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 and event-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 document FLIP-12: Asynchronous I/O Design and Implementation.
When interacting with external systems (for example when enriching stream events with data stored in a database), one needs to take care that 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 waiting
makes 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 and receive the responses concurrently. That way, the waiting time can be overlayed with sending other requests and receiving responses. At the very least, the waiting time is amortized over multiple requests. This leads in most cased to much higher streaming throughput.
Note: Improving throughput by just scaling the
MapFunction to a very high parallelism is in some cases possible as well, but usually
comes at a very high resource cost: Having many more parallel MapFunction instances means more tasks, threads, Flink-internal network
connections, network connections to the database, buffers, and general internal bookkeeping overhead.
As illustrated in the section above, implementing proper asynchronous I/O to a database (or key/value store) requires a client to 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 creating multiple clients and handling the synchronous calls with a thread pool. However, this approach is usually less efficient than a proper asynchronous client.
Flink’s Async I/O API allows users to use asynchronous request clients with data streams. The API handles the integration with data 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 transformation with asynchronous I/O against the database:
AsyncFunctionthat dispatches the requests
The following code example illustrates the basic pattern:
Important note: The
ResultFuture is completed with the first call of
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 parameter guards 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 in the streaming application. Limiting the number of concurrent requests ensures that the operator will not accumulate an ever-growing backlog of pending requests, but that it will trigger backpressure once the capacity is exhausted.
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
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.
AsyncDataStream.unorderedWait(...) for this mode.
Ordered: In that case, the stream order is preserved. Result records are emitted in the same order as the asynchronous
requests are triggered (the order of the operators input records). To achieve that, the operator buffers a result record
until 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 maintained
in the checkpointed state for a longer time, compared to the unordered mode.
AsyncDataStream.orderedWait(...) for this mode.
When the streaming application works with event time, watermarks will be handled correctly by the asynchronous 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 management overhead as the ordered mode does. The amount of that overhead depends on the watermark frequency.
Ordered: Order of watermarks and records is preserved, just like order between records is preserved. There is no significant 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 that are based on the sources processing time.
The asynchronous I/O operator offers full exactly-once fault tolerance guarantees. It stores the records for in-flight asynchronous requests in checkpoints and restores/re-triggers the requests when recovering from a failure.
For implementations with Futures that have an Executor (or ExecutionContext in Scala) for callbacks, we suggests to use a
DirectExecutor, because the
callback typically does minimal work, and a
DirectExecutor avoids an additional thread-to-thread handover overhead. The callback typically only hands
the result to the
ResultFuture, which adds it to the output buffer. From there, the heavy logic that includes record emission and interaction
with the checkpoint bookkeeping happens in a dedicated thread-pool anyways.
DirectExecutor can be obtained via
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 partition
of the stream. Unless the
asyncInvoke(...) method returns fast and relies on a callback (by the client), it will not result in
proper 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
The operator for AsyncFunction (AsyncWaitOperator) must currently be at the head of operator chains for consistency reasons
For the reasons given in issue
FLINK-13063, we currently must break operator chains for the
AsyncWaitOperator to prevent
potential consistency problems. This is a change to the previous behavior that supported chaining. User that
require the old behavior and accept potential violations of the consistency guarantees can instantiate and add the
AsyncWaitOperator manually to the job graph and set the chaining strategy back to chaining via