Featured

I analyzed the performance bottlenecks that appeared when concurrent requests increased on an Ollama-based LLM server, then moved the serving layer to vLLM so it could stably handle around 50 concurrent users.

A record of setting up the conditions needed to run the Code Place AI assistant on an operations cluster, including the NVIDIA driver, NVIDIA runtime, k3s containerd configuration, RuntimeClass, and device plugin.

A record of recovering a disk pressure failure in Code Place that started with WAL growth, including Longhorn replica adjustment, CNPG instance recovery, multipathd interference, adding a Longhorn disk, and checking replication slots, WAL state, and RW/RO endpoints afterward.

A record of tracing a Longhorn volume mount failure during recovery after WAL growth, checking Kubernetes events, share-manager state, ext4 messages, and possible multipathd interference.

A record of tracing why JWT authentication filter errors in GRIT were returned as Tomcat 500 errors instead of the common API error response, and how I handled exceptions based on where they occurred.

A record of using Redis Pub/Sub in UniSchedule to solve a problem where chat messages were not delivered to some users in a multi-instance environment.