Invariants via immutability

Clojure’s emphasis on immutable bindings and data structures lead us to write simpler, more obvious code. We don’t need to worry about immutable things changing over time. This reduces the cognitive load of both writing and reading code.

It is common to assume invariants for mutable structures, however, even in Clojure. Sometimes refactoring such logic in terms of immutable things can be clearer, and also help verification tools like Typed Clojure infer and check interesting invariants.

Local bindings

Local bindings in Clojure are immutable. Furthermore, local bindings often point to data structures which are themselves immutable.

This combination enables Typed Clojure to infer invariants and are both uncomplicated and inexpensive to compute.

For example, if a collection implements clojure.lang.IPersistentCollection (ie. it is immutable), Typed Clojure can track its count.

In this snippet, a local a of type (Coll Number) is in scope. seq returns a true value if its argument is non-empty. Typed Clojure remembers this down the “then” branch, because a is immutable.

(if (seq a)
  ; `a` is (NonEmptyColl Number)
  ; `a` is (Coll Number)

Note: Coll and NonEmptyColl are type aliases defined in clojure.core.typed.

This also works sequentially with assertions.

  (assert (seq a))
  ; `a` is (NonEmptyColl Number)

The program will fail at runtime if the expected invariant is violated. Because of this, Typed Clojure is safe to assume the invariant on our immutable binding for the rest of its scope.

Dynamic vars

Vars in Clojure are mutable. Typed Clojure does not attempt to track any interesting invariants about Vars as programs progress, but, as we will see, we can fall back on the obvious safety of immutable local bindings.

In this example, *atom-or-nil* has type (U nil (Atom1 Number)).

This code results in a type error — Typed Clojure does not refine the type of mutable bindings, so it cannot rule out a null pointer exception. (Typed Clojure guarantees typed code cannot throw null pointer exceptions).

(when *atom-or-nil*
  (swap! *atom-or-nil* inc))

We can convince Typed Clojure that a null pointer exception is impossible by using an intermediate local variable. See the definition of inc-dynamic below.

(ns blog.immutable.dynamic
  (:require [clojure.core.typed 
             :refer [Atom1 Int check-ns ann]]))

(ann *atom-or-nil* (U nil (Atom1 Int)))
(def ^:dynamic *atom-or-nil* nil)

(ann inc-dynamic [-> Int])
(defn inc-dynamic []
  (if-let [a *atom-or-nil*]
    (swap! a inc)

; (check-ns)
; => :ok

inc-dynamic enjoys the strong invariants provided by immutable bindings, at the cost of a local binding. It is also obviously clear to the reader of such code that it is safe from null pointer exceptions. There is no question about mutating a.

Note: For presentational purposes, we return 0 if *atom-or-nil* is nil. See when-let-fail, which throws an exception if the binding is a false value.


Immutability reduces the cognitive load of programming. It also helps analysis tools verify invariants about your code without particularly sophisticated techniques.

Refactoring code to take advantage of immutability often results in clearer, more obviously correct code.

Your readers will thank you for it.

Read more

Occurrence Typing by @samth and Matthias Felleisen. This inference technique helps Typed Clojure infer better types at branches and assertions.


See the code for this blog.


16 Aug 2013