Software Engineer building scalable backend systems.

A 1,000-email batch required roughly 8 minutes, blocking downstream campaign work.

Introduced bounded ThreadPoolTaskExecutor execution with controlled concurrency and CallerRunsPolicy backpressure.

At-least-once event delivery could produce duplicate customer emails.

Made Kafka consumers idempotent using uniquely constrained processed-event records and transactional patterns.

Publishing events after a DB commit risked lost messages or dual-write inconsistency.

Persisted events to an outbox inside the DB transaction; lease-based publishers drain it to Kafka with oldest-unpublished-age monitoring.

A GC pause let a lock lease expire, so a second scheduler acquired the lock — the original resumed and dispatched twice.

Reproduced the race, then added fencing tokens, an idempotency guard and monitoring to make dispatch safe under pauses.

Three sequential model calls added up to ~6 seconds of blocking latency per request.

Fanned the calls out concurrently and joined them with CompletableFuture.allOf().

Analytics endpoints were slow under reporting load and contended with transactional traffic.

Analyzed query plans, added composite indexes and isolated the reporting datasource.