What is the Liskov Substitution Principle (the "L" in SOLID), and what does it require of a subtype?

## The principle The Liskov Substitution Principle (LSP) is the **L** in SOLID. It comes from Barbara Liskov and Jeannette Wing's work on *behavioural subtyping*. The usual phrasing: > If S is a subtype of T, then objects of type T may be replaced with objects of type S without altering any of the desirable properties of the program (correctness, task performed, etc.). Plainly: **any code written against the general type must still be correct when it is secretly handed a specialised type.** ### Terms defined - **Supertype / base type**: the general type — an interface, abstract class, protocol, or trait — that client code is written against (e.g. `Account`, `Shape`, `Collection`). - **Subtype**: a type declared to be usable in place of the supertype — via class inheritance, interface implementation, or structural conformance. - **Client code**: any code that holds a reference typed as the supertype and calls methods on it. Crucially, the client was written and tested **knowing only the supertype's documented behaviour**. - **Substitutability**: the property that swapping in a subtype instance changes nothing the client can observe as a *violation of the contract* (it may change performance or which concrete algorithm runs — that's fine). ### Why signatures are not enough Most languages check only the *syntactic* part: does the subtype have a method with a compatible name and parameter/return types? A compiler will happily accept a subtype that is behaviourally nonsense: - `Square extends Rectangle` where setting the width silently also changes the height — compiles, breaks any client that sets width and height independently. - `ImmutableList implements List` where `add()` throws `UnsupportedOperationException` — compiles, breaks any client that was told `List` supports `add`. - `FastCache extends Cache` where `get()` requires `connect()` to have been called first — compiles, breaks clients that never heard of `connect()`. LSP is about the **contract**, the behaviour promised in documentation, tests, and the type's semantics — not the method table. ### The three rules (Design by Contract framing) Bertrand Meyer's Design by Contract gives the operational checklist. For each overridden operation: 1. **Preconditions may not be strengthened.** A precondition is what the caller must guarantee before calling. If `Base.withdraw(amount)` accepts any positive amount, a subtype that additionally demands `amount <= 100` will reject calls the client was told were legal. Subtypes may *weaken* (accept more) — that never breaks an existing caller. 2. **Postconditions may not be weakened.** A postcondition is what the method guarantees on return. If `Base.find(id)` promises a non-null result or an exception, a subtype returning `null` breaks clients. Subtypes may *strengthen* (promise more) — e.g. also promise the result is sorted. 3. **Invariants must be preserved.** An invariant is a property true of the object before and after every public operation (`balance >= 0`; `width` and `height` are independent). A subtype may add invariants of its own, but may not break the base's. Liskov and Wing add a **history rule**: the subtype must not allow state changes the supertype's clients believe impossible — e.g. an "immutable point" base type subclassed by a mutable point. A memorable summary: **be liberal in what you accept, conservative in what you promise.** ### Signature variance (the type-system half) Languages encode part of the rule in their variance rules for overriding: - **Return types are covariant**: an override may return a *narrower* type (`Animal` → `Dog`). That is postcondition strengthening — safe. - **Parameter types should be contravariant**: an override could safely accept a *wider* type (`Dog` → `Animal`). That is precondition weakening — safe. Most mainstream OO languages (Java, C#, Kotlin, C++) don't actually support contravariant parameters; they require invariant parameter types, and a differing parameter type just creates an overload, not an override. - **Exceptions are covariant**: an override may throw fewer or narrower checked exceptions, never new broader ones — throwing something the client was never told about is postcondition weakening. Variance handles only the parts the compiler can see. Preconditions/postconditions/invariants expressed in *values* ("amount must be positive", "never returns empty") remain the developer's responsibility. ### How violations show up in real code - `throw new UnsupportedOperationException()` in an override. - The client doing `instanceof` / `is` / `typeof` checks and branching per subtype. - Base-class tests that fail when re-run against a subclass. - Extra required call order ("you must call `init()` first on this one"). - Configuration flags on the base whose meaning changes per subclass. ### How to fix - **Extract the true common abstraction**: if only some shapes can resize independently, `Shape` shouldn't have `setWidth`/`setHeight`; put them on `Rectangle` only. - **Prefer composition over inheritance**: `Square` *has a* side length and can expose `area()`, without pretending to be a `Rectangle`. - **Split the interface** (Interface Segregation): `ReadableList` vs `MutableList`, so immutability isn't a broken promise. - **Make objects immutable**: with no setters, the Rectangle/Square conflict evaporates because there is no mutation to be inconsistent about. - **Verify with contract tests**: one test suite written against the supertype's contract, executed against every implementation. ### Trade-offs and nuance - LSP is about *observable contract*, not perfect equivalence. Different performance, different logging, different persistence backend — all fine. - The contract is whatever clients were *told*. If the base type documents "implementations may reject amounts over their configured limit", a limiting subtype is compliant; the precondition was always that broad. Much of "LSP compliance" is therefore a documentation decision made up front. - Strict LSP is expensive to enforce in dynamic or plugin ecosystems; contract tests are the pragmatic substitute for formal proofs.