CHANGELOG.md

Changelog

0.7.2

First successful crates.io release of the split workspace crates (capp-config, capp-queue, capp-cache, capp-router, capp-urls) alongside capp. No functional changes since 0.7.1 — the 0.7.x publishes failed because the sub-crates had never been pushed to the registry and the publish workflow's dry-run loop couldn't resolve them; the workflow now publishes the workspace in dependency order.

0.7.1

Fixed

• FjallRoundRobinTaskQueue::pop() no longer evicts a tag from the rotation the moment its sub-queue transiently empties. It now scans forward from the cursor, skipping empty tags without removing them, and only advances the cursor past a tag it actually serves. Previously an evicted tag only rejoined at the tail of the rotation on its next push(), which starved tags whose backlog repeatedly grazed zero under concurrent consumption (e.g. concurrency_limit close to the number of distinct tags) in favor of tags with deeper, steadier backlogs. QueueEmpty is now returned only after a full lap of all known tags finds no queued work, matching the other TaskQueue backends' contract.

0.7.0

Major architecture change: the mailbox worker pool is gone. The runtime is now one dispatcher task driving one tower service; concurrency, rate-limiting, retry, etc. all live in the caller's ServiceBuilder chain. Stock tower middleware composes without Clone shims because there is exactly one service instance.

Changed

• Breaking: MailboxConfig lost worker_count and prefetch_per_worker. Add .concurrency_limit(N) to your ServiceBuilder chain instead. Remaining fields: producer_buffer, result_buffer, max_retries, dequeue_backoff, stop_when_idle.
• Breaking: MailboxService<D, Ctx> is now BoxService<ServiceRequest<D, Ctx>, (), BoxError> (was BoxCloneService). Services no longer need to be Clone.
• Breaking: build_service_stack drops the Clone bound on the inner service; signature otherwise unchanged.
• Breaking: ServiceRequest lost the worker_id field.
• Breaking: StatsSnapshot is now flat — processed, succeeded, failed, last_latency, in_flight (new), terminal_failures. The per-worker HashMap<usize, WorkerMetrics> is gone, along with WorkerMetrics. The internal StatsEvent::Worker variant was replaced by Call { success, latency } + InFlight(u64).
• Breaking: Envelope<D> type removed (was only used between the dispatcher and the removed worker pool).
• Breaking: OpenTelemetry metrics no longer carry a worker_id attribute — they're emitted as plain counters/histogram.
• examples/hackernews/main.rs replaces the inline tokio::sleep(POLITE_DELAY) with a tower concurrency_limit(2).rate_limit(2, 1s) chain.
• examples/httpbin_tower.rs drops the Buffer + BoxCloneService shim that was previously needed to make the rate-limited service Clone.

Added

• MailboxConfig::stop_when_idle: bool — auto-stop the dispatcher once the queue is empty and all in-flight futures have completed. Suitable for finite crawls; existing tests cover the spawn-future drain semantics.
• StatsSnapshot::in_flight: u64 — live gauge of dispatched-but-not- yet-completed futures.
• New in-tree tests: concurrency_limit_caps_in_flight, stop_when_idle_with_spawned_futures, service_without_clone_compiles.

Removed

• async-channel dependency. The shared MPMC channel that fed envelopes to the worker pool is gone with the pool.

Migration from 0.6.x

// before
MailboxConfig {
    worker_count: 4,
    prefetch_per_worker: 1,
    producer_buffer: 128,
    result_buffer: 128,
    max_retries: 2,
    dequeue_backoff: Duration::from_millis(50),
    stop_when_idle: false,
}

// after
let inner = ServiceBuilder::new()
    .concurrency_limit(4)              // ← where worker_count went
    .service(my_service);
let service = build_service_stack(inner, ServiceStackOptions::default());

MailboxConfig {
    producer_buffer: 128,
    result_buffer: 128,
    max_retries: 2,
    dequeue_backoff: Duration::from_millis(50),
    stop_when_idle: false,
}

If you logged req.worker_id inside service handlers, drop the field. If you read StatsSnapshot.workers, read the flat counters instead.

0.6.2

Added

• Internal capp-testkit workspace crate for reusable local HTTP fixtures used by tests and examples.
• Hyper/Tokio-backed local fixture server with request logging and /stats reporting.
• Built-in local fixture scenarios: blog, blog_with_posts(n), catalog, infinite.
• local_blog_crawl mailbox example that crawls a generated local 100-post blog and validates coverage against fixture stats.
• TaskQueue::recover_inflight trait method (default returns Ok(0)) for reclaiming tasks owned by a previous, crashed process at startup.
• FjallTaskQueue and InMemoryTaskQueue now track an inflight set: pop moves the task there, and ack / nack / push clear it.

Changed

• Reintroduced a realistic crawler example path using the mailbox runtime and a local in-process site instead of only external-service demos.
• Expanded test coverage around capp-testkit, healthcheck timeout/failure behavior, and queue serializer success/error paths.
• healthcheck::internet no longer panics on reqwest errors and no longer depends on a 404 + content_length == 9 magic check. New contract: returns true iff any HTTP response is received within 1 second; any error or timeout yields false.
• Breaking: TaskId::from_string now returns Result<Self, uuid::Error> instead of panicking on malformed input. TaskId also implements FromStr so callers can s.parse::<TaskId>().
• Breaking: MailboxConfig.inbox_capacity renamed to prefetch_per_worker. The mailbox runtime now uses a single shared async-channel MPMC channel of capacity prefetch_per_worker * worker_count instead of N per-worker mpsc inboxes. Eliminates head-of-line blocking when one worker is slow. (Note: the worker pool was removed entirely in 0.7.0.)
• Breaking: WorkerResult::Return variant removed. Workers no longer produce "return" results; on shutdown the dispatcher closes the shared channel and any tasks not picked up remain in the queue's inflight set for the next process to recover.
• Breaking: ServiceStackOptions reduced to a single field (timeout: Option<Duration>). The buffer and concurrency_limit fields were removed.
• Breaking: build_service_stack no longer adds ConcurrencyLimitLayer, .load_shed(), or .buffer() to the service stack — only .timeout() remains. MailboxConfig.worker_count is now the sole concurrency knob; the previous defaults (concurrency_limit = Some(1) + buffer = 1) silently serialized the worker pool through a shared semaphore and a redundant mpsc. Callers that need a per-resource sub-cap should compose their own ServiceBuilder chain before passing the service in.
• Breaking: StatsSnapshot.queue_depth field, StatsEvent::QueueDepth variant, and the OTel capp_queue_depth observable gauge were removed. Nothing was sampling queue depth, so the metric was always zero. Real depth observability needs TaskQueue::len() + periodic sampling and is not part of this release.
• Mailbox dispatcher gives ControlCommand priority over other events via biased; selection.
• handle_worker_result::Ack no longer writes the task before acking; ack removes the row.

Fixed

• Crashed-mid-flight tasks on the Fjall backend are no longer lost. On startup, spawn_mailbox_runtime calls recover_inflight to move any leased-but-unacked rows back to the queue.