The Cost of Drop on Deep Ownership Chains

The Cost of Drop on Deep Ownership Chains

Thesis

Drop is not free. A long chain of Box nodes runs destructors one by one; mem::forget skips them and returns in nanoseconds — which is also how you leak. This chapter times dropping chains of 1 k / 10 k / 100 k nodes versus forgetting the large chain.

Mental model

  drop(head) → Drop::drop → drop(child) → …  (recursive / iterative cleanup)
  mem::forget(head) → no destructors; memory leaked unless into a global allocator you never reclaim

Worked examples

Verified with rustc 1.99.1, cargo run --release.

// chain of Box<Node>; time drop vs forget

Output:

drop chain of    1000 nodes: 21.888µs
drop chain of   10000 nodes: 267.357µs
drop chain of  100000 nodes: 4.211461ms
forget chain of  100000 nodes: 487ns

Drop scales with chain length (~4 ms for 100 k nodes here). forget on the same shape was ~0.5 µs — and leaked the graph.

Trap

Using forget to “speed up shutdown”. You trade CPU for leaks and skipped safety hooks (File flush, lock poisoning cleanup, etc.).

Boring rule

Let Drop run. If shutdown profiles show destructor cost, batch into arenas or reuse pools — do not forget application state.

Try this

1. Replace the linked list with a Vec<Node> and compare drop time.
2. Implement Drop for a type that prints, and watch ordering.
3. Read std::mem::ManuallyDrop.

Sources

• Drop in the Rust book / std::ops::Drop
• mem::forget