This page is meant as a template for writing a FLIP. To create a FLIP choose Tools->Copy on this page and modify with your content and replace the heading with the next FLIP number and a description of your issue. Replace anything in italics with your own description.

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

...

Motivation

A common UDF type is the ScalarFunction.  This works well for CPU-intensive operations, but less well for IO bound or otherwise long-running computations.  One example of this is remote calls to external systems where networking, serialization, and database lookups might dominate the execution time. StreamTask has a single thread serially executing operators and their contained calls, which happen synchronously in the remote call case. Since each call can take, say, 1 second or more, that limits throughput and the overall performance, potentially accumulating backpressure to the upstream operator. The solution is to either: increase the parallelism of the query (resulting in a higher resource cost, overhead, etc.) or asynchronously fire off many requests concurrently and receive results as they complete. This FLIP aims to address the latter solution by introducing AsyncScalarFunction, a new UDF type which allows for issuing concurrent function calls.

Scope

There are lots of combinations of modes and Job types in Flink such as the changelog mode and streaming vs batch.  To make clear the scope this FLIP intends to cover, the functionality will be limited to the following:

• Ordered Async operators: Much discussion has been centered around which changelog modes, SQL queries could be compatible with an operator which allowed unordered results, since there is a performance benefit. For now we'll only consider an operator that retains the input ordering.
• Streaming mode: Some of the design considerations we're considering are focused on streaming.  To get good performance on batch, it's possible we might want to allow batching of async calls, but we're not addressing this at the moment.

Public Interfaces

The primary public class is AsyncScalarFunction, for being the base class of all async scalar functions.  
The type is parameterized with a return type for the eval call.  This is similar to the definition of AsyncTableFunction.

public class AsyncScalarFunction<T>AsyncScalarFunction extends UserDefinedFunction {
    @Override
    public final FunctionKind getKind() {
        return FunctionKind.ASYNC_SCALAR;
    }

    @Override
    public TypeInference getTypeInference(DataTypeFactory typeFactory) {
        TypeInference val = TypeInferenceExtractor.forAsyncScalarFunction(typeFactory, getClass());
        return val;
    }
}

...

public class RemoteCallFunction extends AsyncScalarFunction<String>AsyncScalarFunction {

    private ExternalClient client;
    private ExecutorService executor;

    public RemoteCallFunction() {
    }

    @Override
    public void open(FunctionContext context) throws Exception {
        client = new Client();
        executor = Executors.newFixedThreadPool(
            context.getJobParameter("in-flight-requests", 10));
    }

    @Override
    public void close() throws Exception {
        client.close();
        executor.shutdownNow();
    }

    public final void eval(
            CompletableFuture<String> future,
            String param1,
            int param2) {
        executor.submit(() -> {
            try {
                String resp = client.rpc(param1, param2);
                future.complete(resp);
            } catch (Throwable t) {
                future.completeExceptionally(t);
            }
        });
    }
}

As with the standard ScalarFunction, there is an eval method, but with a 0th parameter of the type   CompletableFuture<String> future.
This is the primary method used to invoke the async functionality.  The generic parameter of the future is used to infer the return type for the type system.

New configurations will be introduced for the functionality, similar in nature to table.exec.async-lookup.* :

table.exec.async-scalar.catalog.db.func-name.buffer-capacity: 10
table.exec.async-scalar.catalog.db.func-name.timeout: 30s
table.exec.async-scalar.catalog.db.func-name.output-mode: ORDERED
table.exec.async-scalar.catalog.db.func-name.retry-strategy: FIXED_DELAY
table.exec.async-scalar.catalog.db.func-name.fixed-delay: 10s
table.exec.async-scalar.catalog.db.func-name.max-attempts: 3
table.exec.async-scalar.system.func-name.buffer-capacity: 10

These options ideally would be function scoped, but since `ConfigOption` doesn't make it easy to have a per-function config, they are global.  Future work could allow these to be overridden on a per definition basis.

The following configurations will be availableThese options are scoped with the catalog, db, and function name as registered in the table environment so that any given definition can be configured.  Similarly, a system function can be configured with the special prefix system, as in the last example.The options have the following meanings:

Proposed Changes

Planner Changes

Split Rules

One of the areas that have been used as inspiration for planner changes are the python calc rules.  Most of the split rules (rules for complex calc nodes being split into multiple simpler calc nodes) will be generalized and
shared between the two, since remote python calls and async calls more generally share much of the same structure.  If done correctly, the intention is to simplify the async operator to handle only FlinkLogicalCalcs
which contain async UDF calls in projections and no other calc logic (non async calls, field accesses, conditions).

The high level motivation is that anything that comes after an async call is easier to chain as a series of operators rather than internally within a single operator.

...

FlinkLogicalCalc
Projections:
Condition: func(...) AND I0

FlinkLogicalCalc
Projections: I0 AS F0, func(...) AS F1
Condition:

FlinkLogicalCalc
Projections:
Condition: F1 AND F0

FlinkLogicalCalc
Projections: Concat(func(...), I1)
Condition: 

FlinkLogicalCalc
Projections: I1 AS F0, func(...) as F1
Condition:

FlinkLogicalCalc
Projections: Concat(F1, F0) 
Condition:

FlinkLogicalCalc
Projections:func(...).foobar
Condition: 

FlinkLogicalCalc
Projections: func(...) as F0
Condition:

FlinkLogicalCalc
Projections: F0.foobar
Condition:

FlinkLogicalCalc
Projections:
Condition: func(...).foobar

FlinkLogicalCalc
Projections: 
Condition: func(...)

FlinkLogicalCalc
Projections: 
Condition: I0.foobar

FlinkLogicalCalc
Projections: func(I5.foobar)
Condition: 

FlinkLogicalCalc
Projections: I5.foobar as F0
Condition:

FlinkLogicalCalc
Projections: func(F0)
Condition: 

FlinkLogicalCalc
Projections: func(...)
Condition: C1

FlinkLogicalCalc
Projections: 
Condition: C1

FlinkLogicalCalc
Projections: func(...)
Condition:

FlinkLogicalCalc
Projections: func(func(...))
Condition:

FlinkLogicalCalc
Projections: func(...) as F0
Condition:

FlinkLogicalCalc
Projections: func(F0)
Condition:

FlinkLogicalCalc
Projections: func(...) as F0
Condition:

FlinkLogicalCalc
Projections: F0, func(...)
Condition:F0, func(...)
Condition:

Physical Rules

In additional the split rules, there will also need to be a PhysicalAsyncCalcRule which converts FlinkLogicalCalcs to PhysicalAsyncCalcs. This will check for the existence of any async calls in the calc, using the same AsyncCalcCallFinder logic above.

Disallowing Async functionality when not supported

It is most prudent to only allow async behavior where it is known to not violate SQL semantics.  
To  To do this, rules will be introduced which contain query structures which we don’t want to allow and if found, all of the async calls will be executed in synchronous mode.

This can be done by introducing a new trait AsyncOperatorModeTrait, which comes in sync mode and async mode (default), and which will be attached to a FlinkLogicalCalc
if it contains async calls which we would prefer to execute in sync mode.  Execution Execution in synchronous mode just utilizes the same stack of as async, but waits on the result immediately after issuing the request.

...

• Queries with ordering semantics:
  • e.g. SELECT func(f1) FROM Table ORDER BY f2;
    Here, we expect that the results of the query will be ordered by f2.  We wouldn't want to return results in the completion order from async function func.  
    We can solve it by either utilizing output-mode as outputting in ORDERED, and ensuring that we return the results in the input order, or by putting it into synchronous mode and ensuring ordering by doing one at a time.
• Others?  Would be great to get feedback on other cases that should be consideredWould be great to get feedback on other cases that should be considered.

For the first version of this functionality where the operator outputs only in ordered mode, synchronous mode may not need to be enabled.

Runtime Changes

Code Generation

...

Since the code generator already supports generating AsyncFunctions (currently used by lookup joins),  it this will be used with the fetching main logic in the method asyncInvoke(RowData input, ResultFuture<Collection<RowData> result).  The body of that method will use existing code generation to call the UDF and do the appropriate casting for the various arguments.  Additional logic will capture the UDF result Future, set a callback, convert results, and complete the AsyncFunction ResultFutureoutput row.

Utilizing a class AsyncDelegatingResultFuture similar to the existing DelegatingResultFuture (used for lookup joins), the generated method could look similar to the following:

@Override
public void asyncInvoke(RowData input, ResultFuture<RowData> resultFuture) throws Exception {
    {
   // Invokes callbacks on resultFuture once the async call is complete.
   final AsyncDelegatingResultFuture delegatingFuture = AsyncDelegatingResultFuture(resultFuture);

   try {
       java.util.function.Function<Object, GenericRowData> outputFactory = new java.util.function.Function<Object, GenericRowData>() {
           @Override
           public GenericRowData apply(Object udfResult) {
			   // Gather the results and return the output object
               final GenericRowData out = new GenericRowData(2);
               out.setField(0, fdelegatingFuture.getSynchronousResult(0));
               out.setField(1, udfResult);
               return out;
           }
       };

	   // Once it sees that the async future is done, the factory will be used to get the resulting output row
       delegatingFuture.setOutputFactory(outputFactory);

       // If an input is needed in the next operator, pass it along
       int passThroughField = input.getInt(0);
       delegatingFuture.addSynchronousResult(passThroughField);

	   // Create a new future object and invoke the UDF.
       // The result will be converted to the internal type before calling the output factory.
       CompletableFuture<?> udfResultFuture = delegatingFuture.createAsyncFuture(typeConverter);
       udfInstanceasyncScalarFunctionUdf.eval(udfResultFuture);
    } catch (Throwable e) {
      resultFuture.completeExceptionally(e);
    }
}

...

Since the call to the AsyncScalarFunction is wrapped in a AsyncFunction taking input rows, we have the benefit of using the existing class AsyncWaitOperator,
which handles ordering, checkpointing, timeouts and other implementation details.  Since only ordered results are handled in this scope, ORDERED will be the default behavior.

The PhysicalAsyncCalcs mentioned in the planning phase will translate to an exec node, which creates the transformation containing this operator.

Compatibility, Deprecation, and Migration Plan

...