Lazy evaluation for python 3
I will write this later…
What is lazy?
lazy is a module for making python lazily evaluated (kinda).
lazy runs under python 3.5 and 3.4.
Why not lazy?
I think lazy computation is pretty cool, I also think python is pretty cool; combining them is double cool.
How to lazy?
There are 3 means of using lazy code:
- IPython cell and line magics
lazy_function takes a python function and returns a new function that is the lazy version. This can be used as a decorator.
@lazy_function def f(a, b): return a + b
Calling f(1, 2) will return a thunk that will add 1 and 2 when it needs to be strict. Doing anything with the returned thunk will keep chaining on more computations until it must be strictly evaluated.
Lazy functions allow for lexical closures also:
@lazy_function def f(a): def g(b): return a + b return g
When we call f(1) we will get back a thunk like we would expect; however, this thunk is wrapping the function g. Because g was created in a lazy context, it will also be a lazy_function implicitly. This means that type(f(1)(2)) is thunk; but, f(1)(2) == 3.
We can use strict to strictly evaluate parts of a lazy function, for example:
>>> @lazy_function ... def no_strict(): ... print('test') ... >>> strict(no_strict())
In this example, we never forced print, so we never saw the result of the call. Consider this function though:
>>> @lazy_function ... def with_strict(): ... strict(print('test')) ... >>> strict(with_strict()) test >>> result = with_strict() >>> strict(result) test
Here we can see how strict works inside of a lazy function. strict causes the argument to be strictly evaluated, forcing the call to print. We can also see that just calling with_strict is not enough to evaluate the function, we need to force a dependency on the result.
This is implemented at the bytecode level to frontload a large part of the cost of using the lazy machinery. There is very little overhead at function call time as most of the overhead was spent at function creation (definiton) time.
We can convert normal python into lazy python with the run_lazy function which takes a string, the ‘name’, globals, and locals. This is like exec or eval for lazy python. This will mutate the provided globals and locals so that we can access the lazily evaluated code.
>>> code = """ print('not lazy') strict(print('lazy')) """ >>> run_lazy(code) lazy
This also uses the same bytecode manipulation as lazy_function so they will give the same results.
IPython cell and line magics
If you have IPython installed, you may use the cell and line magic machinery to write and evaluate lazy code. For example:
In : from lazy import strict In : %lazy 2 + 2 # line magic acts as an expression Out: 4 In : type(_2) Out: lazy._thunk.thunk In : %%lazy # cell magic is treated as a statement ...: print('lazy') ...: strict(print('strict')) ...: strict
At its core, lazy is just a way of converting expressions into a tree of deferred computation objects called thunks. thunks wrap normal functions by not evaluating them until the value is needed. A thunk wrapped function can accept thunks as arguments; this is how the tree is built. Some computations cannot be deferred because there is some state that is needed to construct the thunk, or the python standard defines the return of some method to be a specific type. These are refered to as ‘strict points’. Examples of strict points are str and bool because the python standard says that these functions must return an instance of their own type. Most of these converters are strict; however, some other things are strict because it solves recursion issues in the interpreter, like accessing __class__ on a thunk.
Custom Strictness Properties
strict is actually a type that cannot be put into a thunk. For example:
>>> type(thunk(strict, 2)) int
Notice that this is not a thunk, and has been strictly evaluated.
To create custom strict objects, you can subclass strict. This prevents the object from getting wrapped in thunks allowing you to create strict data structures.
Objects may also define a __strict__ method that defines how to strictly evalueate the object. For example, an object could be defined as:
class StrictFive(object): def __strict__(self): return 5
This would make strict(StrictFive()) return 5 instead of an instance of StrictFive.
undefined is a value that cannot be strictly evaluated. It is useful as a placeholder for computations.
We can imagine undefined in python as:
@thunk.fromexpr @Exception.__new__ class undefined(Exception): def __strict__(self): raise self
This object will raise an instance of itself when it is evaluated. This is presented as an equivalent definition, though it is actually in c to make nicer stack traces.
Currently, the following things are known to not work:
Recursively defined thunks
A recursively defined thunk is a thunk that appears in its own graph twice. For example:
>>> a = thunk(lambda: a) >>> strict(a)
This will cause an infinite loop because in order to strictly evaluate a, we will call the function which returns a which we will try to strictly evaluate.
Status: Bug, might fix.
This is basically correct, for example:
>>> a = lambda: a() >>> a() ... RuntimeError: maximum recursion depth exceeded
The difference in the thunk example is that we will drop into c code to preform the recursion so it will not terminate in a reasonable amount of time.
The potential fix could be to try to detect these cycles and raise some Exception; however, this might be a very expensive check in the good case making thunk evaluation much slower.
I opened it up in the repl, everything is strict!
Because the python spec says the __repr__ of an object must return a str, a call to repr must strictly evaluate the contents so that we can see what it is. The repl will implicitly call repr on things to display them. We can see that this is a thunk by doing:
>>> a = thunk(operator.add, 2, 3) >>> type(a) lazy.thunk.thunk >>> a 5
Again, because we need to compute something to represent it, the repl is a bad use case for this, and might make it appear at first like this is always strict.
print didn’t do anything!
Um, what did you think it would do?
If we write:
@lazy_function def f(a, b): print('printing the sum of %s and %s' % (a, b)) return a + b
Then there is no reason that the print call should be executed. No computation depends on the results, so it is casually skipped.
The solution is to force a dependency:
@lazy_function def f(a, b): strict(print('printing the sum of %s and %s' % (a, b))) return a + b
strict is a function that is used to strictly evaluate things. Because the body of the function is interpreted as lazy python, the function call is converted into a thunk, and therefore we can strict it.
This is true for any side-effectful function call.
x is being evaluated strictly when I think it should be lazy
There are some cases where things MUST be strict based on the python language spec. Because this is not really a new language, just an automated way of writing really inefficient python, python’s rules must be followed.
For example, __bool__, __int__, and other converters expect that the return type must be a the proper type. This counts as a place where strictness is needed1.
This might not be the case though, instead, I might have missed something and you are correct, it should be lazy. If you think I missed something, open an issue and I will try to address it as soon as possible.
Some stateful thing is broken
Sorry, you are using unmanaged state and lazy evaluation, you deserve this. thunks cache the normal form so that calling strict the second time will refer to the cached value. If this depended on some stateful function, then it will not work as intended.
I tried to do x with a thunk and it broke!
The library is probably broken. This was written on a whim and I barely thought through the use cases.
Please open an issue and I will try to get back to you as soon as possible.
- The function call for the constructor will be made lazy in the LazyTransformer (like thunk(int, your_thunk)), so while this is a place where strictness is needed, it can still be ‘optimized’ away.