TODO.md

seda-bus — TODO

Done

• Event-driven worker pool with per-stage concurrency permits (1.3.0).
• Retry + dead-letter; atomic-write + ordered-replay persistence; ExactlyOnce dedup-on-replay (1.3.0).
• JUnit suite (1.3.0).
• Compile target lowered to Java 11 (1.3.1).

Next

2.0 — the adaptive controller

The headline missing piece of the SEDA design.

• Per-stage instrumentation: queue depth, wait time, service time, throughput, nack rate — sampled cheaply on the drain path.
• A controller thread that periodically re-tunes each stage's concurrency permit from measured latency vs. a target.
• Load shedding: when a stage is saturated, reject at admission (already the mechanism) and optionally signal upstream stages to slow.
• Config to cap total threads and per-stage min/max concurrency.
• A pluggable controller policy so callers can supply their own.

Smaller

• Metrics hooks (a ChannelMetrics listener) so hosts can export to their own telemetry without the controller.
• Configurable BATCH size per channel (currently a constant 64).
• Backoff between retry attempts (currently immediate re-queue).
• Dead-letter file rotation (channel deadLetter.json grows unbounded).
• receive(timeout) / pull-model ergonomics: a typed poll that returns the handler result.
• Optional priority within a stage (currently strict FIFO per channel).
• Document the persistence file format and the replay ordering guarantees.

Housekeeping

• Publish to a real repository (currently local / jitpack only).
• Property-based tests for the routing-slip engine and for replay under simulated crashes.
• A couple of the capacity / retry / dead-letter tests are timing-sensitive and flake under load (pass reliably in isolation). Make them deterministic (inject a clock / use latches instead of sleeps).