferLoadFactor
                           throttler: Throttler,
                           time: Time,

                           checkDone: (TopicAndPartition) => Unit) 

Cleaner.readBuffer

Cleaner.writeBuffer

这两个buff大小为config.ioBufferSize / config.numThreads / 2

在做某个topicAndPartition清理时,需要从老segmernt中读到Cleaner.readBuffer,之后在把符合的message写入心segment时,要先把数据写到Cleaner.writeBuffer中

cleaner.clean(cleanable)函数作用

1.buildOffsetMap

清理消息的起始位置是0,结束位置为endOffset

buildOffsetMap函数返回的偏移量记作endOffset,由如下两个因素决定

这个偏移量不会超过( cleanable.firstDirtyOffset + map.slots * this.dupBufferLoadFactor)

在从segment往offsetMap写message.key, entry.offset时,写到map.utilization < this.dupBufferLoadFactor位置的offset

1).得到需要清理的segment集合,取出 cleanable.firstDirtyOffset到 log.activeSegment.baseOffset的所有segment,记作 dirty

2).通过offsetMap参数的大小,来计算一次清理的结束的offset,记作minStopOffset

minStopOffset = (start + map.slots * this.dupBufferLoadFactor).toLong

3).遍历dirty

满足两个条件其中之一, segment.baseOffset <= minStopOffset || map.utilization < this.dupBufferLoadFactor

就开始调用buildOffsetMapForSegment来把该segment信息保存在offsetMap中

segment把消息读到Cleaner.readBuffer中,之后利用Cleaner.readBuffer创建 ByteBufferMessageSet

entry类型为MessageAndOffset(message: Message, offset: Long)

offsetMap保存的内容是 map.put(message.key, entry.offset)

2.得到需要删除的时间戳,比这个时间戳小的,就直接删除,不计入归并计算,记作 deleteHorizonMs

1)把offset从0到cleanable.firstDirtyOffset的segment集合

2)取出该集合最后一个segment,这个segment是离当前时间最近的segment,

deleteHorizonMs = seg.lastModified - log.config.deleteRetentionMs

3.把offset从0到endOffset,进行分组,每组segment字节大小不超过log.config.segmentSize,每组index大小不能超过log.config.maxIndexSize

groupSegmentsBySize(log.logSegments(0, endOffset), log.config.segmentSize, log.config.maxIndexSize)

如果写消息不频繁,就会根据时间间隔产生过多的段,通过这样分组合并,可以减少物理文件的数量

4.遍历每个分组

cleanSegments(log: Log,

     segments: Seq[LogSegment],

     map: OffsetMap,

     deleteHorizonMs: Long) 

1)遍历分组中的每个段segment

retainDeletes = old.lastModified > deleteHorizonMs retainDeletes为true,表示保留需要这个消息

2)cleanInto(topicAndPartition: TopicAndPartition, source: LogSegment,

                             dest: LogSegment, map: OffsetMap, retainDeletes: Boolean) 

开始把分组中的段写入到新段segment中

遍历每个段的每个消息

满足下面两个条件的写入到新段中

**这个消息在OffsetMap存在,并且offset和OffsetMap里的foundOffset一致

**消息为不空,并且需要保留

利用老的segment的 segments.last.lastModified来设置新段 cleaned.lastModified = modified

3)保留参数newSegment的LogSegment

删除掉参数oldSegments列表中baseOffset和参数newSegment.baseOffset不想等的seg

先newSegment文件名.cleaned改成swap,完成删除后再恢复成.cleaned