CurrentModule = ExtensibleEffects
DocTestSetup = quote
using ExtensibleEffects
end
ExtensibleEffects.jl comes packed with a bunch of effects which can be used right away and in addition provide helpful insights into how to create your own effects. A lot is possible within ExtensibleEffects, but there are also limitations.
Vector, Iterable, and also Writer only need to define eff_applies
# Vector is supported generically as AbstractArray
ExtensibleEffects.eff_applies(handler::Type{T}, effectful::T) where {T<:AbstractArray} = true
ExtensibleEffects.eff_applies(handler::Type{<:Iterable}, effectful::Iterable) = true
ExtensibleEffects.eff_applies(handler::Type{<:Writer}, effectful::Writer) = trueeff_applies always needs to be given explicitly. This is especially needed for the autorun feature, and explicitness is also in general useful.
The definitions for eff_pure and eff_flatmap follows their default implementations, which use functionality from TypeClasses.
ExtensibleEffects.eff_pure(T, value) = TypeClasses.pure(T, value)
function ExtensibleEffects.eff_flatmap(continuation, a)
a_of_eff_of_a = map(continuation, a)
eff_of_a_of_a = flip_types(a_of_eff_of_a)
eff_of_a = map(flatten, eff_of_a_of_a)
eff_of_a
endIdentity and Const, i.e. Option, Try and Either, need a bit more interaction with the effect system.
The design decision of unifying Option/Try/Either by separating normal behaviour into Identity and stopping behaviour into Const has many advantages. One difficulty, however, is that Const does not have a TypeClasses.pure implementation, so how do we define eff_pure? The answer is an important insight into how ExtensibleEffects work: The pure function is only called on the final values within all the nested eff_flatmap calls. After destructuring all nested effects into plain values, the effects get rewrapped around the plain values. A Const, however, is always constant - there is no inner value to work on. That is why we can freely use any eff_pure definition, as it will never be run in case of a Const. However it will be run within other handlers for reconstructing their stack of wrappers. Hence the simplest and sound eff_pure implementation is to do nothing but leave the value unchanged. All in all this is the implementation for Const
ExtensibleEffects.eff_flatmap(continuation, effectful::Const) = effectful
ExtensibleEffects.eff_pure(handler::Type{<:Const}, value) = valueYou see, the continuation is ignored and eff_pure just returns the very same value.
The interaction between Const and Identity needs to be handled in addition. If something would return a Const the default implementation of eff_pure for Identity would wrap it into an Identity layer, resulting into Identity(Const(...)). This differs from the TypeClasses.flatmap implementation which would return just a Const(...). This interaction is crucial for Option/Try/Either. We can implement it by special casing the eff_pure
ExtensibleEffects.eff_pure(handler::Type{<:Identity}, value) = Identity(value)
ExtensibleEffects.eff_pure(handler::Type{<:Identity}, value::Const) = value
ExtensibleEffects.eff_flatmap(continuation, effectful::Identity) = continuation(effectful.value)Take a look at the definition of eff_flatmap for Identity. eff_flatmap gets a continuation which when called returns an Eff{YourEffectType{Value}} and should always return an Eff{YourEffectType{Value}} as well, i.e. Eff{Identity{Value}} in our case. Usually it is quite difficult to apply the continuation and return the same type, but for Identity the case is super simple, as we can just strip away the outer Identity layer.
Finally, in case you disable the default autorun feature, you may also want to use Option/Try/Either as the handlers instead of specifying the two handlers Identity and Const separately. This is enabled by explicitly defining eff_applies and forwarding the eff_pure to the case for Identity. The eff_flatmap will automatically fallback to those for Identity or Const respectively.
ExtensibleEffects.eff_applies(handler::Type{<:Either}, effectful::Either) = true
ExtensibleEffects.eff_pure(handler::Type{<:Either}, value) = ExtensibleEffects.eff_pure(Identity, value) # Const would never reach this
All these effects can still be run automatically, without any further context parameters. In ExtensibleEffect terms, this means that Option, Vector, Iterable and the like are handlers by its own and can be autorun.
We acknowledge this everytime we define eff_applies like ExtensibleEffects.eff_applies(handler::Type{<:HandlerType}, value::HandlerType) = true.
The Writer monad is actually similar to Vector when using its default accumulator TypeClasses.neutral.
julia> @syntax_eff begin
a = Writer("hello.", 1)
b = Writer("world.", 2)
@pure a + b
end
Writer{String, Int64}("hello.world.", 3)
However, if you want to use another accumulator as the default one, extra adaptations are required. Unlike Option such a parameter cannot be added by simply overloading eff_pure or the like, but we need an extra handler type which can carry all additional information needed.
julia> @runhandlers (WriterHandler("pure-accumulator."), Vector) @syntax_eff_noautorun begin
a = Writer("hello.", 1)
b = Writer("world.", 2)
c = [3, 4]
@pure a + b + c
end
Writer{String, Vector{Int64}}("hello.world.pure-accumulator.", [6, 7])
julia> @runhandlers (Vector, WriterHandler("pure-accumulator.")) @syntax_eff_noautorun begin
a = Writer("hello.", 1)
b = Writer("world.", 2)
c = [3, 4]
@pure a + b + c
end
2-element Vector{Writer{String, Int64}}:
Writer{String, Int64}("hello.world.pure-accumulator.", 6)
Writer{String, Int64}("hello.world.pure-accumulator.", 7)
The example may be slightly artifical, however you can see nicely how the eff_pure function works differently when using an explicit WriteHandler handler. Also, you do not see multiple "pure-accumulator." accumulating for each single value. Instead you have tight control about whether first the indeterminism by Vector should be run and then the accumulator by Writer or the other way arround. Either way, in both cases everything is sound and nice.
The handler is defined by
struct WriterHandler{Acc}
pure_acc::Acc
end
ExtensibleEffects.eff_applies(handler::WriterHandler, effectful::Writer) = true
ExtensibleEffects.eff_pure(handler::WriterHandler, value) = Writer(handler.pure_acc, value)eff_flatmap is the same as for pure Writer and defined as
function ExtensibleEffects.eff_flatmap(continuation, a::Writer)
eff_of_writer = continuation(a.value)
map(eff_of_writer) do b
Writer(a.acc ⊕ b.acc, b.value)
end
endThere are other effects which really require external parameters in order to work at all. One such monad is Callable, i.e. plain functions. A function needs to be called to get at its value, however with which args and kwargs should the function be called? These need to be given somehow, and hence we require for a specific handler like seen for WriterHandler. It is called CallableHandler and is defined as follows
struct CallableHandler{Args, Kwargs}
args::Args
kwargs::Kwargs
end
ExtensibleEffects.eff_applies(handler::CallableHandler, value::Callable) = true
ExtensibleEffects.eff_pure(handler::CallableHandler, a) = a
function ExtensibleEffects.eff_flatmap(handler::CallableHandler, continuation, a::Callable)
continuation(a(handler.args...; handler.kwargs...))
endIn addition, we would like to construct a Callable as the return value, i.e. wrapping everything into a Function itself which asks for the args and kwargs. We do this by a special macro @runcallable which can be composed with other such outer wrappers thanks to the ExtensibleEffects.@insert_into_runhandlers helper macro.
@runcallable efftranslates to
Callable(function(args...; kwargs...)
@insert_into_runhandlers CallableHandler(args...; kwargs...) eff
end)Types like Callable cannot be autorun, and thankfully, ExtensibleEffects can recognize this automatically. All in all we get
julia> f = @runcallable @syntax_eff begin
a = Callable(x -> x*x)
b = Callable(x -> 2x)
@pure a + b
end;
julia> f(7)
63
Another example for a Monad which requires an outer wrapper is the ContextManager type from DataTypesBasic.jl. It needs a continuation to be run, and ideally should return everything wrapped into a new ContextManager so that the executions can be lazy. Very analogous to the requirements of Callable, and indeed there is ContextManagerHandler(continuation) and @runcontextmanager, just like for Callable. There is one extra caveat which is very special to ContextManagers and that is the execution order.
It turns out, ContextManager is actually one of the most difficult computational contexts to be represented within Extensible Effects. This is because, within Eff the continuations always return another Eff, within which other effects still need to be handled. The computation inside was not fully executed and is hold frozen for later execution. However, the ContextManager semantics really depend on the assumption, that whatever uses the internal value has finished before the ContextManager exits. Otherwise loaded resources may have been destroyed before they are actually used.
The only way to solve this conflict is that the ContextManager handler needs to run last. This is actually checked by the ContextManagerHandler, however as @runcontextmanager wraps everything again into a lazy ContextManager, you need to run it to see the error. The macro version @runcontextmanager_ (with an underscore) will immediately execute the ContextManager and hence would also directly throw the error.
There is further a special handler ContextManagerCombinedHandler which improves the execution order for memory performance.
Let's have a simple factory for ContextManagers for example purposes
julia> create_context(x) = @ContextManager continuation -> begin
println("before $x")
result = continuation(x)
println("after $x")
result
end
create_context (generic function with 1 method)
julia> println_return(x) = (println(x); x)
println_return (generic function with 1 method)
Using ContextManagerHandler we get
julia> eff = @syntax_eff begin
a = [100,200]
b = create_context(a)
c = [5,6]
d = create_context(a + c)
@pure a, b, c, d
end;
julia> @runhandlers (ContextManagerHandler(println_return),) eff
before 100
before 105
before 106
before 200
before 205
before 206
[(100, 100, 5, 105), (100, 100, 6, 106), (200, 200, 5, 205), (200, 200, 6, 206)]
after 206
after 205
after 200
after 106
after 105
after 100
4-element Vector{NTuple{4, Int64}}:
(100, 100, 5, 105)
(100, 100, 6, 106)
(200, 200, 5, 205)
(200, 200, 6, 206)
Alternatively, we can use a combined handler ContextManagerCombinedHandler which runs both handlers at once. Intuitively you may think this cannot change anything, but indeed, quite a lot is changed.
julia> eff = @syntax_eff noautorun(Vector) begin
a = [100,200]
b = create_context(a)
c = [5,6]
d = create_context(a + c)
@pure a, b, c, d
end
Eff(effectful=[100, 200], length(cont)=1)
julia> handlers = (ContextManagerCombinedHandler(Vector, println_return),)
(ContextManagerCombinedHandler{Type{Vector{T} where T}, typeof(println_return)}(Vector{T} where T, ContextManagerHandler{typeof(println_return)}(println_return)),)
julia> @runhandlers handlers eff
before 100
before 105
(100, 100, 5, 105)
after 105
before 106
(100, 100, 6, 106)
after 106
after 100
before 200
before 205
(200, 200, 5, 205)
after 205
before 206
(200, 200, 6, 206)
after 206
after 200
4-element Vector{NTuple{4, Int64}}:
(100, 100, 5, 105)
(100, 100, 6, 106)
(200, 200, 5, 205)
(200, 200, 6, 206)
The difference lies in the different continuations (stored within Eff.cont) which are created internally during the run of the handler. When the Vector handler was executed independently beforehand, the ContextManagerHandler worked with far larger continuations, which actually reached up to the end of the entire computation. Now when running both handlers at once, the continuation is more intuitive, namely the one which only goes up to the end of the for loop iteration (thinking of Vector handling like for loop execution).
The most advanced possibility of defining effects is showcased by State. It needs a custom handler which provides the initial state information StateHandler(state), similar to Callable. In addition, every effect actually adapts the state, and hence the handler itself must be updated accordingly so that subsequent effects are actually handled with the correct current state. This can be done by implementing your individual ExtensibleEffects.runhandler method. Let's take a look on how StateHandler is implemented.
struct StateHandler{T}
state::T
end
ExtensibleEffects.eff_applies(handler::StateHandler, effectful::State) = true
ExtensibleEffects.eff_pure(handler::StateHandler, value) = (value, handler.state)The type is very simple, it just stores the initial or current state. eff_applies is defined the standard way, and eff_pure is not returning a State object itself, but instead what a standard State function state -> (value, state) would return. This is similar to Callable where eff_pure also does not construct a Callable. A difference though is that here we need to change the return value, appending the state, so that outer wrappers can construct a fully valid State object.
Now comes the trick to pass on the internal states
function ExtensibleEffects.runhandler(handler::StateHandler, eff::Eff)
eff_applies(handler, eff.effectful) || return runhandler_not_applies(handler, eff)
value, nextstate = eff.effectful(handler.state)
nexthandler = StateHandler(nextstate)
if isempty(eff.cont)
_eff_pure(nexthandler, value)
else
runhandler(nexthandler, eff.cont(value))
end
endExtensibleEffects.runhandler is the key function to overwrite if you would like to define what happens between evaluations of your effectful type.
- The first line checks whether our handler
StateHandleractually applies to the given effectful, and if not, we return the result of a helper functionrunhandler_not_applies. - Everything which follows implements the case that our handler applies and hence assumes that
eff.effectfulis of typeState. - We run the
State, given our previous state from the handler, and get the nextstate which we use to construct the nexthandler. - Depending on whether we already reached the end of our effect program
eff, we either stop with_eff_pure, or recurse intorunhandleragain.
Finally, similar to Callable and ContextManager there also exists a run macro for State, called runstate, which will wrap everything into a State within which the StateHandler is run. Let's see everything in action
julia> state_eff = @runstate @syntax_eff begin
a = State(x -> (x+2, x*x))
b = State(x -> (a + x, x+1))
@pure a, b
end;
julia> state_eff(3)
((5, 14), 10)
Note that unlike Callable and ContextManager, a State always has to ensure that it is the first outer wrapper being run,
because it returns the inner state as an additional argument.
If you would nest it within @runcallable, e.g. like @runstate @runcallable eff it wouldn't work,
as now the appended state is within the Callable and not directly within the State, violating the definition of State.
Within the ExtensibleEffects framework, different handlers naturally compose nicely in order to execute all the given effects. Constructing wrappers around the solution, like it does @runcallable for example, is not part of the standard extensible effects framework though. Nevertheless, especially if you want to replace Monads (TypeClasses.jl) with ExtensibleEffects.jl, being able to construct these wrappers easily and composable makes for a very intuitive plug-and-play interface.
In order to compose wrappers, we currently use a macro approach. Each wrapper (including custom ones) constructs its wrapper-code using macros. Take a look at @runcallable
again. It is defined like
macro runcallable(expr)
esc(:(Callable(function(args...; kwargs...)
@insert_into_runhandlers CallableHandler(args...; kwargs...) ($expr)
end)))
endi.e.
@runcallable efftranslates to
Callable(function(args...; kwargs...)
@insert_into_runhandlers CallableHandler(args...; kwargs...) eff
end)@insert_into_runhandlers will expand all inner macros and search for a call to runhandlers in order to insert the newly constructed CallableHandler. If no runhandlers is found, it will call it itself.
Using this style, all wrappers can be written very concisely and combined in arbitrary order (in principle). There is an alternative way of implementating it using nested functions instead of macros, however the syntax would look more polluted and we haven't seen crucial disadvantages of the macro approach yet.
Despite this nice generic composability, as @runstate needs always be run first and @runcontextmanager needs to run last, in our concrete example the order is given by the semantics of the wrappers: It would be @runcontextmanager @runcallable @runstate
julia> contextmanager_callable_state = @runcontextmanager @runcallable @runstate @syntax_eff begin
co = create_context(42)
st = State(s -> (co+s, s*s))
ca = Callable(x -> x + st + co)
@pure co, st, ca
end;
julia> typeof(contextmanager_callable_state).name
typename(ContextManager)
julia> typeof(contextmanager_callable_state(println_return)).name
typename(Callable)
julia> typeof(contextmanager_callable_state(println_return)(3)).name
typename(State)
julia> contextmanager_callable_state(println_return)(3)(9)
before 42
((42, 51, 96), 81)
after 42
((42, 51, 96), 81)
ExtensibleEffects can do quite a lot, however some things are just not possible.
One such limitation are infinite iterables. They are not supportable. The reason is that ExtensibleEffects always run effects by translating them to operations on plain values. With an infinite iterable there is no way to extract all the values though. If you know more about your infite iterable, it may become possible again. For instance it could be a simple infinite repetition of the same element. In such cases we indeed can extract "all" values at once, operate on it, and reconstruct the repeating iterable around it.
A second limitation is that we need to know a way to wrap a plain value into our handled effect. While we can always make an a into an [a], it is not possible to wrap it into a Dict for instance. Which key would you choose? In general, a function like this is called pure, and it is needed because ExtensibleEffects work by first breaking down everything into values before reconstructing the respective effects from scratch.
Compare both limitations to the fully supported Callable type. Also here, the value is hidden within the respective function, however we know how to principally get the value, namely by applying correct args and kwargs to the callable. Hence we can indeed operate on the value within a Callable, which lets us support Callable by ExtensibleEffects. Something like this is not possible with infinite iterables. Also we can wrap a value into a Callable by creating a constant function (args...; kwargs...) -> a, which is not possible for a dictionary.