#observability
7 posts.
AI Learned Code, but Software Learns in Operation
AI can absorb enormous written knowledge, but usable software improves through running instances, operational feedback, and production reality.
Trace Should Not Break When Kafka Is in the Middle
When Kafka sits inside an MSA flow, tracing does not continue automatically like synchronous HTTP calls. This post explains how to propagate trace context through Kafka headers and how to think about producer and consumer instrumentation.
JVM Metrics Alone Cannot Explain a Container
Introducing Pletor node-metrics-agent, an open source JVM agent that exposes host and container node metrics through JMX and works well with Prometheus JMX exporter.
Can Kafka Client Metrics Really Close the Observability Gap?
A practical look at Kafka client telemetry from KIP-714: how it works, which metrics it can collect, how to configure it, and where the operational limits are.
The Kafka Broker Slowed Down, but Kafka Was Not the Cause
A practical incident-style walkthrough of a Kafka broker throughput drop, follower replicas falling out of ISR, and the storage path contention below the VM.
Consumer Lag Is Not a Health Score: Thinking in Kafka Consuming Pressure
A practical way to read Kafka Consumer Lag together with producer rate and consumer group capacity instead of treating lag as an absolute health signal.
When Logging Becomes the Bottleneck: Keeping Heavy Appender Work Off the Request Path
A practical guide to Log4j2 Custom Appenders through hot-path protection, bounded queues, AsyncAppender, throttling, and operational trade-offs.