Output Slot: How null->fn->Set Saves 90% Memory

Chronicle 14 - Arc 5: Architecture v4 - Performance Without Compromise

The easiest way to waste memory in a reactive library is to allocate collections for nodes that never need them.

That was our v3 reality. Every producer created an output Set up front. It was simple and predictable, but most nodes had 0 or 1 downstream consumer for most of their lifetime. We were paying for the general case on every instance.

In v4, we replaced eager allocation with an output slot state machine:

• null when a node has no subscribers
• fn when exactly one subscriber is attached
• Set<fn> when fan-out grows beyond one

Why this matters in real graphs

Large app graphs are sparse. Even with many derived nodes, most edges are linear chains or temporary subscriptions.

When every node starts with new Set(), you allocate:

• an object for the Set
• internal hash storage
• bookkeeping for iteration and growth

Multiplied across thousands of nodes, this dominated baseline memory before useful work even started.

The slot transition model

The output slot behaves like a tiny adaptive container:

1. Subscribe first sink: null -> fn
2. Subscribe second sink: fn -> Set([fn1, fn2])
3. Unsubscribe to one sink: Set -> fn
4. Unsubscribe last sink: fn -> null

This keeps the hot path branchy but cheap, and avoids object churn in the common 0/1 subscriber case.

Performance trade-off we accepted

We intentionally traded a little code complexity for much lower resident memory:

• More branch checks in subscribe/unsubscribe
• More careful handling of iteration when the slot can change shape
• Fewer allocations and fewer GC-triggering objects

In practice, the branch overhead was noise compared with allocation savings. The graph spends far more time existing than changing topology.

Correctness hazards we had to avoid

Adaptive containers are easy to get wrong. The dangerous bugs were:

• Mutation while iterating: removing a sink during dispatch can skip callbacks
• Shape drift: forgetting to downshift Set -> fn leaks structure and loses benefits
• Duplicate fan-out: accidental double-subscribe corrupts cardinality assumptions

The fix was to centralize transitions and keep all add/remove operations in one path instead of inline branching in many callers.

The architectural lesson

v4 was not "micro-optimizations week." It was about aligning data structures with observed graph shape.

The output slot model works because it mirrors reality:

• most nodes are disconnected or singly connected
• high fan-out exists, but is exceptional

If your graph is sparse, your core storage should be sparse-first too.

That single change unlocked a lot of v4's memory profile improvements without changing user APIs at all.