Assertion density

Assertion density

An assertion is an executable claim about what must be true at a point in the program, and its worth is proximity: it fails at the cause, not three stack frames later where the bad value finally breaks something. Scattering asserts at random adds noise; placing them where invariants actually live turns a class of debugging sessions into one clear failure. The skill is knowing those places.

Method

1. Assert preconditions at every public boundary. At the top of an exported function, check the arguments the caller must honor: assert n >= 0, non-null, index in range. A caller's mistake becomes a message at the door.
2. Establish class invariants in the constructor. After construction, assert the object sits in a legal state: assert self.balance >= 0, both ends of a range ordered. An object that cannot exist badly never spreads a bad state.
3. Guard the loop invariant inside the loop. Assert the property the loop maintains each pass: an index stays in bounds, a running sum stays non-negative, a heap stays ordered. The assert catches the pass that breaks it.
4. Check postconditions before returning. Ahead of return result, assert what you promised: the output is sorted, its length matches the input, no null slipped through. This defends the contract callers are about to rely on.
5. Assert the impossible branch. In an else or default you believe unreachable, write assert False, "unreachable: " + state rather than a silent pass. If reality disagrees, you learn at once and precisely.
6. Keep assertions cheap and side-effect free. Never assert queue.pop(): builds that strip asserts, like python -O or NDEBUG, would drop the pop too. Assert conditions; compute nothing the program itself depends on.

Litmus tests

• When this assert fires, does the message point at the cause or a symptom?
• Would stripping assertions in a release build change any behavior? It must not.
• Can you name the invariant each assert defends in a sentence? If not, cut it.

Boundaries

Assertions guard programmer errors, not user input or I/O failure: validate untrusted data with real error handling that survives in production, since asserts may compile out. Hot loops where an assert measurably costs throughput are a judgment call; some teams gate expensive checks behind a debug flag.