The pod that did not survive Tuesday's deploy

A Vercel AI SDK chatbot deployed to Google Kubernetes Engine throws 500 errors on 8% of requests after a Tuesday deploy. Using Pydantic Logfire’s new Kubernetes observability view, developers identify a pod stuck in an out-of-memory restart loop caused by a tokenizer dependency that doubled memory usage in the new image.