#arena-allocator #tree #arena #ecs #generation


A safe tree using an arena allocator that allows deletion without suffering from the ABA problem by using generational indices

3 releases

0.1.2 Nov 30, 2018
0.1.1 Nov 29, 2018
0.1.0 Nov 29, 2018

#362 in Memory management

46 downloads per month

MPL-2.0 license

449 lines


A safe tree using an arena allocator that allows deletion without suffering from the ABA problem by using generational indices.

It uses generational-arena under the hood, made by fitzgen, special thanks to him.

generational-arena is itself inspired by Catherine West's closing keynote at RustConf 2018, where these ideas were presented in the context of an Entity-Component-System for games programming.

What? Why?

When you are working with a tree and you want to add and delete individual nodes at a time, or you are writing a game and its world consists of many inter-referencing objects with dynamic lifetimes that depend on user input. These are situations where matching Rust's ownership and lifetime rules can get tricky.

It doesn't make sense to use shared ownership with interior mutability (ie Rc<RefCell<T>> or Arc<Mutex<T>>) nor borrowed references (ie &'a T or &'a mut T) for structures. The cycles rule out reference counted types, and the required shared mutability rules out borrows. Furthermore, lifetimes are dynamic and don't follow the borrowed-data-outlives-the-borrower discipline.

In these situations, it is tempting to store objects in a Vec<T> and have them reference each other via their indices. No more borrow checker or ownership problems! Often, this solution is good enough.

However, now we can't delete individual items from that Vec<T> when we no longer need them, because we end up either

  • messing up the indices of every element that follows the deleted one, or

  • suffering from the ABA problem. To elaborate further, if we tried to replace the Vec<T> with a Vec<Option<T>>, and delete an element by setting it to None, then we create the possibility for this buggy sequence:

    • obj1 references obj2 at index i

    • someone else deletes obj2 from index i, setting that element to None

    • a third thing allocates obj3, which ends up at index i, because the element at that index is None and therefore available for allocation

    • obj1 attempts to get obj2 at index i, but incorrectly is given obj3, when instead the get should fail.

By introducing a monotonically increasing generation counter to the collection, associating each element in the collection with the generation when it was inserted, and getting elements from the collection with the pair of index and the generation at the time when the element was inserted, then we can solve the aforementioned ABA problem. When indexing into the collection, if the index pair's generation does not match the generation of the element at that index, then the operation fails.


  • Zero unsafe
  • There is different iterators to traverse the tree
  • Well tested


First, add vec-tree to your Cargo.toml:

vec-tree = "0.1"

Then, import the crate and use the vec-tree::Tree

extern crate vec_tree;
use vec_tree::VecTree;

let mut tree = VecTree::new();

// Insert some elements into the tree.
let root_node = tree.insert_root(1);
let child_node_1 = tree.insert(10, root_node);
let child_node_2 = tree.insert(11, root_node);
let child_node_3 = tree.insert(12, root_node);
let grandchild = tree.insert(100, child_node_3);

// Inserted elements can be accessed infallibly via indexing (and missing
// entries will panic).
assert_eq!(tree[child_node_1], 10);

// Alternatively, the `get` and `get_mut` methods provide fallible lookup.
if let Some(node_value) = tree.get(child_node_2) {
    println!("The node value is: {}", node_value);
if let Some(node_value) = tree.get_mut(grandchild) {
    *node_value = 101;

// We can remove elements.

// Insert a new one.
let child_node_4 = tree.insert(13, root_node);

// The tree does not contain `child_node_3` anymore, but it does contain
// `child_node_4`, even though they are almost certainly at the same index
// within the arena of the tree in practice. Ambiguities are resolved with
// an associated generation tag.

// We can also move a node (and its descendants).
tree.append_child(child_node_1, child_node_4);

// Iterate over the children of a node.
for value in tree.children(child_node_1) {
    println!("value: {:?}", value);

// Or all the descendants in depth first search order.
let descendants = tree
    .map(|node| tree[node])

assert_eq!(descendants, [1, 10, 13, 11]);