JAEB - Just Another Event Bus

In-process, snapshot-driven event bus for Tokio applications.

JAEB provides:

• sync + async listeners via a unified subscribe API
• automatic dispatch-mode selection based on handler trait
• explicit listener unsubscription via Subscription handles or RAII SubscriptionGuard
• dependency injection via handler structs
• retry policies with configurable strategy for async handlers
• dead-letter stream for terminal failures
• explicit Result-based error handling
• graceful shutdown with in-flight task completion
• idempotent shutdown
• optional Prometheus-compatible metrics via the metrics crate
• structured tracing with per-handler spans
• summer-rs integration via summer-jaeb and #[event_listener] macro support summer-jaeb-macros

Installation

[dependencies]
jaeb = { version = "0.3.2" }
tokio = { version = "1", features = ["macros", "rt-multi-thread"] }

To enable metrics instrumentation:

[dependencies]
jaeb = { version = "0.3.2", features = ["metrics"] }

Quick Start

use std::time::Duration;

use jaeb::{
    DeadLetter, EventBus, EventBusError, EventHandler, FailurePolicy,
    HandlerResult, RetryStrategy, SyncEventHandler,
};

#[derive(Clone)]
struct OrderCheckoutEvent {
    order_id: i64,
}

struct AsyncCheckoutHandler;

impl EventHandler<OrderCheckoutEvent> for AsyncCheckoutHandler {
    async fn handle(&self, event: &OrderCheckoutEvent) -> HandlerResult {
        println!("async checkout {}", event.order_id);
        Ok(())
    }
}

struct SyncAuditHandler;

impl SyncEventHandler<OrderCheckoutEvent> for SyncAuditHandler {
    fn handle(&self, event: &OrderCheckoutEvent) -> HandlerResult {
        println!("sync audit {}", event.order_id);
        Ok(())
    }
}

struct DeadLetterLogger;

impl SyncEventHandler<DeadLetter> for DeadLetterLogger {
    fn handle(&self, dl: &DeadLetter) -> HandlerResult {
        eprintln!(
            "dead-letter: event={} listener={} attempts={} error={}",
            dl.event_name, dl.subscription_id, dl.attempts, dl.error
        );
        Ok(())
    }
}

#[tokio::main]
async fn main() -> Result<(), EventBusError> {
    let bus = EventBus::new(64)?;

    let retry_policy = FailurePolicy::default()
        .with_max_retries(2)
        .with_retry_strategy(RetryStrategy::Fixed(Duration::from_millis(50)));

    // Async handler -- dispatch mode inferred from EventHandler impl
    let checkout_sub = bus
        .subscribe_with_policy(AsyncCheckoutHandler, retry_policy)
        .await?;

    // Sync handler -- dispatch mode inferred from SyncEventHandler impl
    let _audit_sub = bus.subscribe(SyncAuditHandler).await?;

    // Dead-letter handler (convenience method, auto-sets dead_letter: false)
    bus.subscribe_dead_letters(DeadLetterLogger).await?;

    bus.publish(OrderCheckoutEvent { order_id: 42 }).await?;
    bus.try_publish(OrderCheckoutEvent { order_id: 43 })?;

    checkout_sub.unsubscribe().await?;
    bus.shutdown().await?;
    Ok(())
}

API Overview

EventBus

• EventBus::new(buffer) -> Result<EventBus, EventBusError> -- create a new bus with the given channel capacity
• EventBus::builder() -- builder for fine-grained configuration (buffer size, timeouts, concurrency limits)
• subscribe(handler) -> Result<Subscription, EventBusError> -- dispatch mode inferred from trait
• subscribe_with_policy(handler, policy) -> Result<Subscription, EventBusError> -- policy type is compile-time checked (FailurePolicy for async, NoRetryPolicy for sync)
• subscribe_dead_letters(handler) -> Result<Subscription, EventBusError>
• publish(event) -> Result<(), EventBusError>
• try_publish(event) -> Result<(), EventBusError> -- non-blocking, returns ChannelFull when immediate dispatch capacity is unavailable
• unsubscribe(subscription_id) -> Result<bool, EventBusError>
• shutdown() -> Result<(), EventBusError> -- idempotent, drains in-flight tasks
• async fn is_healthy() -> bool -- checks if the internal control loop is still running

EventBus is Clone -- all clones share the same underlying runtime state.

Handler Traits

• EventHandler<E> -- async handler, dispatched on a spawned task (requires E: Clone)
• SyncEventHandler<E> -- sync handler, awaited inline during dispatch

The dispatch mode is selected automatically based on which trait is implemented. The IntoHandler<E, Mode> trait performs the conversion; the Mode parameter (AsyncMode / SyncMode) is inferred, so callers simply write bus.subscribe(handler).

Subscription & SubscriptionGuard

Subscription holds a SubscriptionId and a bus handle. Call subscription.unsubscribe() to remove the handler, or use bus.unsubscribe(id) directly.

SubscriptionGuard is an RAII wrapper that automatically unsubscribes the listener when dropped. Convert a Subscription via subscription.into_guard(). Call guard.disarm() to prevent the automatic unsubscribe.

Types

• Event -- blanket trait implemented for all T: Send + Sync + 'static
• HandlerResult = Result<(), Box<dyn Error + Send + Sync>>
• FailurePolicy { max_retries, retry_strategy, dead_letter } -- for async handlers
• NoRetryPolicy { dead_letter } -- for sync handlers (or async handlers that don't need retries)
• IntoFailurePolicy<M> -- sealed trait enforcing compile-time policy/handler compatibility
• DeadLetter { event_name, subscription_id, attempts, error, event, failed_at, listener_name }
• SubscriptionId -- opaque handler ID (wraps u64)

Feature Flags

When the metrics feature is enabled, the bus records:

• eventbus.publish (counter, per event type)
• eventbus.handler.duration (histogram, per event type)
• eventbus.handler.error (counter, per event type)
• eventbus.handler.join_error (counter, per event type)

Observability

JAEB uses the tracing crate throughout. Key spans and events:

• event_bus.publish / event_bus.subscribe / event_bus.shutdown -- top-level operations
• eventbus.handler span -- per-handler execution with event, mode, and listener_id fields
• handler.retry (warn) -- logged on each retry attempt
• handler.failed (error) -- logged when retries are exhausted
• async handler panics are surfaced as task failures and follow retry/dead-letter policy

Architecture

JAEB uses a split control/data architecture:

• A snapshot registry (ArcSwap<RegistrySnapshot>) stores listeners and middleware in immutable per-type slots for low-overhead publish-path reads.
• A lightweight control loop handles async failure notifications, dead-letter routing, and shutdown coordination.

Dispatch uses two lanes per event type:

• sync lane: serialized by a per-type gate (FIFO for sync dispatch)
• async lane: spawned in background and not blocked by sync backlog

Semantics

• publish waits for dispatch and sync listeners to finish.
• publish does not wait for async listeners to finish.
• async handler failures can be retried based on FailurePolicy.
• sync handlers execute exactly once -- retries are not supported; passing a FailurePolicy with retries to a sync handler is a compile error via IntoFailurePolicy<M>.
• after retries are exhausted (async) or on first failure (sync), dead-letter events are emitted when dead_letter: true.
• dead-letter handlers themselves cannot trigger further dead letters (recursion guard).
• shutdown waits for in-flight async listeners (with optional timeout).
• after shutdown, all operations return EventBusError::Stopped.

Error Variants

• EventBusError::Stopped -- shutdown has started or the bus is stopped
• EventBusError::ChannelFull -- publish saturation limit reached (try_publish only)
• EventBusError::InvalidConfig(ConfigError) -- invalid builder/constructor configuration (zero buffer size, zero concurrency)
• EventBusError::MiddlewareRejected(String) -- a middleware rejected the event before it reached any listener
• EventBusError::ShutdownTimeout -- the configured shutdown_timeout expired and in-flight async tasks were forcibly aborted

Examples

See examples/jaeb-demo for a working demo that includes:

• async and sync handlers with retry policies
• dead-letter logging
• Prometheus metrics exporter
• structured tracing setup

Run it with:

cd examples/jaeb-demo
RUST_LOG=info,jaeb=trace cargo run

Notes

• JAEB requires a running Tokio runtime.
• Events must be Send + Sync + 'static. Async handlers additionally require events to be Clone.
• Events are in-process only (no persistence, replay, or broker integration).
• The crate enforces #![forbid(unsafe_code)].

License

jaeb is distributed under the MIT License.

Copyright (c) 2025-2026 Linke Thomas

This project uses third-party libraries. See THIRD-PARTY-LICENSES for the full list of dependencies, their versions, and their respective license terms.