Version 0.1.2
This document is written to serve as a reference for developers who are developing a future backend to the future framework as implemented in the future package for R that available on CRAN. The Future Application Programming Interface (API) has three fundamental functions at its core:
f <- future(expr) - create a future from an R expression (non-blocking but may be blocking) r <- resolved(f) - check whether a future is resolved or not (non-blocking)v <- value(f) - retrieve the value of a future (blocking)With these three functions alone, it is possible to evaluate one or more R expressions synchronously and asynchronously. How and where these expressions are resolved depends on which “future backend” is in use. For example, one backend may evaluated the expressions sequentially (synchronously) while another may evaluated them in parallel (asynchronously). Regardless of backend, the value of a future expression is always the same.
It is fundamental to the future ecosystem that all future backends conform to the Future API specification. Conformance serves as a guarantor of correctness and behavior for both the developer who use futures in their software as well as the end-user of their software. A future backend that meets the requirements can be used in any software that use futures internally.
For example, the above three functions serve as building blocks in several higher-level map-reduce APIs. One example is the [future.apply] package on CRAN that provides future_lapply(), which is a futurized version of lapply() available in the ‘base’ package. This function can be used to perform the lapply-like processing in parallel using a parallel backend. The implementation of the ‘future.apply’ package is 100% invariant to the parallel backend used. This is possible because all future backends conform to a set of rules. Rules that are documented below.
A supplement to the specification herein is the ‘Test Suite for Future API Backends’, which consists of a set of tests that can be used to validated that a future backend meets the minimal requirements of the Future API. These tests run from the command-line, from the R prompt, or as part of the package tests of a backend package. This test suite is documented and implemented in the future.tests package available on CRAN.
If you find that something in this document to be missing, unclear, or faulty, please report your feedback using the official issue tracker for the ‘future’ package at https://github.com/HenrikBengtsson/future. If you have feedback that is specific to the test suite, please use the official issue tracker for the ‘future.tests’ package at https://github.com/HenrikBengtsson/future.tests
The Future API has three fundamental functions at its core:
f <- future(expr) - create a future from an R expression (non-blocking but may be blocking) r <- resolved(f) - check whether a future is resolved or not (non-blocking)v <- value(f) - retrieve the value of a future (blocking)The implementation of a future backend for these involves several steps. For simplicity, lets say we call our future backend ‘myparallel’. In summary, a future backend needs to implement four API components:
myparallel() that inherits from class future. This function should return a Future object (as defined in the future package) that also inherits from S3 class MyParallelFuture if a non-lazy future is created (the default lazy = FALSE). The default should be that this function calls run() on the Future object before returning, unless a lazy is created.run() for MyParallelFuture that starts the evaluation of the future R expression part of the Future object. This method is often non-blocking for parallel backends, but may be blocking if all compute resources are exhausted. It is typically blocking for sequential backends.resolved() for MyParallelFuture that, in a non-blocking fashion, returns TRUE if the future is resolved and FALSE if not.result() for MyParallelFuture that returns a FutureResult object (as defined by the future package) when the future is resolved or otherwise fails to resolve. If the future is not yet resolved, this method should block until the future is resolved.With this in place, the selection of using this backend as the future plan, will be done as plan(myparallel) with the option of specifying certain arguments to be passed to myparallel(). With the plan set, a call to f <- future(expr) will then correspond to a f <- myparallel(expr) call. With the defaults, myparallel() will then launch the evaluation of expr asynchronously before returning the MyParallelFuture object. When calling resolved(f) to query whether the future expression is resolved or not, the underlying S3 method for this class will then check in with the parallel worker whether the expression is resolved or not. When calling value(f), the S3 method for the Future class calls result(f), which will return the FutureResult object for this future. If the future is not yet resolved, this call will block until it is. If no errors occurred while resolving the future expression, then value(f) will return the value of the expression, which is recorded by the backend in the FutureResult object. If there was an evaluation error, then value(f) will resignal (“relayed”) that error. Any captured conditions or standard output will also be relayed at this point.
This section describes in detail what the requirements of the above four components are. The requirements are given as a continuation of the above ‘myparallel’ example. If otherwise not specified, all functions mentioned below are from the future package.
The constructor function myparallel() for creating a Future object must inherits from class future such that inherits(myparallel, "future") is true.
The constructor function should have explicit arguments expr, substitute, envir, and ..., where argument expr is an expression, argument substitute is a logical (TRUE or FALSE), and envir is an environment. All or parts of the ... arguments should be passed along to the Future() function. If substitute is TRUE (default), then expr should be re-evaluated as expr <- substitute(expr). Environment envir should default to parent.frame() as it is used to identify global variables.
If the backend supports more than one worker, then it should also have an explicit workers argument.
Any currying arguments that can be specified when setting the future plan must be explicit arguments of the constructor function such that they appear as named elements in formals(myparallel). Currying arguments are arguments that can be “tweaked” by the end-user, e.g. plan(myparallel, workers = 2).
The value of the constructor function should be invisible and a MyParallelFuture object that inherits from Future. This is achieved by calling Future(), with all matching arguments passed along and then prepending "MyParallelFuture" to the class attribute of the returned Future object.
Before returning the Future object, the constructor function should launch the future if, and only if, the lazy element of the Future object (e.g. f$lazy) is FALSE (default). The constructor function must never evaluate the expression expr if lazy is TRUE.
The constructor function must not update the RNG state.
An S3 method run() for MyParallelFuture that takes a Future object as its first argument is required. It should accept additional arguments via ..., which are currently not used.
The run() method should invisibly return the Future object, which may be a modified version of the input Future object.
The run() method is responsible for not launching the same future twice. If run() is called on an already launched or a resolved future, then an informative FutureError error constructed by the FutureError() function should be produced.
The run() method is responsible for evaluation the expression returned by getExpression() with the Future object as the first argument. The evaluation of this expression should respect any global variables in the FutureGlobals object returned by globals() with the Future object as the first argument. The evaluation should also respect any package names returned by packages() with the Future object as the first argument.
If the backend provides parallel processing, then run() should return the future as soon as possible and without waiting for it to be resolved. If all workers are occupied, then run() is responsible for waiting until a worker becomes available and then launch the future on that worker and immediatedly return the future.
The run() method may produce FutureError error as created by FutureError() in case it fails to launch the future on the worker or the worker has terminated unexpectedly.
The run() method must not update the RNG state.
An S3 method resolved() for MyParallelFuture that takes a Future object as its first argument and return either TRUE or FALSE is required. It should accept additional arguments via ..., which are currently not used.
The method may be called zero or more times.
The method should return FALSE as long as the future is unresolved. It may also return FALSE if it fail to establish the state of the future within a reasonable time period (“timeout”). It should return TRUE as soon as it can be established that the future is resolved. After it has returned TRUE once, any succeeding calls should return TRUE.
If resolved() is called on a future that yet has not been launched, it should launch the future by calling run(). This is the only occasion when resolved() may block. In all other cases, it should return promptly.
The resolved() method may produce FutureError error as created by FutureError() in case communication with the worker has broken down or the worker has terminated unexpectedly.
The resolved() method must not update the RNG state.
An S3 method result() for MyParallelFuture that takes a Future object as its first argument and return a FutureResult object is required. It should accept additional arguments via ..., which are currently not used.
The method may be called zero or more times.
If result() is called on a future that yet has not been launched, it should launch the future by calling run().
If result() is called on a future that is not yet resolved, it should block until the future is resolved.
The value of result() should be the value from evaluating the getExpression() expression that run() launched.
The result() method may produce FutureError error as created by FutureError() in case communication with the worker has broken down or the worker has terminated unexpectedly.
The result() method must not update the RNG state.