Status
...
Page properties | ||
---|---|---|
Document the state by adding a label to the FLIP page with one of "discussion", "accepted", "released", "rejected".
|
...
...
|
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
Apache
...
JIRA: tbd
Released: tbd
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
Apache Flink has a rich connector ecosystem that can persist data in various destinations. Flink natively supports Apache Kafka, Amazon Kinesis Data Streams, Elasticsearch, HBase, and many more destinations. Additional connectors are maintained in Apache Bahir or directly on GitHub. The basic functionality of these sinks is quite similar. They batch events according to user defined buffering hints, sign requests and send them to the respective endpoint, retry unsuccessful or throttled requests, and participate in checkpointing. They primarily just differ in the way they interface with the destination. Yet, all the above-mentioned sinks are developed and maintained independently.
...
There are two user-facing aspects of the generic sink. First, an abstract class that is used to implement a new sink for a concrete destination. Second, the interface that is used by end-users, who want to leverage an existing sink to persist events in a destination. Appendix A contains a simplified sample implementation for a Kinesis Data Stream sink.
The async sink is based on FLIP-143 and FLIP-177. It is based on the following generic types to be extensible and remain agnostic to the destination.
...
Code Block | ||||
---|---|---|---|---|
| ||||
public abstract class AsyncSinkWriter<InputT, RequestEntryT extends Serializable> implements SinkWriter<InputT, Collection<CompletableFuture<?>>Void, Collection<RequestEntryT>> { /** * This method specifies how to persist buffered request entries into the * destination. It is implemented when support for a new destination is * added. * <p> * The method is invoked with a set of request entries according to the * buffering hints (and the valid limits of the destination). The logic then * needs to create and execute the request against the destination (ideally * by batching together multiple request entries to increase efficiency). * The logic also needs to identify individual request entries that were not * persisted successfully and resubmit them using the {@code * requeueFailedRequestEntryrequestResult} method. * <p> * TheDuring methodcheckpointing, returnsthe asink futureneeds thatto indicates,ensure oncethat completed,there thatare allno * requestoutstanding entries that have been passed to the method on invocation have * either been successfully persisted in the destination or have beenin-flight requests. * * @param requestEntries a set of request entries that should be sent to the * re-queued. * <p> * During checkpointing, the sink needs to ensure that there are nodestination * outstanding in-flight requests. Ie,@param requestResult a ResultFuture that allneeds to futuresbe returnedcompleted byonce thisall * method are completed. * * @param requestEntries a set ofrequest requestsentries that shouldhave bebeen sentpassed to the APImethod * endpoint on invocation have either * @return a future that completes when all request entriesbeen successfully * persisted in the destination or have been * successfully persisted to the API or were re-queued */ protected abstract CompletableFuture<?>void submitRequestEntries(List<RequestEntryT> requestEntries, ResultFuture<RequestEntryT> requestResult); ... } |
Internally, the AsyncSinkWriter
buffers RequestEntryT
s and invokes the submitRequestEntries
method with a set of RequestEntryT
s according to user specified buffering hints. The AsyncSinkWriter
also tracks in-flight requests, ie, calls to the API that have been sent but not completed. During a commit, the sink enforces that all in-flight requests have completed and currently buffered RequestEntryT
s are persisted in the application state snapshot.
Code Block | ||||
---|---|---|---|---|
| ||||
/** * The ElementConverter provides a mapping between for the elements of a * stream to request entries that can be sent to the destination. * <p> * The resulting request entry is buffered by the AsyncSinkWriter and sent * to the destination when the {@code submitRequestEntries} method is * invoked. */ private final ElementConverter<InputT, RequestEntryT> elementConverter; /** * Buffer to hold request entries that should be persisted into the * destination. * <p> * A request entry contain all relevant details to make a call to the * destination. Eg, for Kinesis Data Streams a request entry contains the * payload and partition key. * <p> * It seems more natural to buffer InputT, ie, the events that should be * persisted, rather than RequestEntryT. However, in practice, the response * of a failed request call can make it very hard, if not impossible, to * reconstruct the original event. It is much easier, to just construct a * new (retry) request entry from the response and add that back to the * queue for later retry. */ private final BlockingDeque<RequestEntryT>Deque<RequestEntryT> bufferedRequestEntries = new LinkedBlockingDeque<>(...)ArrayDeque<>(); /** * Tracks all pending async calls that have been executed since the last * checkpoint. Calls that already completed (successfully or unsuccessfully) are * are automatically removeddecrementing from the queuecounter. Any request entry that was not * successfully persisted needneeds to be handled and retried by the logic in * {@code submitRequestsToApi}. * <p> * There is a limit on the number of concurrent (async) requests that can be * handled by the client library. This limit is enforced by checking the * size of this queue before issuing new requests. * <p> * To complete a checkpoint, we need to make sure that no requests are in * flight, as they may fail, which could then lead to data loss. */ private BlockingDeque<CompletableFuture<?>> inFlightRequests = new LinkedBlockingDeque<>(...)int inFlightRequestsCount; @Override public void write(InputT element, Context context) throws IOException, InterruptedException { // blocks if too many elements have bufferedRequestEntries.putLast(elementConverter.apply(element, context)); been buffered } /** * The entire request may fail or single request entries that are part of while (bufferedRequestEntries.size() >= MAX_BUFFERED_REQUESTS_ENTRIES) { mailboxExecutor.yield(); *} the request may not be persisted successfully, eg, because of network bufferedRequestEntries.add(elementConverter.apply(element, context)); * issues// orblocks serviceif sidetoo throttling.many Allasync requestrequests entriesare thatin failed withflight * transient failures needflush(); to be re-queued with this method so that aren't } /** * lost and can be retried later.Persists buffered RequestsEntries into the destination by invoking {@code * <p> * Request entries that are causing the same error in a reproducible manner,submitRequestEntries} with batches according to the user specified * buffering hints. * * eg, ill-formed request entries, must not be re-queued but the error needsThe method blocks if too many async requests are in flight. */ to be handled inprivate thevoid logic of {@code submitRequestEntries}. Otherwiseflush() throws InterruptedException { * these request entries will be retried indefinitely, always causing thewhile (bufferedRequestEntries.size() >= MAX_BATCH_SIZE) { * same error. *// create a batch protectedof voidrequest requeueFailedRequestEntry(RequestEntryT requestEntry) throws InterruptedException { entries that should be persisted in the destination bufferedRequestEntries.putFirst(requestEntry); } ArrayList<RequestEntryT> batch = new ArrayList<>(MAX_BATCH_SIZE); /** * Persists buffered RequestsEntries into the destination by invoking {@codewhile (batch.size() <= MAX_BATCH_SIZE && !bufferedRequestEntries.isEmpty()) { * submitRequestEntries} with batches according to the user specified * buffering hints.try { */ public void flush() throws InterruptedException { while batch.add(bufferedRequestEntries.sizeremove() >= MAX_BATCH_SIZE) {); // create a batch of} requestcatch entries(NoSuchElementException thate) should{ be persisted in the destination ArrayList<RequestEntryT> batch = new ArrayList<>(MAX_BATCH_SIZE); for (int i = 0; i < MAX_BATCH_SIZE; i++) { // if there are not enough elements, just create a smaller batch break; try { } batch.add(bufferedRequestEntries.remove());} ResultFuture<RequestEntryT> requestResult = } catch (NoSuchElementException e) { failedRequestEntries -> mailboxExecutor.execute( // if there are not enough elements, just create a smaller batch () -> completeRequest(failedRequestEntries), break; } "Mark in-flight request as completed } and requeue %d request entries", logger.info("submit requests for {} elements", batch failedRequestEntries.size()); //while call(inFlightRequestsCount the destination specific code that actually persists the request entries >= MAX_IN_FLIGHT_REQUESTS) { CompletableFuture<?> future = submitRequestEntries(batchmailboxExecutor.yield(); //} keep track of in flight request inFlightRequests.put(future)inFlightRequestsCount++; // remove the request from the tracking queue once it competedsubmitRequestEntries(batch, requestResult); } } /** future.whenComplete((response, err) -> { * Marks an in-flight request as completed and prepends failed requestEntries back to the inFlightRequests.remove(future); * internal requestEntry buffer for later retry. * }); * @param failedRequestEntries requestEntries } that need to be }retried */** private * In-flight requests may fail, but they will be retried if the sink is void completeRequest(Collection<RequestEntryT> failedRequestEntries) { inFlightRequestsCount--; * still healthy. * <p> * To not lose any requests, there cannot be any outstanding in-flight * requests when a commit is initialized. To this end, all in-flight * requests need to be completed as part of the pre commit. */ @Override public List<Collection<CompletableFuture<?>>> prepareCommit(boolean flush) throws IOException { // reuse current inFlightRequests as commitable and create an empty queue // to avoid copy and clearing List<Collection<CompletableFuture<?>>> committable = Collections.singletonList(inFlightRequests); // all// By just iterating through failedRequestEntries, it reverses the order of the // failedRequestEntries. It doesn't make a difference for kinesis:putRecords, as the api // does not make any order guarantees, but may cause avoidable reorderings for other // destinations. failedRequestEntries.forEach(bufferedRequestEntries::addFirst); } /** * In flight requests will be retried if the sink is still healthy. But if in-flight requests * fail after a checkpoint has been triggered and Flink needs to recover from the checkpoint, * the (failed) in-flight requests are handled by the AsyncSinkCommiter gone and newcannot be retried. Hence, there cannot be any // elements cannot be* addedoutstanding toin-flight therequests queuewhen during a commit, so it's save to is initialized. * * //<p>To createthis aend, new queue inFlightRequests = new ConcurrentLinkedQueue<>(); all in-flight requests need to completed before proceeding with the commit. */ return committable;@Override } public List<Void> prepareCommit(boolean /** * All in-flight requests have been completed, but there may still be flush) throws IOException, InterruptedException { if (flush) { * request entries in the internal buffer that are yet to be sent to the flush(); } // *wait endpoint.until Theseall requestin-flight entriesrequests arecompleted stored in the snapshot state so that while (inFlightRequestsCount > 0) *{ they don't get lost in case of a failure/restart of the applicationmailboxExecutor.yield(); */ } @Override public List<Collection<RequestEntryT>>return snapshotStateCollections.emptyList() throws IOException {; } /** * All return Collections.singletonList(bufferedRequestEntries);in-flight requests have been completed, but there may still be } |
Limitations
...
The sink usually persist InputTs in the order they are added to the sink, but reorderings may occur, eg, when RequestEntryTs need to be retried.
...
* request entries in the internal buffer that are yet to be sent to the
* endpoint. These request entries are stored in the snapshot state so that
* they don't get lost in case of a failure/restart of the application.
*/
@Override
public List<Collection<RequestEntryT>> snapshotState() throws IOException {
return Collections.singletonList(bufferedRequestEntries);
}
|
Limitations
- The sink is designed for destinations that provide an async client. Destinations that cannot ingest events in an async fashion cannot be supported by the sink.
The sink usually persist InputTs in the order they are added to the sink, but reorderings may occur, eg, when RequestEntryTs need to be retried.
- We are not focusing on support for exactly-once semantics beyond simple upsert capable and idempotent destinations at this point.
Appendix A – Simplified example implementation of a sink for Kinesis Data Streams
Example ElementConverter for Kinesis Data Streams
This is the functionality that needs to be implemented by end users of the sink. It specifies how an element of a DataStream is mapped to a PutRecordsRequestEntry
that can be submitted to the Kinesis Data Streams API.
Code Block | ||||
---|---|---|---|---|
| ||||
new ElementConverter<InputT, PutRecordsRequestEntry>() {
|
Appendix A – Simplified example implementation of a sink for Kinesis Data Streams
Example ElementConverter for Kinesis Data Streams
This is the functionality that needs to be implemented by end users of the sink. It specifies how an element of a DataStream is mapped to a PutRecordsRequestEntry
that can be submitted to the Kinesis Data Streams API.
Code Block | ||||
---|---|---|---|---|
| ||||
new ElementConverter<InputT, PutRecordsRequestEntry>() {
@Override
public PutRecordsRequestEntry apply(InputT element, SinkWriter.Context context) {
return PutRecordsRequestEntry
.builder()
.data(SdkBytes.fromUtf8String(element.toString()))
.partitionKey(String.valueOf(element.hashCode()))
.build();
}
} |
...
Code Block | ||||
---|---|---|---|---|
| ||||
private class AmazonKinesisDataStreamWriter extends AsyncSinkWriter<InputT, PutRecordsRequestEntry> { @Override protected CompletableFuture<?> submitRequestEntries(List<PutRecordsRequestEntry> requestEntries) { private class AmazonKinesisDataStreamWriter extends AsyncSinkWriter<InputT, PutRecordsRequestEntry> { @Override protected void submitRequestEntries(List<PutRecordsRequestEntry> requestEntries, ResultFuture<PutRecordsRequestEntry> requestResult) { // create a batch request PutRecordsRequest batchRequest = PutRecordsRequest .builder() .records(requestEntries) // create a batch request .streamName(streamName) PutRecordsRequest batchRequest = PutRecordsRequest .build(); .builder() // call api with batch request CompletableFuture<PutRecordsResponse> future = client.recordsputRecords(requestEntriesbatchRequest); // re-queue elements of failed requests .streamName(streamName) future.whenComplete((response, err) -> { if (response.buildfailedRecordCount(); > 0) { // call api with batch request ArrayList<PutRecordsRequestEntry> CompletableFuture<PutRecordsResponse>failedRequestEntries future= =new clientArrayList<>(response.putRecordsfailedRecordCount(batchRequest)); // re-queue elements of failed requests CompletableFuture<PutRecordsResponse>List<PutRecordsResultEntry> handleResponserecords = futureresponse.records(); .whenComplete((response, err) -> { for (int i = 0; i < if (response.failedRecordCountrecords.size() > 0; i++) { List<PutRecordsResultEntry> if (records.get(i).errorCode() != response.records(); null) { for (int i = 0; i < records.size(); i++) { failedRequestEntries.add(requestEntries.get(i)); } if (records.get(i).errorCode() != null) { } requeueFailedRequest(requestEntries.get(i)); requestResult.complete(failedRequestEntries); } else }{ }requestResult.complete(Collections.emptyList()); } //TODO: handle errors of the entire request... }); } // return future to track completion of async request return handleResponse; } ... } |
...
...
} |
Rejected Alternatives
The sink needs a way to build back pressure, eg, if the throughput limit of the destination is exceeded. Initially we were planning to adopt the isAvailable pattern from the source interface. But the benefits are too vague at this point and it would require substantial changes to the sink API. We'll hence start with a blocking implementation of the write
function and see how far we get.
Code Block | ||||
---|---|---|---|---|
| ||||
/** * Signals if enough RequestEntryTs have been buffered according to the user * specified buffering hints to make a request against the destination. This * functionality will be added to the sink interface by means of an * additional FLIP. * * @return a future that will be completed once a request against the * destination can be made */ public CompletableFuture<Void> isAvailable() { ... } |
...