It's the end of observability as we know it (and I feel fine)

Original Article ↗ Hacker News Discussion ↗

Cost, Data Volume, and Architecture

  • Many see the proposed “AI-first” observability model as a cost amplifier: unified sub‑second stores, anomaly detection, and constant analysis imply huge telemetry and compute bills.
  • Several argue that LLMs don’t remove the need for graphs, alerts, or careful logging strategy; they just sit on top of an already-expensive stack.
  • There’s concern that using LLMs to proactively scan all telemetry for issues would be far more expensive than traditional threshold-based alerting.

What LLMs Actually Add

  • Strong support for using LLMs to accelerate root cause analysis once you know something is wrong: given a starting signal (alert, spike), an agent can traverse logs/metrics/traces, test hypotheses, and propose narratives.
  • Others note the blog’s demo was closer to “LLM as smart pivoting UI” than a full agentic workflow; the human still framed the question and knew where to look.
  • Some see LLMs as valuable integrators across disparate tools (traces, logs, metrics) without deep product-level integration.

Skepticism, Hype, and Marketing

  • Many call the post a thin or not-at-all-veiled product pitch, with grandiose language (“end of observability”, “speed of AI”) that doesn’t match the incremental reality.
  • Critics stress that anomaly detection and RCA remain intrinsically hard; framing AI as paradigm-ending is seen as overselling.

Reliability, Determinism, and Correlation Traps

  • A recurring theme: nondeterministic, occasionally-confidently-wrong systems are dangerous for RCA. People want tools that surface hypotheses but that also quantify uncertainty or actively try to disprove themselves.
  • Several warn about spurious correlations in time-series and “AI that correlates everything with everything”; statistical metrics (r², p‑values) are easily abused by both humans and LLMs.

Skills, Responsibility, and Over-Reliance

  • Debate over whether AI will help people learn or encourage shallow, copy‑paste understanding; concern that less-expert staff plus AI will be “good enough” for management.
  • Strong view that humans must remain accountable for decisions; AI is best treated like a powerful but error-prone intern.
  • Some see real upside for small SRE/IT teams and SMBs: LLMs can lower the bar to “big-league” observability setups and faster incident triage, without staffing large expert teams.

Tooling and UX Frustrations

  • Multiple comments say if you need an LLM to pivot between traces, logs, and metrics, the observability product probably has UX/feature gaps.
  • Others counter that most observability UIs are bad enough that a natural-language layer is a net win, even if it doesn’t replace graphs.