9 posts in Architecture.
9 min
At a 90% hit rate your database is provisioned for 10% of your traffic. Nobody decided that — it happened gradually. Which is why 'fall back to the DB' turns a 40-second Redis blip into a database outage, and what to build instead in Spring Boot.
10 min
A two-hour outage leaves a million events in the outbox. Flushing them is how you turn one incident into two. Why a fixed drain rate is a guess, and how to build a closed-loop drain in Spring Boot that lets downstream latency set the pace.
10 min
Lag is climbing, so you raise the concurrency. Now it is worse. Why consumer count is the wrong lever when downstream is the constraint, how slow processing triggers a rebalance death spiral, and what to tune instead in Spring Boot.
8 min
Project Loom gets discussed as one thing. It is three, at three very different maturity levels — and one of them has been in preview seven times, with breaking API changes as recently as JDK 26. What that means for a Spring Boot codebase you actually have to ship.
11 min
Counting errors is the wrong signal — it alerts on your busiest merchant every afternoon and stays silent when a bank route quietly dies. A windowed failure rate in Kafka Streams and Spring Boot, and the operational details that decide whether it survives production.
9 min
Running cache, queue, locks, sessions and rate limiting on a single Redis is the right call for most young systems. It also quietly builds one failure domain spanning all of them — and one of those jobs fails as silent data corruption rather than an outage.
11 min
The dual-write bug survives @Transactional, and it quietly loses money. Here is the architecture that fixes it at a million transactions a second — a hash-partitioned outbox on Oracle, SKIP LOCKED relays and Kafka, in Java 21 and Spring Boot — plus some honest scrutiny of the numbers.
12 min
Once your data spans services, ACID stops at the service boundary. Sagas trade atomicity for availability — but the compensations, the outbox, and the isolation anomalies are where the real engineering is.
10 min
Connection exhaustion takes down a database that isn't even busy. Here's why the arithmetic guarantees it at scale, and how connection multiplexing with ProxySQL breaks the link between application pools and database threads.