KubeCon North America in Salt Lake City set the pace this week. OpenTelemetry published migration-heavy news for Collector and JavaScript users, and engineering teams kept putting AI to work on root-cause analysis and on instrumenting AI agents. This issue also picks up two notable releases that landed just after the last one.
Releases
- Deprecating the Collector Batch Processor — OpenTelemetry blog. The
batchprocessor is deprecated in favour of exporter-level batching viasending_queue(or the newqueue_batchprocessor). Exporter batching becomes the default around v0.171.0 and the processor is removed around v0.177.0. Read it for the queue-sizing gotcha: the queue now holds unbatched requests, so the defaultqueue_sizecan hold far less data than before. - Announcing the OpenTelemetry JavaScript SDK 3.0 — OpenTelemetry blog. SDK 3.0 removes deprecated packages such as the Jaeger exporter and the OpenTracing/OpenCensus shims, adds dual ESM/CommonJS exports, and marks Logs stable. The API package has no breaking changes, but the minimum is now Node.js 22.15, so check your runtimes before you upgrade.
- Prometheus 3.16.0 — GitHub. This release adds experimental query cost estimation and execution limits (
--enable-feature=query-cost) and an experimentalintegral()function. Smoothed and anchored vector selectors are now stable. It also fixes out-of-order sample loss after WAL checkpoints and several OTLP exemplar bugs. - Meet the new Elastic Streams: managed data pipelines for telemetry — Elastic Observability Labs. Elastic presents Streams as a managed pipeline layer for routing and processing telemetry, including data arriving via S3. It’s worth a look if you’re weighing vendor-run pipelines against running your own Collector fleet.
Deep dives
- When missing telemetry is the signal — OpenTelemetry blog. An empty dashboard can mean a quiet system or a broken pipeline. The post shows how to tell them apart: set expectations for when activity should appear, compare event counts stage by stage, push a synthetic canary event end to end, and watch the Collector’s accepted, sent and queued metrics.
- Building an agent harness for OTel instrumentation with GenAI semantic conventions and Weaver — OpenTelemetry blog. Coding agents plan and write GenAI framework instrumentation, and Weaver’s live-check validates the emitted telemetry against the semantic-convention registry, failing the run on violations. It’s a practical pattern for keeping instrumentation correct while the frameworks keep changing.
- Netflix turns operational data into a knowledge graph for AI-driven observability — InfoQ. Netflix models services, deployments, alerts, incidents and dependencies as an ontology-backed RDF graph, with LLM-assisted enrichers gated by strict validation. The graph feeds its agentic AutoRCA system, and humans still approve the fixes. It’s a good look at the data foundations AI SRE tooling needs.
- Three model calls, 92% of the wait: AI agent observability with OpenTelemetry — Elastic Observability Labs. A worked example of tracing an LLM agent with OpenTelemetry, showing where its latency actually goes. It’s useful if you’re starting on GenAI spans and want a concrete picture of what they reveal.
- Two changes to the OpenTelemetry Kafka receiver that stop one slow tenant stalling the rest — Elastic Observability Labs. A write-up of upstream Kafka receiver changes that isolate tenants from each other in shared Collector pipelines. It’s relevant to anyone running Kafka-buffered, multi-tenant telemetry ingestion.
Incidents
- Microsoft 365 outage: timeline, failure mode and incident response breakdown — Xurrent. A breakdown of the Exchange Online degradation that began on 12 November. Servers failing health checks were pulled from rotation, the rest overloaded, and restored capacity buckled under the backlog. Microsoft hasn’t published a root cause yet, but the lessons on readiness checks versus liveness checks, capping health-check ejections, staged recovery and retry budgets apply widely.
Industry news
- Google open-sources Ambient Quality Agent (AQuA) to diagnose production AI agent failures — InfoQ. Google has released an open-source tool for diagnosing why AI agents fail in production. It’s another sign that agent reliability is becoming its own observability discipline.
— END OF ISSUE —More issues →