Real-World ScenariosTopic 11 of 18

ProblemHardAsked Often

In-Process Pub-Sub Bus

In one line

An in-memory pub-sub: publishers send to topics, subscribers receive matching messages. Decisions: per-subscriber bounded queue (slow consumer doesn't block fast); ordered or fan-out; sync or async delivery. Trades latency for backpressure.

The problem

Build an in-memory pub-sub: publishers send messages to topics; subscribers register for topics and receive matching messages. Used in: Spring's ApplicationEventPublisher, GUI event buses, in-process plugin systems.

Constraints:

Many publishers, many subscribers.
Slow subscriber must not block publisher.
Dynamic subscribe/unsubscribe.
Thread-safe under concurrent publish + subscribe.

Note

What this tests Composing ConcurrentHashMap (topics → subscribers) + bounded queues (per-subscriber buffering) + worker threads (subscriber drain) + atomic counters (drop tracking). Four primitives, one design.

The architectural decision

Tip

Per-subscriber queue pattern Don't make publishers iterate-and-deliver synchronously, slow handler blocks publisher.

Instead: each subscriber owns a queue + worker thread. Publish = enqueue to each subscriber's queue. Subscriber's worker drains and runs the handler. Slow handler only delays its own queue.

The queue capacity is the tuning knob: small (drop fast on backpressure) vs large (absorb bursts but use memory).

Slow-consumer policy

This is the most contested design choice. Three options:

Policy	Semantics	Use when
Drop on full	Publisher succeeds; slow consumer loses events	Telemetry, metrics, optional notifications
Block on full	Publisher waits; backpressure flows up	Orders, financial events, anything that can't be lost
Block with timeout	Publisher waits N ms, then drops	Most production buses, middle ground

Warning

What kills naive implementations "Synchronous handler dispatch", publisher calls each subscriber inline. One slow handler delays every subscriber + the publisher. This is the #1 production-incident shape: a logging subscriber stalled the order-processing publisher.

When NOT to roll a custom one

For most use cases:

Spring ApplicationEventPublisher, Spring-aware.
Guava EventBus, simple, in-process.
Reactor / RxJava, full reactive streams with backpressure.
Kafka / NATS / Redis Streams, when durability or cross-service delivery is required.

Roll a custom bus only after measuring the alternatives and finding they don't fit. Most "in-process pub-sub" interview answers should cite the right library and discuss its tradeoffs, not build from scratch.

Implementations

ConcurrentHashMap + per-subscriber bounded queue

Each subscriber owns a dedicated thread that drains its own queue. Publisher iterates topic's subscribers and offers to each queue. Drop-on-full prevents slow-consumer backpressure.

 1  import java.util.concurrent.*;
 2  import java.util.concurrent.atomic.AtomicLong;
 3  
 4  class EventBus<T> {
 5      private final ConcurrentHashMap<String, List<Subscriber<T>>> topics = new ConcurrentHashMap<>();
 6      private final AtomicLong dropped = new AtomicLong();
 7  
 8      public Subscription subscribe(String topic, Consumer<T> handler) {
 9          Subscriber<T> sub = new Subscriber<>(handler);
10          topics.computeIfAbsent(topic, k -> new CopyOnWriteArrayList<>()).add(sub);
11          return () -> topics.get(topic).remove(sub);
12      }
13  
14      public void publish(String topic, T event) {
15          List<Subscriber<T>> subs = topics.get(topic);
16          if (subs == null) return;
17          for (Subscriber<T> sub : subs) {
18              if (!sub.queue.offer(event)) dropped.incrementAndGet();
19          }
20      }
21  
22      static class Subscriber<T> {
23          final BlockingQueue<T> queue = new ArrayBlockingQueue<>(1024);
24          final Thread worker;
25  
26          Subscriber(Consumer<T> handler) {
27              worker = new Thread(() -> {
28                  while (!Thread.currentThread().isInterrupted()) {
29                      try { handler.accept(queue.take()); }
30                      catch (InterruptedException e) { return; }
31                      catch (Exception e) { /* log; don't kill the worker */ }
32                  }
33              });
34              worker.setDaemon(true);
35              worker.start();
36          }
37      }
38  
39      interface Subscription { void unsubscribe(); }
40  }

Key points

•Topics → subscriber list; publish iterates and delivers (or queues per subscriber)
•Each subscriber has bounded queue → slow consumer doesn't backpressure publisher
•Drop-on-full vs block-on-full, slow-consumer policy is THE design decision
•Concurrent map for topic→subscribers (CHM, sync.Map, dict + RLock)
•Used by Spring Events, Akka Streams, in-process EventBus libraries

Follow-up questions

▸Why per-subscriber queues instead of one global queue?

Per-subscriber queues isolate slow consumers, one slow handler doesn't block others. Single global queue: every subscriber processes serially → throughput limited by slowest. Per-subscriber: each runs at its own pace, can be drop-on-full independently.

▸Drop-on-full vs block-on-full?

Drop-on-full preserves publisher throughput (publisher never blocks). Block-on-full applies backpressure, publisher slows when consumer can't keep up. Drop is right for telemetry/metrics; block is right for guaranteed delivery (orders, billing).

▸How are subscriber failures handled?

Each worker wraps the handler in try/catch, exceptions log and continue. Otherwise a thrown exception kills the worker thread, leaving the subscriber silently dead. Always isolate handler bugs from the worker loop.

▸How does this differ from Kafka?

Kafka has persistent log; messages survive restart; consumers track offsets. In-process bus is ephemeral, restart loses messages. Kafka is for cross-service durable pub-sub; in-process bus is for in-app event-driven decoupling.

In-Process Pub-Sub Bus

In one line

The problem

Constraints:

Many publishers, many subscribers.
Slow subscriber must not block publisher.
Dynamic subscribe/unsubscribe.
Thread-safe under concurrent publish + subscribe.

Note

The architectural decision

Tip

Per-subscriber queue pattern Don't make publishers iterate-and-deliver synchronously, slow handler blocks publisher.

Instead: each subscriber owns a queue + worker thread. Publish = enqueue to each subscriber's queue. Subscriber's worker drains and runs the handler. Slow handler only delays its own queue.

The queue capacity is the tuning knob: small (drop fast on backpressure) vs large (absorb bursts but use memory).

Slow-consumer policy

This is the most contested design choice. Three options:

Policy	Semantics	Use when
Drop on full	Publisher succeeds; slow consumer loses events	Telemetry, metrics, optional notifications
Block on full	Publisher waits; backpressure flows up	Orders, financial events, anything that can't be lost
Block with timeout	Publisher waits N ms, then drops	Most production buses, middle ground

Warning

When NOT to roll a custom one

For most use cases:

Spring ApplicationEventPublisher, Spring-aware.
Guava EventBus, simple, in-process.
Reactor / RxJava, full reactive streams with backpressure.
Kafka / NATS / Redis Streams, when durability or cross-service delivery is required.

Implementations

ConcurrentHashMap + per-subscriber bounded queue

Each subscriber owns a dedicated thread that drains its own queue. Publisher iterates topic's subscribers and offers to each queue. Drop-on-full prevents slow-consumer backpressure.

 1  import java.util.concurrent.*;
 2  import java.util.concurrent.atomic.AtomicLong;
 3  
 4  class EventBus<T> {
 5      private final ConcurrentHashMap<String, List<Subscriber<T>>> topics = new ConcurrentHashMap<>();
 6      private final AtomicLong dropped = new AtomicLong();
 7  
 8      public Subscription subscribe(String topic, Consumer<T> handler) {
 9          Subscriber<T> sub = new Subscriber<>(handler);
10          topics.computeIfAbsent(topic, k -> new CopyOnWriteArrayList<>()).add(sub);
11          return () -> topics.get(topic).remove(sub);
12      }
13  
14      public void publish(String topic, T event) {
15          List<Subscriber<T>> subs = topics.get(topic);
16          if (subs == null) return;
17          for (Subscriber<T> sub : subs) {
18              if (!sub.queue.offer(event)) dropped.incrementAndGet();
19          }
20      }
21  
22      static class Subscriber<T> {
23          final BlockingQueue<T> queue = new ArrayBlockingQueue<>(1024);
24          final Thread worker;
25  
26          Subscriber(Consumer<T> handler) {
27              worker = new Thread(() -> {
28                  while (!Thread.currentThread().isInterrupted()) {
29                      try { handler.accept(queue.take()); }
30                      catch (InterruptedException e) { return; }
31                      catch (Exception e) { /* log; don't kill the worker */ }
32                  }
33              });
34              worker.setDaemon(true);
35              worker.start();
36          }
37      }
38  
39      interface Subscription { void unsubscribe(); }
40  }

Key points

•Topics → subscriber list; publish iterates and delivers (or queues per subscriber)
•Each subscriber has bounded queue → slow consumer doesn't backpressure publisher
•Drop-on-full vs block-on-full, slow-consumer policy is THE design decision
•Concurrent map for topic→subscribers (CHM, sync.Map, dict + RLock)
•Used by Spring Events, Akka Streams, in-process EventBus libraries

Follow-up questions

▸Why per-subscriber queues instead of one global queue?

▸Drop-on-full vs block-on-full?

▸How are subscriber failures handled?

▸How does this differ from Kafka?

In-Process Pub-Sub Bus

The problem

The architectural decision

Slow-consumer policy

When NOT to roll a custom one

Implementations

Key points

Follow-up questions

Related reading

In-Process Pub-Sub Bus

The problem

The architectural decision

Slow-consumer policy

When NOT to roll a custom one

Implementations

Key points

Follow-up questions

Related reading