Event data has a way of humbling neat architecture diagrams. It arrives late. It arrives twice. It arrives with incorrect attributes.
AI applications are messy data producers. They create prompts, completions, tool calls, retrieval traces, ranking signals, evaluation scores, safety.
Agents retry. Networks fail. Jobs time out after doing some work. APIs return ambiguous responses. Schedulers run the same workflow twice.
By Alex Merced, Head of Developer Relations at Dremio
