目录
SourceReader 概念
SourceReader 源码方法
void start();
InputStatus pollNext(ReaderOutput output) throws Exception;
List snapshotState(long checkpointId);
CompletableFuture isAvailable();
void addSplits(List splits);
参考
SourceReader是一个运行在Task Manager上的组件,主要是负责读取 SplitEnumerator 分配的source split。
SourceReader 提供了一个拉动式(pull-based)处理接口。Flink任务会在循环中不断调用 pollNext(ReaderOutput) 轮询来自 SourceReader 的记录。 pollNext(ReaderOutput) 方法的返回值指示 SourceReader 的状态。
pollNext(ReaderOutput) 会使用 ReaderOutput 作为参数,为了提高性能且在必要情况下, SourceReader 可以在一次 pollNext() 调用中返回多条记录。例如:有时外部系统的工作系统的工作粒度为块。而一个块可以包含多个记录,但是 source 只能在块的边界处设置 Checkpoint。在这种情况下, SourceReader 可以一次将一个块中的所有记录通过 ReaderOutput 发送至下游。
然而,除非有必要,SourceReader 的实现应该避免在一次 pollNext(ReaderOutput) 的调用中发送多个记录。这是因为对 SourceReader 轮询的任务线程工作在一个事件循环(event-loop)中,且不能阻塞。
在创建 SourceReader 时,相应的 SourceReaderContext 会提供给 Source,而 Source 则会将对应的上下文传递给 SourceReader 实例。 SourceReader 可以通过 SourceReaderContext 将 SourceEvent 传递给相应的 SplitEnumerator 。 Source 的一个典型设计模式是让 SourceReader 发送它们的本地信息给 SplitEnumerator,后者则会全局性地做出决定。
SourceReader API 是一个底层(low-level)API,允许用户自行处理分片,并使用自己的线程模型来获取和移交记录。为了帮助实现 SourceReader,Flink 提供了 SourceReaderBase 类,可以显著减少编写 SourceReader 所需要的工作量。
强烈建议连接器开发人员充分利用 SourceReaderBase 而不是从头开始编写 SourceReader。
这里简单说一下,如何通过 Source 创建 DataStream ,有两种方法(感觉上没啥区别):
// fromSource 这个返回的是source
final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
Source mySource = new MySource(....);
DataStream stream = env.fromSource(
mySource,
WatermarkStrategy.noWatermarks(),// 无水标
"MySourceName");
..
// addSource 这个返回的是Source function
final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
DataStream<..> stream = env.addSource(new MySource(...));
判断是否有splits了,如果当前没有已经分配的splits了就发送请求获取。
/** Start the reader. */
void start();
// FileSourceReader的实现
@Override
public void start() {
// we request a split only if we did not get splits during the checkpoint restore
if (getNumberOfCurrentlyAssignedSplits() == 0) {
context.sendSplitRequest(); // 发送split的读取请求给SplitEnumerator,在handleSplitRequest方法中被调用
}
}
主要负责拉取下一个可读取的记录到SourceOutput,确保这个方法是非阻塞的,并且最好一次调用只输出一条数据。
/**
* Poll the next available record into the {@link SourceOutput}.
*
* The implementation must make sure this method is non-blocking.
*
*
Although the implementation can emit multiple records into the given SourceOutput, it is
* recommended not doing so. Instead, emit one record into the SourceOutput and return a {@link
* InputStatus#MORE_AVAILABLE} to let the caller thread know there are more records available.
*
* @return The InputStatus of the SourceReader after the method invocation.
*/
InputStatus pollNext(ReaderOutput output) throws Exception;
// FileSourceReader读取数据的pollNext方法位于父类SourceReaderBase中
@Override
public InputStatus pollNext(ReaderOutput output) throws Exception {
// make sure we have a fetch we are working on, or move to the next
// 获取当前从fetcher中读取到的一批split
// RecordsWithSplitIds代表了从fetcher拉取到SourceReader的数据
// RecordsWithSplitIds可以包含多个split,但是对于FileRecords而言,只代表一个split
RecordsWithSplitIds recordsWithSplitId = this.currentFetch;
if (recordsWithSplitId == null) {
// 如果没有,获取下一批split
recordsWithSplitId = getNextFetch(output);
if (recordsWithSplitId == null) {
// 如果还没有获取到,需要检查后续是否还会有数据到来。
return trace(finishedOrAvailableLater());
}
}
// we need to loop here, because we may have to go across splits
while (true) {
// Process one record.
// 从split中获取下一条记录
final E record = recordsWithSplitId.nextRecordFromSplit();
if (record != null) {
// emit the record.
// 如果获取到数据
// 记录数量计数器加1
numRecordsInCounter.inc(1);
// 发送数据到Output
// currentSplitOutput为当前split对应的下游output
// currentSplitContext.state为reader的读取状态
recordEmitter.emitRecord(record, currentSplitOutput, currentSplitContext.state);
LOG.trace("Emitted record: {}", record);
// We always emit MORE_AVAILABLE here, even though we do not strictly know whether
// more is available. If nothing more is available, the next invocation will find
// this out and return the correct status.
// That means we emit the occasional 'false positive' for availability, but this
// saves us doing checks for every record. Ultimately, this is cheaper.
// 总是发送MORE_AVAILABLE
// 如果真的没有可用数据,下次调用会返回正确的状态
return trace(InputStatus.MORE_AVAILABLE);
} else if (!moveToNextSplit(recordsWithSplitId, output)) {
// 如果本次fetch的split已经全部被读取(本批没有更多的split),读取下一批数据
// The fetch is done and we just discovered that and have not emitted anything, yet.
// We need to move to the next fetch. As a shortcut, we call pollNext() here again,
// rather than emitting nothing and waiting for the caller to call us again.
return pollNext(output);
}
// else fall through the loop
}
}
getNextFetch方法获取下一批 split 。
@Nullable
private RecordsWithSplitIds getNextFetch(final ReaderOutput output) {
// 检查fetcher是否有错误
splitFetcherManager.checkErrors();
LOG.trace("Getting next source data batch from queue");
// elementsQueue中缓存了fetcher线程获取的split
// 从这个队列中拿出一批split
final RecordsWithSplitIds recordsWithSplitId = elementsQueue.poll();
// 如果队列中没有数据,并且接下来这批split已被读取完毕,返回null
if (recordsWithSplitId == null || !moveToNextSplit(recordsWithSplitId, output)) {
// No element available, set to available later if needed.
return null;
}
// 更新当前的fetch
currentFetch = recordsWithSplitId;
return recordsWithSplitId;
}
finishedOrAvailableLater 方法检查后续是否还有数据,返回对应的状态。
private InputStatus finishedOrAvailableLater() {
// 检查所有的fetcher是否都已关闭
final boolean allFetchersHaveShutdown = splitFetcherManager.maybeShutdownFinishedFetchers();
// 如果reader不会再接收更多的split,或者所有的fetcher都已关闭
// 返回NOTHING_AVAILABLE,将来可能会有记录可用。
if (!(noMoreSplitsAssignment && allFetchersHaveShutdown)) {
return InputStatus.NOTHING_AVAILABLE;
}
if (elementsQueue.isEmpty()) {
// 如果缓存队列中没有数据,返回END_OF_INPUT
// We may reach here because of exceptional split fetcher, check it.
splitFetcherManager.checkErrors();
return InputStatus.END_OF_INPUT;
} else {
// We can reach this case if we just processed all data from the queue and finished a
// split,
// and concurrently the fetcher finished another split, whose data is then in the queue.
// 其他情况返回MORE_AVAILABLE
return InputStatus.MORE_AVAILABLE;
}
}
moveToNextSplit
方法前往读取下一个split。
private boolean moveToNextSplit(
RecordsWithSplitIds recordsWithSplitIds, ReaderOutput output) {
// 获取下一个split的ID
final String nextSplitId = recordsWithSplitIds.nextSplit();
if (nextSplitId == null) {
// 如果没获取到,则当前获取过程结束
LOG.trace("Current fetch is finished.");
finishCurrentFetch(recordsWithSplitIds, output);
return false;
}
// 获取当前split上下文
// Map> splitStates它保存了split ID和split的状态
currentSplitContext = splitStates.get(nextSplitId);
checkState(currentSplitContext != null, "Have records for a split that was not registered");
// 获取当前split对应的output
// SourceOperator在从SourceCoordinator获取到分片后会为每个分片创建一个OUtput,currentSplitOutput是当前分片的输出
currentSplitOutput = currentSplitContext.getOrCreateSplitOutput(output);
LOG.trace("Emitting records from fetch for split {}", nextSplitId);
return true;
}
主要是负责创建 source 的 checkpoint 。
/**
* Checkpoint on the state of the source.
*
* @return the state of the source.
*/
List snapshotState(long checkpointId);
public List snapshotState(long checkpointId) {
List splits = new ArrayList();
this.splitStates.forEach((id, context) -> {
splits.add(this.toSplitType(id, context.state));
});
return splits;
}
/**
* Returns a future that signals that data is available from the reader.
*
* Once the future completes, the runtime will keep calling the {@link
* #pollNext(ReaderOutput)} method until that methods returns a status other than {@link
* InputStatus#MORE_AVAILABLE}. After that the, the runtime will again call this method to
* obtain the next future. Once that completes, it will again call {@link
* #pollNext(ReaderOutput)} and so on.
*
*
The contract is the following: If the reader has data available, then all futures
* previously returned by this method must eventually complete. Otherwise the source might stall
* indefinitely.
*
*
It is not a problem to have occasional "false positives", meaning to complete a future
* even if no data is available. However, one should not use an "always complete" future in
* cases no data is available, because that will result in busy waiting loops calling {@code
* pollNext(...)} even though no data is available.
*
* @return a future that will be completed once there is a record available to poll.
*/
// 创建一个future,表明reader中是否有数据可被读取
// 一旦这个future进入completed状态,Flink一直调用pollNext(ReaderOutput)方法直到这个方法返回除InputStatus#MORE_AVAILABLE之外的内容
// 在这之后,会再次调isAvailable方法获取下一个future。如果它completed,再次调用pollNext(ReaderOutput)。以此类推
public CompletableFuture isAvailable() {
return this.currentFetch != null ? FutureCompletingBlockingQueue.AVAILABLE : this.elementsQueue.getAvailabilityFuture();
}
/**
* Adds a list of splits for this reader to read. This method is called when the enumerator
* assigns a split via {@link SplitEnumeratorContext#assignSplit(SourceSplit, int)} or {@link
* SplitEnumeratorContext#assignSplits(SplitsAssignment)}.
*
* @param splits The splits assigned by the split enumerator.
*/
// 添加一系列splits,以供reader读取。这个方法在SplitEnumeratorContext#assignSplit(SourceSplit, int)或者SplitEnumeratorContext#assignSplits(SplitsAssignment)中调用
void addSplits(List splits);
其中,SourceReaderBase类的实现,fetcher的作用是从拉取split缓存到SourceReader中。
@Override
public void addSplits(List splits) {
LOG.info("Adding split(s) to reader: {}", splits);
// Initialize the state for each split.
splits.forEach(
s ->
splitStates.put(
s.splitId(), new SplitContext<>(s.splitId(), initializedState(s))));
// Hand over the splits to the split fetcher to start fetch.
splitFetcherManager.addSplits(splits);
}
addSplits
方法将fetch任务交给 SplitFetcherManager
处理,它的 addSplits
方法如下:
@Override
public void addSplits(List splitsToAdd) {
// 获取正在运行的fetcher
SplitFetcher fetcher = getRunningFetcher();
if (fetcher == null) {
// 如果没有,创建出一个fetcher
fetcher = createSplitFetcher();
// Add the splits to the fetchers.
// 将这个创建出的fetcher加入到running fetcher集合中
fetcher.addSplits(splitsToAdd);
// 启动这个fetcher
startFetcher(fetcher);
} else {
// 如果获取到了正在运行的fetcher,调用它的addSplits方法
fetcher.addSplits(splitsToAdd);
}
}
最后我们查看SplitFetcher
的addSplits
方法:
public void addSplits(List splitsToAdd) {
// 将任务包装成AddSplitTask,通过splitReader兼容不同格式数据的读取方式
// 将封装好的任务加入到队列中
enqueueTask(new AddSplitsTask<>(splitReader, splitsToAdd, assignedSplits));
// 唤醒fetcher任务,使用SplitReader读取数据
// Split读取数据并缓存到elementQueue的逻辑位于FetcherTask,不再具体分析
wakeUp(true);
}
数据源 | Apache Flink
Flink 源码之新 Source 架构 - 简书
Flink新Source架构(下) - 知乎