The canonical BFS keeps a Set of visited nodes and a queue of frontier nodes. Marking a node visited at the moment it enters the queue (not when it dequeues) is the easy mistake to avoid: forgetting to mark on enqueue can re-enqueue the same node many times and blow up runtime. Array.shift() is O(n) for large arrays; for big graphs use a real queue or an index-pointer trick (see next accordion). On small graphs the simplicity wins.

Some problems care which 'level' (distance from start) each node belongs to: level-order tree traversal, computing the shortest distance to every reachable node, finding all leaves at the same depth. Building the next frontier from the current one in a single pass groups nodes by level naturally, without storing per-node depths. The trade-off is two arrays per level instead of a queue, but the level-grouped output is exactly the shape the caller usually wants.

Multi-source BFS seeds the queue with every starting node at once, then runs a normal BFS. Because all sources are at distance 0, the level-by-level frontier still respects shortest-path semantics. This is the trick that solves 'rotting oranges' (every rotten orange spreads at the same rate), 'walls and gates' (every gate is distance 0), and most grid-distance problems. The pattern generalises beyond grids to any graph where multiple anchor nodes are equally close to themselves.