Signals vs messages: broadcast vs point-to-point

Motto — A message finds the one instance that owns the conversation; a signal reaches everyone who cares — confuse them and you either drop deliveries or stampede your engine.

Part of Phase 07 — Events, timers & messaging. Concept lesson — no code required.

The Problem

Two events land in your loan platform on the same afternoon. First, the bureau’s callback for application APP-002: exactly one instance must receive it, and delivering it twice or to a sibling would be a correctness bug. Second, RBI revises the repo rate: every in-flight application holding a floating-rate offer must recalculate. Same engine, same “something happened outside” shape, but opposite delivery semantics. BPMN gives each its own primitive. Models that use the wrong one work fine in the demo, then fail at the first duplicate or the first thousand-instance broadcast.

The Concept

	Message	Signal
Addressing	name + correlation key	name only
Receivers	exactly one	zero to N — all of them
Zero receivers means	a problem (late/mis-keyed) — surface it	nothing to do — normal
Carries payload	yes, into the one instance’s variables	engine-level: use sparingly; receivers usually re-read state
Scope	always engine-wide, keyed	engine-wide, or instance-scoped (a signal thrown and caught within one instance)
Typical use	bureau callback, e-sign webhook, payment confirmation	rate change, “compliance freeze all onboarding”, cache invalidation between branches

Two design rules fall out:

If you need a correlation key, it’s a message. The moment “which instance?” has an answer, broadcast is wrong. A signal named bureauCallback-APP-002 — a real anti-pattern found in real models — is a message wearing a signal costume, and it defeats duplicate detection.
If zero listeners is fine, it’s a signal. Messages must land; signals are fire-and-forget by design. Here is the operational warning too: a signal caught by 10,000 instances moves 10,000 tokens, on a synchronous throw, in one transaction. Broadcast to large populations needs async continuations (Phase 2) between the catch and any heavy work.

Both also come as event subprocess triggers (lesson 05) — “while this scope is active, react to X.” This is where signals shine: a complianceFreeze signal event subprocess in every onboarding process is one drawn element per model, not a boundary on every task.

Ship It

This lesson ships outputs/signal-message-cheatsheet.md — the decision table plus the model-review flags.

Check Yourself

Q1. “When the checker approves batch B-77, continue that batch’s process.” Message or signal?

A) signal — approvals are broadcasts
B) message — there’s a correlation key (B-77) and exactly one intended receiver
C) either works
D) timer

<details><summary>Answer</summary>B — “that batch’s” is a correlation key wearing prose. One intended receiver always means a message.</details>

Q2. A signal is thrown and no instance is waiting. What should happen?

A) an error — deliveries must land
B) nothing, by design — zero listeners is a normal outcome for broadcast
C) the signal is queued until someone subscribes
D) a new instance starts

<details><summary>Answer</summary>B — this is the semantic core of the split. If zero receivers worries you, you wanted a message — or a signal start event, which does create instances.</details>

Q3. The model review finds <signal name="paymentConfirmed-${orderId}"/>. The verdict?

A) fine — dynamic names are expressive
B) an anti-pattern: a keyed, single-receiver event is a message; the costume defeats correlation-error detection
C) fine if orders are unique
D) prefer a timer

<details><summary>Answer</summary>B — per-key signal names re-implement correlation without its safety rails: no ambiguity refusal, no unmatched-delivery surface.</details>

Challenge. Sweep a process model you own (or the capstone, next phase) and label every external interaction M or S using the two rules. Then find the one that’s neither — a conditional event or a polling loop — and write down why it resisted classification. That’s usually where an event registry (next lesson) belongs.

Next: The event registry
Previous: Message events

Signals vs messages: broadcast vs point-to-point

The Problem

The Concept

Ship It

Check Yourself

Related