SNS & Pub/Sub

SNS & pub/sub is the fan-out pattern — one event published to a topic (L247), many subscribers receive it (L247): the queues (L245), the functions (L266), the endpoints (L220) — the decoupled broadcast behind the event-driven AI platform (L248), with the filtering (L247) deciding which subscriber gets which event (L247).

The one-sentence interview answer: "Pub/sub is the fan-out (L247). The publisher — a service, a workflow (L217) — publishes an event to a topic (L247): 'job completed', 'error raised', 'model updated' (L248). The subscribers — the queues (L245), the Lambda functions (L266), the webhooks (L220), the email — each subscribed to the topic (L247), each receiving a copy of the event (L247). SNS is the managed topic service (L247): the topic, the subscriptions, the fan-out, and the filtering (L247) — the subscriber's filter decides which events it gets (L247). The value: the decoupling (L248) — the publisher doesn't know the subscribers (L247), and new subscribers join without changing the publisher (L248). The AI platform's events (L248) — the jobs, the errors, the changes — flow through the topics (L247), the nervous system of the platform (L260)."

Think of SNS as the building's PA system (public address). The announcements (the events, L247) are made over the PA (the topic, L247): "the job is done", "the error occurred" (L248). The people (the subscribers, L247) hear the announcements relevant to them: the operations team hears the errors (the alert, L208), the billing hears the jobs' completions (L332), the webhooks (L220) hear the changes. The PA doesn't know who's listening (the decoupling, L248) — anyone can listen in (the subscription, L247), and the listeners filter for what matters to them (L247). The building works because the announcements are broadcast, and each listener takes what it needs (L247).

   the PA system (SNS, L247)
   ┌────────────────────────────────────────────────────────┐
   │ the announcement (the event, L247)                     │
   │ the topic (L247) → the listeners (the subscribers,     │
   │ L247) — each takes what it needs (the filter, L247)    │
   │ the PA doesn't know who's listening (L248)             │
   └────────────────────────────────────────────────────────┘

The mental model is the PA system: the broadcast, the listeners, and each listener taking what it needs (L247).

   an event happens (L248)
        │
        ▼
   ┌──────────────────────────────────────────────────────────┐
   │ 1 · THE PUBLISH (L247)                                   │
   │     the service publishes to the topic (L247)            │
   └──────────────────┬───────────────────────────────────────┘
                      ▼
   ┌──────────────────────────────────────────────────────────┐
   │ 2 · THE FAN-OUT (L247)                                   │
   │     the topic delivers to every subscriber (L247)        │
   │     the queues (L245) · the functions (L266) · the       │
   │     webhooks (L220) · the email (L224)                   │
   └──────────────────┬───────────────────────────────────────┘
                      ▼
   ┌──────────────────────────────────────────────────────────┐
   │ 3 · THE FILTERS (L247)                                   │
   │     each subscriber's filter (L247) decides which        │
   │     events it receives (L247)                            │
   └──────────────────────────────────────────────────────────┘
                      ▼
   each subscriber processes its copy (L245, L266)

The flow is the fan-out: publish → fan out → filter → process — the PA's broadcast (L247).

• The topic (L247). The event's channel: the publisher publishes to the topic (L247), and the topic delivers to the subscribers (L247). The topic is the decoupling point (L248): the publisher doesn't know the subscribers (L247).
• The subscribers (L247). The consumers of the topic: the queues (L245), the Lambda functions (L266), the HTTP endpoints (the webhooks, L220), the email (L224). Each subscriber receives a copy of the event (L247).
• The fan-out (L247). The topic's delivery: one publish, many copies (L247) — the point-to-point queue's (L245) counterpart (L247).
• The filtering (L247). The subscriber's filter (L247): the subscription policy (L247) decides which events the subscriber receives (L247) — the operations team's topic subscription filters for the errors (L247).
• The AI shape (L248). The platform's events (L248): the job completed (L245), the error raised (L211), the model updated (L341), the tenant changed (L357) — published to the topics (L247), the nervous system of the platform (L260).

• The job completion (L245). The worker publishes "job completed" (L247) → the fan-out (L247) → the webhook (L220), the billing (L332), and the analytics (L328) each get a copy (L247).
• The error alert (L211). The service publishes "error" (L247) → the operations' subscription (L247) with the filter (L247) → the alert (L208).
• The model update (L341). The model registry publishes "model changed" (L247) → the cache invalidation (L244) and the config reload (L247).
• The tenant events (L357). The tenant's changes (L357) published (L247) → the per-tenant subscribers (L320).
• Anything decoupled (L260). The topics (L247) are the L260 platform's nervous system (L260) — the events flowing (L248).

The through-line: the PA system — the events broadcast, the subscribers filtering, the decoupling carrying the platform's signals (L247).

Say it in four moves:

1. The pattern. "One event, many subscribers (L247) — the fan-out (L247)."
2. The topic. "The publisher publishes; the topic delivers (L247)."
3. The filter. "Each subscriber's filter (L247) decides what it gets (L247)."
4. The value. "The decoupling (L248) — new subscribers join without the publisher's change (L247)."

• The topic is the seam (L248). The senior answer places the topics at the platform's seams (L248) — the services publish, the systems subscribe (L247) — the event-driven architecture (L248).
• The fan-out is the one-to-many (L247). The point-to-point queue (L245) is one-to-one; the topic (L247) is one-to-many (L247) — the pattern by the consumers' count (L247).
• The filter is the subscriber's contract (L247). The subscription policy (L247) — the subscriber receives only what it needs (L247) — the fan-out's efficiency (L247).
• The decoupling is the resilience (L248). The subscriber's failure doesn't affect the publisher (L248) — the queues (L245) buffer (L245), the functions (L266) retry (L169).
• The ordering is the topic's limit (L247). The fan-out is best-effort ordered (L247) — the strictly-ordered flows use the FIFO queue (L246), not the topic (L247).

• The queue for the fan-out (L245). The one-to-one for the one-to-many (L247) — each subscriber re-implementing the publish (L247).
• The topic for the point-to-point (L245). The one-to-many for a single consumer (L247) — the queue (L245) is simpler (L247).
• No filters (L247). Every subscriber receives everything (L247) — the fan-out's noise (L247).
• The publisher knows the subscribers (L248). The coupling (L248) — the decoupling point missed (L247).
• The ordering expected (L247). The best-effort order (L247) — the FIFO (L246) for the strict (L247).
• The event as the payload (L248). The full data in the event (L248) — the reference (L248) vs the event's size (L247).

• Place the topics at the seams (L248) — the services publish, the systems subscribe (L247).
• Filter the subscriptions (L247) — each subscriber receives what it needs (L247).
• Keep the publisher decoupled (L247) — it knows nothing about the subscribers (L248).
• Use the queues for the buffering (L245) — the subscribers' queues (L247) absorb the bursts (L222).
• Keep the events lean (L248) — the reference, not the payload (L247).
• Use the FIFO for the strict order (L246) — not the topic (L247).

Q: What is pub/sub?

A: The fan-out pattern (L247): one event published to a topic (L247), many subscribers receive it (L247) — the queues (L245), the functions (L266), the webhooks (L220). SNS is the managed topic service (L247): the topic, the subscriptions, the fan-out, and the filtering (L247). The value is the decoupling (L248) — the publisher doesn't know the subscribers (L247).

Q: Topic or queue?

A: The consumers' count decides (L247). The queue (L245) is point-to-point — one consumer (L245). The topic (L247) is one-to-many — the fan-out (L247): a job's completion reaching the webhook, the billing, and the analytics (L247). The queue for the one-to-one (L245); the topic for the one-to-many (L247).

Q: How do the filters work?

A: The subscription policy (L247). Each subscriber's subscription has a filter (L247) — the operations team's subscription filters for the errors, the billing's for the completions (L247). The subscriber receives only what its filter matches (L247) — the fan-out stays efficient (L247).

Q: What's the AI platform's use?

A: The events (L248): the job completed (L245), the error raised (L211), the model updated (L341). The services publish to the topics (L247); the downstream — the webhooks (L220), the alerts (L208), the analytics (L328) — subscribe with the filters (L247). The topics are the platform's nervous system (L260) — the events flowing between the decoupled parts (L248).

• What's the fan-out (L247)?
• Topic vs queue (L247)?
• How do the filters work (L247)?
• What's the decoupling's value (L248)?
• What are the platform's events (L248)?

The senior read: the consumer count is the choice — the queue for the one, the topic for the many (L247).

// SNS & pub/sub: the topic, the subscribers, the filter (L247).
import { SNSClient, PublishCommand } from '@aws-sdk/client-sns';

const sns = new SNSClient({ region: 'us-east-1' });

// THE PUBLISH (L247) — the publisher knows only the topic (L248).
await sns.send(new PublishCommand({
  TopicArn: process.env.JOB_EVENTS_TOPIC,
  Message: JSON.stringify({ type: 'job.completed', jobId, tenant, cost }),  // L248
  // the message attributes — the filter's inputs (L247):
  MessageAttributes: { type: { DataType: 'String', StringValue: 'job.completed' } },
}));

// THE SUBSCRIBERS (L247) — each with its filter (L247).
//   the webhook subscription:  filter policy { type: ["job.completed"] }   → L220
//   the billing subscription:  filter policy { type: ["job.completed"] }   → L332
//   the operations:            filter policy { type: ["error.raised"] }    → L208
//   the analytics:             filter policy { type: ["*"] }               → L328

// THE CONSUMER — the Lambda's queue (L247, L266).
//   the topic fans out to the queues (L245) → the Lambda (L266) processes.

What the reader must SEE — the PA system:

  PublishCommand to the topic → the publisher knows only the topic (L247, L248)
  MessageAttributes           → the filters' inputs (L247)
  the subscription policies   → each listener takes what it needs (L247)
  the queues → the Lambda     → the fan-out's consumers (L266)

  The broadcast, the listeners, each taking what it needs.

• The fan-out is the broadcast's cost (L150). One publish, many copies (L247) — the delivery cost scales with the subscribers (L247).
• The filters are the efficiency (L151). The subscription policies (L247) — the subscriber receives only what it needs (L247).
• The queues buffer the bursts (L245). The subscribers' queues (L245) — the Lambda (L266) and the workers absorb the fan-out's bursts (L222).
• The ordering is best-effort (L247). The topic's order (L247) — the FIFO queue (L246) for the strict (L247).

• Pub/sub = the fan-out (L247): one event, many subscribers.
• The topic: the decoupling point (L247) — the publisher knows only the topic (L248).
• The subscribers: the queues (L245), the functions (L266), the webhooks (L220).
• The filters: the subscription policies (L247).
• The ordering: best-effort (L247) — the FIFO (L246) for the strict (L247).
• The AI shape: the platform's events (L248) — the nervous system (L260).

SNS & PUB/SUB = the fan-out — the PA system

THE PATTERN (L247)
  the publisher publishes to the topic (L247)
  the topic delivers to every subscriber (L247)
  the subscribers: the queues (L245), the functions (L266),
  the webhooks (L220), the email (L224)

THE DECOUPLING (L247, L248)
  the publisher knows only the topic (L247)
  the new subscribers join without the publisher's change (L248)
  the subscriber's failure doesn't affect the publisher (L248)

THE FILTERS (L247)
  the subscription policies (L247)
  each subscriber receives what its filter matches (L247)

THE ORDERING (L247)
  best-effort on the topic (L247)
  the FIFO queue (L246) for the strictly-ordered flows (L247)

THE AI SHAPE (L248)
  the events: the job completed (L245), the error (L211),
  the model updated (L341)
  the topics = the platform's nervous system (L260)

INTERVIEW, 4 MOVES
  1 pattern  "one event, many subscribers (L247)"
  2 topic    "the decoupling point (L247, L248)"
  3 filters  "the subscription policies (L247)"
  4 AI shape "the platform's events (L248, L260)"

You now hold the fan-out: the topic that decouples, the subscribers that filter, and the events that flow between the platform's parts. The AI platform now has a nervous system — every signal broadcast, each listener taking what it needs (L247).

Next: the architecture the events enable — event-driven architecture (L248), events as the contract between AI and the business.