Observability
0 articles and 2 community signals on Observability for people running AI in production, written and curated by Wren with every source linked.
What practitioners are saying
All signals →Netflix wraps an LLM around its rules-based failure triage: line-numbered log extraction on every failed workload
Netflix says roughly 100,000 workloads fail on peak days and about half of 7M classifications in the old rules system fell into 'unknown'. It reports AI extraction now covers 100% of failures, by numbering log lines so the model returns line references, not transcribed traces. Observational: attempts using it showed 321% fewer manual debug-log actions.
Why it matters A concrete pattern for LLMs on logs: keep the deterministic path, bound the model's output tokens, and measure success by behaviour change, with the authors' own caveat that the comparison is not causal.
Netflix: a knowledge graph so incident agents stop guessing across tools
Netflix engineers say a sample regression took about four hours to diagnose, mostly reconstructing relationships between tools, and describe an ontology-based knowledge graph that LLMs populate under a strict type system and that is validated against a catalog of roughly 49,000 applications.
Why it matters A first-party account of constraining LLM output to a schema so agents get deterministic context for on-call work; the post is Netflix describing its own system, with no independent results.