From HaskellWiki
Jump to navigation Jump to search

Rhine is an arrowized FRP library with type level clocks. It is designed with multi-rate FRP in mind, combining synchronous arrowized FRP features (as in Yampa) with asynchronous building blocks.

Building blocks in Rhine

There are two kinds of separation of concerns: First, temporal aspects (clocking/sampling) are separated from dataflow aspects. Second, synchronous components are separated from asynchronous ones. The situation is summarised in this table:

Aspect separation
Time Data
Synchronous Clocks Clocked signal functions
Asynchronous Schedules Resampling buffers


In Rhine, you can annotate each (synchronous) signal function with a clock type. This clock type gives clock safety since throughout the framework it is ensured that every signal function is run at the speed of this clock.

A 'running clock' is an effectful stream of time stamps. (The side effect can include polling the system time, a device, or waiting/blocking until a certain time.) The clock is said to tick whenever a time stamp is emitted.

Rhine has a number of ready-to-use clocks, for example fixed-rate clocks like

Millisecond n
is a type-level natural number.


Event sources are also clocks in a natural sense that tick every time an event occurs.

Signal functions

Signal functions are aware of time and internal state. Standard arrowized FRP constructions are implemented in Rhine:

type Pos =(Double, Double, Double)

freeFallFrom10Meters ::ClSF m cl a Pos
freeFallFrom10Meters =arr (const (0,0,-9.81))
                   >>> integral
                   >>> integralFrom (0,0,10)

Here, m is an arbitrary monad and cl is an arbitrary clock type. a is the input data type (in this case it is arbitrary because any input is discarded), and Pos is the output data type.

Being based on Dunai, signal functions can also perform side effects in a monad m:

exitWhenNegative :: ClSF (ExceptT () m) cl Pos ()
exitWhenNegative = throwOn' <<< arr (\(_, _, z)->(z < 0, ()))

Like in standard Haskell, monad transformers need to be handled. In Rhine (and Dunai), interesting things happen when handling monad transformers. In the case of ExceptT, we have control flow!

freeFallThenOnTheGround :: ClSF m cl a Pos
freeFallThenOnTheGround =safely$ do
  try $ freeFallFrom10Meters >>> (id &&& exitWhenNegative)>>>arr fst
  safe $arr $ const (0, 0, 0)

This little signal function simulates an object that falls from 10 meters, throws an exception when the height drops below 0, and then eternally stays at the coordinates (0, 0, 0). This example is discussed at length in the article (link on the bottom of the page).

We can also use other monads, like IO:

printHeight :: ClSF IO cl a ()
printHeight =freeFallThenOnTheGround >>> arrMCl print

Rhine does not hide any side effects under the hood. All effects that are used are visible in the type.

Clock safety 

Signal functions are "clocked" and can only be composed synchronously when their clocks match. The following will not typecheck:

doubleAt1Hz :: ClSF m (Millisecond 1000) Integer Integer
doubleAt1Hz =arr (* 2)

doubleAt2Hz :: ClSF m (Millisecond  500) Integer Integer
doubleAt2Hz =arr (* 2)

quadrupleAtNonsensicalRate =doubleAt1Hz >>> doubleAt2Hz

error: * Couldn't match type `500' with `1000' [...]

In general, clock type errors will occur when trying to compose Rhine components at non-matching clocks. This feature is called 'clock-safety'. Subsystems on different clocks have to communicate over explicitly specified resampling buffers, coordinated by explicit schedules (see sections below).

Behaviours and clock-polymorphism

Signal functions can be polymorphic in the clock. In that case, they are called Behaviours, since they model the original FRP idea of a value (or function) varying with time, oblivious of the sampling/clocking strategy.


A schedule for two clocks cl1 and cl2 is a universal clock such that cl1 and cl2 are subclocks. The schedule ticks exactly whenever either cl1 or cl2 would tick.

Rhine has a number of predefined schedules, so you rarely need to implement your own. For example, you can use concurrency and let the framework run the two clocks in separate background threads, while all concurrent communication is encapsulated and hidden from the library user. For certain clocks, such as fixed-step clocks, it is also possible to schedule them 'deterministically', which is very useful when e.g. audio and video sampling ratios need to be kept stable.

Resampling buffers

A resampling buffer is the fundamental asynchronous data component in Rhine. It accepts put and get calls that put data into, and get data from, the buffer. Usually, these methods need not be called explicitly. Instead, resampling buffers connect synchronous signal functions that work at different rates.

The Rhine library implements standard buffering and resampling techniques such as FIFO queues, linear/cubic/sinc interpolation, first order holds and others. Some buffers, like fixed-rate up-/downsamplers, are annotated with clock types in order to be used only at the correct rate. Others are clock-polymorphic.

Main loops

Signal functions and resampling buffers are composed to 'signal networks'. In turn, signal networks together with compatible clocks form the main high-level components, called Rhines themselves. The library has several combinators that allow for easy creation of Rhines from basic components.

A 'closed' Rhine, i.e. one that has the trivial input and output type (), is a main loop. It is run by starting its clock, feeding the time stamps into the signal network and let it perform its side effects.

All this is done automatically by the framework without inserting any additional side effects. This makes Rhine suitable for reproducible testing, since it hides no IO under the hood.


  • Rhine is developed on https://github.com/turion/rhine. The development version usually stays up to date with the latest GHC. Support requests for applications developed with Rhine are very welcome there.
  • Rhine is on stackage.

General development guideline

To build an application with Rhine, you can usually follow this guideline:

  1. Ask yourself what the different synchronous subsystems of the whole application should be. You will recognise them by the rate they work ("tick") at. For example in a game, you might have a user event system, a physical simulation, and a graphics system.
  2. Ask yourself whether you have any special clocks, e.g. external devices, event machines you need to connect to, external loops you have to connect to. Implement those as clocks in Rhine, ideally using existing components.
  3. Implement each synchronous subsystem as a clocked signal function.
  4. Decide how to schedule the different clocks. In most cases, you will be able to use deterministic or concurrent schedules from the library.
  5. Decide how to resample the data from one subsystem to the other. Take existing resampling buffers from the library, or build your own buffers from existing ones and signal functions.
  6. Build the main Rhine program from all the components. The clock types will tell you whether you have correctly combined everything.
  7. Run the main loop and possibly add initialisation and clean up actions.

Further resources