What is the difference between an Idempotent and a Deterministic function?
A deterministic function is just a function in the mathematical sense. Given the same input, you always get the same output. On the other hand, an idempotent function is a function which satisfies the identity
f(f(x)) = f(x)
As a simple example. If UCase()
is a function that converts a string to an upper case string, then clearly UCase(Ucase(s)) = UCase(s)
.
Idempotent functions are a subset of all functions.
In more simple terms:
- Pure deterministic function: The output is based entirely, and only, on the input values and nothing else: there is no other (hidden) input or state that it relies on to generate its output. There are no side-effects or other output.
- Impure deterministic function: As with a deterministic function that is a pure function: the output is based entirely, and only, on the input values and nothing else: there is no other (hidden) input or state that it relies on to generate its output - however there is other output (side-effects).
- Idempotency: The practical definition is that you can safely call the same function multiple times without fear of negative side-effects. More formally: there are no changes of state between subsequent identical calls.
Idempotency does not imply determinacy (as a function can alter state on the first call while being idempotent on subsequent calls), but all pure deterministic functions are inherently idempotent (as there is no internal state to persist between calls). Impure deterministic functions are not necessarily idempotent.
Pure deterministic | Impure deterministic | Pure Nondeterministic | Impure Nondeterministic | Idempotent | |
---|---|---|---|---|---|
Input | Only parameter arguments (incl. this ) |
Only parameter arguments (incl. this ) |
Parameter arguments and hidden state | Parameter arguments and hidden state | Any |
Output | Only return value | Return value or side-effects | Only return value | Return value or side-effects | Any |
Side-effects | None | Yes | None | Yes | After 1st call: Maybe. After 2nd call: None |
SQL Example | UCASE |
CREATE TABLE |
GETDATE |
DROP TABLE |
|
C# Example | String.IndexOf |
DateTime.Now |
Directory.Create(String) Footnote1 |
- Footnote1 -
Directory.Create(String)
is idempotent because if the directory already exists it doesn't raise an error, instead it returns a newDirectoryInfo
instance pointing to the specified extant filesystem directory (instead of creating the filesystem directory first and then returning a newDirectoryInfo
instance pointing to it) - this is just like how Win32'sCreateFile
can be used to open an existing file.
A temporary note on non-scalar parameters, this
, and mutating input arguments:
(I'm currently unsure how instance methods in OOP languages (with their hidden this
parameter) can be categorized as pure/impure or deterministic or not - especially when it comes to mutating the the target of this
- so I've asked the experts in CS.SE to help me come to an answer - once I've got a satisfactory answer there I'll update this answer).
A note on Exceptions
Many (most?) programming languages today treat thrown exceptions as either a separate "kind" of return (i.e. "return
to nearest catch
") or as an explicit side-effect (often due to how that language's runtime works). However, as far as this answer is concerned, a given function's ability to throw an exception does not alter its pure/impure/deterministic/non-deterministic label - ditto idempotency (in fact: throwing is often how idempotency is implemented in the first place e.g. a function can avoid causing any side-effects simply by throwing right-before it makes those state changes - but alternatively it could simply return
too.).
So, for our CS-theoretical purposes, if a given function can throw an exception then you can consider the exception as simply part of that function's output. What does matter is if the exception is thrown deterministically or not, and if (e.g. List<T>.get(int index)
deterministically throws if index < 0
).
Note that things are very different for functions that catch exceptions, however.
Determinacy of Pure Functions
For example, in SQL UCASE(val)
, or in C#/.NET String.IndexOf
are both deterministic because the output depends only on the input. Note that in instance methods (such as IndexOf
) the instance object (i.e. the hidden this
parameter) counts as input, even though it's "hidden":
"foo".IndexOf("o") == 1 // first cal
"foo".IndexOf("o") == 1 // second call
// the third call will also be == 1
Whereas in SQL NOW()
or in C#/.NET DateTime.UtcNow
is not deterministic because the output changes even though the input remains the same (note that property getters in .NET are equivalent to a method that accepts no parameters besides the implicit this
parameter):
DateTime.UtcNow == 2016-10-27 18:10:01 // first call
DateTime.UtcNow == 2016-10-27 18:10:02 // second call
Idempotency
A good example in .NET is the Dispose()
method: See Should IDisposable.Dispose() implementations be idempotent?
a Dispose method should be callable multiple times without throwing an exception.
So if a parent component X
makes an initial call to foo.Dispose()
then it will invoke the disposal operation and X
can now consider foo
to be disposed. Execution/control then passes to another component Y
which also then tries to dispose of foo
, after Y
calls foo.Dispose()
it too can expect foo
to be disposed (which it is), even though X
already disposed it. This means Y
does not need to check to see if foo
is already disposed, saving the developer time - and also eliminating bugs where calling Dispose
a second time might throw an exception, for example.
Another (general) example is in REST: the RFC for HTTP1.1 states that GET
, HEAD
, PUT
, and DELETE
are idempotent, but POST
is not ( https://www.w3.org/Protocols/rfc2616/rfc2616-sec9.html )
Methods can also have the property of "idempotence" in that (aside from error or expiration issues) the side-effects of N > 0 identical requests is the same as for a single request. The methods GET, HEAD, PUT and DELETE share this property. Also, the methods OPTIONS and TRACE SHOULD NOT have side effects, and so are inherently idempotent.
So if you use DELETE
then:
Client->Server: DELETE /foo/bar
// `foo/bar` is now deleted
Server->Client: 200 OK
Client->Server DELETE /foo/bar
// foo/bar` is already deleted, so there's nothing to do, but inform the client that foo/bar doesn't exist
Server->Client: 404 Not Found
// the client asks again:
Client->Server: DELETE /foo/bar
// foo/bar` is already deleted, so there's nothing to do, but inform the client that foo/bar doesn't exist
Server->Client: 404 Not Found
So you see in the above example that DELETE
is idempotent in that the state of the server did not change between the last two DELETE
requests, but it is not deterministic because the server returned 200
for the first request but 404
for the second request.
A deterministic function will return the same result for the same inputs, regardless of how many times you call it.
An idempotent function may NOT return the same result (it will return the result in the same form but the value could be different, see http example below). It only guarantees that it will have no side effects. In other words it will not change anything.
For example, the GET
verb is meant to be idempotent in HTTP protocol. If you call "~/employees/1" it will return the info for employee with ID of 1 in a specific format. It should never change anything but simply return the employee information. If you call it 10, 100 or so times, the returned format will always be the same. However, by no means can it be deterministic. Maybe if you call it the second time, the employee info has changed or perhaps the employee no longer even exists. But never should it have side effects or return the result in a different format.
My Opinion
Idempotent is a weird word but knowing the origin can be very helpful, idem meaning same and potent meaning power. In other words it means having the same power which clearly doesn't mean no side effects so not sure where that comes from. A classic example of There are only two hard things in computer science, cache invalidation and naming things. Why couldn't they just use read-only? Oh wait, they wanted to sound extra smart, perhaps? Perhaps like cyclomatic complexity?