用于外部数据访问的异步 I/O

This page explains the use of Flink’s API for asynchronous I/O with external data stores.For users not familiar with asynchronous or event-driven programming, an article about Futures andevent-driven programming may be useful preparation.

Note: Details about the design and implementation of the asynchronous I/O utility can be found in the proposal and design documentFLIP-12: Asynchronous I/O Design and Implementation.

The need for Asynchronous I/O Operations

When interacting with external systems (for example when enriching stream events with data stored in a database), one needs to take carethat communication delay with the external system does not dominate the streaming application’s total work.

Naively accessing data in the external database, for example in a MapFunction, typically means synchronous interaction:A request is sent to the database and the MapFunction waits until the response has been received. In many cases, this waitingmakes up the vast majority of the function’s time.

Asynchronous interaction with the database means that a single parallel function instance can handle many requests concurrently andreceive the responses concurrently. That way, the waiting time can be overlayed with sending other requests andreceiving responses. At the very least, the waiting time is amortized over multiple requests. This leads in most cased to much higherstreaming throughput.

异步 I/O - 图1

Note: Improving throughput by just scaling the MapFunction to a very high parallelism is in some cases possible as well, but usuallycomes at a very high resource cost: Having many more parallel MapFunction instances means more tasks, threads, Flink-internal networkconnections, network connections to the database, buffers, and general internal bookkeeping overhead.

Prerequisites

As illustrated in the section above, implementing proper asynchronous I/O to a database (or key/value store) requires a clientto that database that supports asynchronous requests. Many popular databases offer such a client.

In the absence of such a client, one can try and turn a synchronous client into a limited concurrent client by creatingmultiple clients and handling the synchronous calls with a thread pool. However, this approach is usually lessefficient than a proper asynchronous client.

Async I/O API

Flink’s Async I/O API allows users to use asynchronous request clients with data streams. The API handles the integration withdata streams, well as handling order, event time, fault tolerance, etc.

Assuming one has an asynchronous client for the target database, three parts are needed to implement a stream transformationwith asynchronous I/O against the database:

  • An implementation of AsyncFunction that dispatches the requests
  • A callback that takes the result of the operation and hands it to the ResultFuture
  • Applying the async I/O operation on a DataStream as a transformation

The following code example illustrates the basic pattern:

  1. // This example implements the asynchronous request and callback with Futures that have the
  2. // interface of Java 8's futures (which is the same one followed by Flink's Future)
  4. /**
  5.  * An implementation of the 'AsyncFunction' that sends requests and sets the callback.
  6.  */
  7. class AsyncDatabaseRequest extends RichAsyncFunction<String, Tuple2<String, String>> {
  9.     /** The database specific client that can issue concurrent requests with callbacks */
  10.     private transient DatabaseClient client;
  12.     @Override
  13.     public void open(Configuration parameters) throws Exception {
  14.         client = new DatabaseClient(host, post, credentials);
  15.     }
  17.     @Override
  18.     public void close() throws Exception {
  19.         client.close();
  20.     }
  22.     @Override
  23.     public void asyncInvoke(String key, final ResultFuture<Tuple2<String, String>> resultFuture) throws Exception {
  25.         // issue the asynchronous request, receive a future for result
  26.         final Future<String> result = client.query(key);
  28.         // set the callback to be executed once the request by the client is complete
  29.         // the callback simply forwards the result to the result future
  30.         CompletableFuture.supplyAsync(new Supplier<String>() {
  32.             @Override
  33.             public String get() {
  34.                 try {
  35.                     return result.get();
  36.                 } catch (InterruptedException | ExecutionException e) {
  37.                     // Normally handled explicitly.
  38.                     return null;
  39.                 }
  40.             }
  41.         }).thenAccept( (String dbResult) -> {
  42.             resultFuture.complete(Collections.singleton(new Tuple2<>(key, dbResult)));
  43.         });
  44.     }
  45. }
  47. // create the original stream
  48. DataStream<String> stream = ...;
  50. // apply the async I/O transformation
  51. DataStream<Tuple2<String, String>> resultStream =
  52.     AsyncDataStream.unorderedWait(stream, new AsyncDatabaseRequest(), 1000, TimeUnit.MILLISECONDS, 100);
  1. /**
  2.  * An implementation of the 'AsyncFunction' that sends requests and sets the callback.
  3.  */
  4. class AsyncDatabaseRequest extends AsyncFunction[String, (String, String)] {
  6.     /** The database specific client that can issue concurrent requests with callbacks */
  7.     lazy val client: DatabaseClient = new DatabaseClient(host, post, credentials)
  9.     /** The context used for the future callbacks */
  10.     implicit lazy val executor: ExecutionContext = ExecutionContext.fromExecutor(Executors.directExecutor())
  13.     override def asyncInvoke(str: String, resultFuture: ResultFuture[(String, String)]): Unit = {
  15.         // issue the asynchronous request, receive a future for the result
  16.         val resultFutureRequested: Future[String] = client.query(str)
  18.         // set the callback to be executed once the request by the client is complete
  19.         // the callback simply forwards the result to the result future
  20.         resultFutureRequested.onSuccess {
  21.             case result: String => resultFuture.complete(Iterable((str, result)))
  22.         }
  23.     }
  24. }
  26. // create the original stream
  27. val stream: DataStream[String] = ...
  29. // apply the async I/O transformation
  30. val resultStream: DataStream[(String, String)] =
  31.     AsyncDataStream.unorderedWait(stream, new AsyncDatabaseRequest(), 1000, TimeUnit.MILLISECONDS, 100)

Important note: The ResultFuture is completed with the first call of ResultFuture.complete.All subsequent complete calls will be ignored.

The following two parameters control the asynchronous operations:

  • Timeout: The timeout defines how long an asynchronous request may take before it is considered failed. This parameterguards against dead/failed requests.

  • Capacity: This parameter defines how many asynchronous requests may be in progress at the same time.Even though the async I/O approach leads typically to much better throughput, the operator can still be the bottleneck inthe streaming application. Limiting the number of concurrent requests ensures that the operator will notaccumulate an ever-growing backlog of pending requests, but that it will trigger backpressure once the capacityis exhausted.

Timeout Handling

When an async I/O request times out, by default an exception is thrown and job is restarted.If you want to handle timeouts, you can override the AsyncFunction#timeout method.

Order of Results

The concurrent requests issued by the AsyncFunction frequently complete in some undefined order, based on which request finished first.To control in which order the resulting records are emitted, Flink offers two modes:

  • Unordered: Result records are emitted as soon as the asynchronous request finishes.The order of the records in the stream is different after the async I/O operator than before.This mode has the lowest latency and lowest overhead, when used with processing time as the basic time characteristic.Use AsyncDataStream.unorderedWait(…) for this mode.

  • Ordered: In that case, the stream order is preserved. Result records are emitted in the same order as the asynchronousrequests are triggered (the order of the operators input records). To achieve that, the operator buffers a result recorduntil all its preceding records are emitted (or timed out).This usually introduces some amount of extra latency and some overhead in checkpointing, because records or results are maintainedin the checkpointed state for a longer time, compared to the unordered mode.Use AsyncDataStream.orderedWait(…) for this mode.

Event Time

When the streaming application works with event time, watermarks will be handled correctly by theasynchronous I/O operator. That means concretely the following for the two order modes:

  • Unordered: Watermarks do not overtake records and vice versa, meaning watermarks establish an order boundary.Records are emitted unordered only between watermarks.A record occurring after a certain watermark will be emitted only after that watermark was emitted.The watermark in turn will be emitted only after all result records from inputs before that watermark were emitted.

That means that in the presence of watermarks, the unordered mode introduces some of the same latency and managementoverhead as the ordered mode does. The amount of that overhead depends on the watermark frequency.

  • Ordered: Order of watermarks an records is preserved, just like order between records is preserved. There is nosignificant change in overhead, compared to working with processing time.

Please recall that Ingestion Time is a special case of event time with automatically generated watermarks thatare based on the sources processing time.

Fault Tolerance Guarantees

The asynchronous I/O operator offers full exactly-once fault tolerance guarantees. It stores the records for in-flightasynchronous requests in checkpoints and restores/re-triggers the requests when recovering from a failure.

Implementation Tips

For implementations with Futures that have an Executor (or ExecutionContext in Scala) for callbacks, we suggests to use a DirectExecutor, because thecallback typically does minimal work, and a DirectExecutor avoids an additional thread-to-thread handover overhead. The callback typically only handsthe result to the ResultFuture, which adds it to the output buffer. From there, the heavy logic that includes record emission and interactionwith the checkpoint bookkeeping happens in a dedicated thread-pool anyways.

A DirectExecutor can be obtained via org.apache.flink.runtime.concurrent.Executors.directExecutor() orcom.google.common.util.concurrent.MoreExecutors.directExecutor().

Caveat

The AsyncFunction is not called Multi-Threaded

A common confusion that we want to explicitly point out here is that the AsyncFunction is not called in a multi-threaded fashion.There exists only one instance of the AsyncFunction and it is called sequentially for each record in the respective partitionof the stream. Unless the asyncInvoke(…) method returns fast and relies on a callback (by the client), it will not result inproper asynchronous I/O.

For example, the following patterns result in a blocking asyncInvoke(…) functions and thus void the asynchronous behavior:

  • Using a database client whose lookup/query method call blocks until the result has been received back

  • Blocking/waiting on the future-type objects returned by an asynchronous client inside the asyncInvoke(…) method