In-process Event Bus · Analysis

01 · Frame the problem

Understand the central challenge

Design a single-process publish/subscribe library. Listeners can unsubscribe or add subscriptions while an event is being handled. Slow or throwing listeners need documented behavior.

Publication must define which subscriptions belong to an event and what happens when listeners throw or mutate the registry. Reentrancy is a semantic choice, not just an iteration implementation detail.

02 · Find the design pressure

Start simple, then identify the gap

Begin with tokenized subscriptions and a stable publish snapshot. Iterating the live registry lets callbacks unpredictably skip or add recipients. Preserve snapshot order and isolate failures so one listener or error reporter cannot corrupt dispatch.

Use a token-indexed subscription registry and a dispatcher that iterates a stable snapshot in registration order. Isolate listener failures and define synchronous, reentrant and optional asynchronous behavior explicitly. Subscribe validates the topic/callback and stores a new token with a monotonically increasing registration order. Duplicate callback subscriptions remain separate registrations. Publish captures the topic’s subscriptions before invoking user code. Treat event payloads as immutable supported values; snapshot or freeze them according to the library’s explicit payload contract. Invoke snapshot listeners in registration order. Catch each listener error, collect it in PublishResult and report it through a guarded error handler. Error-handler failures are collected without recursive event publication. Unsubscribe removes exactly one token and is idempotent. Enforce a nesting-depth limit for synchronous reentrant publish. If async delivery is added, bound its queue and return an overload result when full.

03 · Define success

Requirements to preserve

Return stable subscription tokens and support idempotent unsubscribe.
Define which subscribers receive an in-flight publication.
Isolate listener errors and specify dispatch order.
Bound asynchronous dispatch if an async adapter is provided.

04 · Identify the correctness boundary

Invariants, representation and mutation boundaries

Snapshot subscriptions before dispatch so subscribe/unsubscribe during a callback has predictable scope. Preserve registration order within that snapshot and isolate listener failures through a protected error reporter. Reentrant publish takes its own snapshot; it must not corrupt the outer iteration or reuse mutable payload state.

05 · Make it concrete

Reason through a worked example

Listeners A and B are subscribed. A unsubscribes B while handling event 1; B still receives event 1 under snapshot semantics, but not event 2. If A throws, B still runs. If the error reporter also throws, the result records that failure rather than recursively redispatching errors forever.