Skip to content

GRAFANA 2026-10-20 Tier 2

Read original ↗

Break the link between data volume and your observability bill with Adaptive Telemetry

Summary

A Grafana Labs product post framing the core observability-economics thesis — your cost should scale with the value you get from telemetry, not with the raw volume you ingest — and rounding up new capabilities across all three signals in the Adaptive Telemetry suite in Grafana Cloud. The default coping strategies (shorter retention, dropping whole services/environments from collection, or flat probabilistic sampling) are "data diets" that cut to control the bill and risk losing the exact data you need during an incident. Adaptive Telemetry instead keeps the signals that earn their place and makes the cut selective. Adaptive Traces adds three retention refinements on top of its existing tail-sampling engine: volumetric sampling (spreads a single sampling-percentage budget fairly across services/request types, ~25% higher "information density" than probabilistic at equal volume), diversity sampling (every 15 min stores one of each unique attribute combination so new services/code paths are never missed), and query dropped traces (every dropped trace stays retrievable by trace ID for 24h — a lookup safety net that lets you sample aggressively). The headline claim: cut trace volume by 90% while keeping the traces that matter. Adaptive Logs adds Fleet-managed Adaptive Logs (push the same pattern-based drop rules down to the collectors you already run via Fleet Management, so matching logs are filtered before they leave your environment — saving egress and network on top of storage) and Archive and replay (route low-value logs to your own low-cost object storage instead of dropping them; bring them back on demand for an investigation/audit/compliance request). Adaptive Metrics adds Auto-Apply (apply aggregation recommendations with no human in the loop; coverage-first or cost-first mode), recapping the GA covered in the earlier 2026-10-15 post.

Key takeaways

  • The unifying thesis: decouple cost from ingested volume. "Your costs shouldn't scale with how much data you ingest. Instead, they should scale with the value you get from it, so you only pay to keep the data that earns its place." This is the same vendor-pricing tension the observability concept names ("vendors typically charge on ingested data volume … more data ≠ better insights") turned into a product line. (Source: sources/2026-10-20-grafana-break-the-link-between-data-volume-and-your-observability-bill)
  • Flat sampling is a blunt instrument; selective retention is the fix. "Many teams sample with a fixed percentage, which treats a slow checkout request the same as a routine health check." Cut the percentage and the bill drops but so does the chance the trace you need survives. Adaptive Traces' three refinements all attack which traces to keep, not how many.
  • Volumetric sampling: one knob, fair budget allocation. You set a single sampling percentage; the system "spreads your sampling budget fairly across services and request types, so low-volume services aren't crowded out, and it keeps adjusting as your traffic changes." No per-service policies to hand-tune. Grafana Labs' own testing: "about 25% higher information density than probabilistic sampling at the same volume."
  • Diversity sampling: guarantee coverage of the long tail. "Every 15 minutes, it stores a unique combination of attributes such as service, route, and status codes." When a new service or code path appears you don't have to write/update custom sampling logic to catch it — at least one representative of each attribute combination is always kept.
  • Query dropped traces: a 24h lookup safety net, not a search index. Adaptive Traces keeps every dropped trace retrievable by trace ID for 24 hours after ingest. If you have the ID (from a log line, support ticket, or testing framework) you can fetch that exact trace even though it was dropped. Critically: "Dropped traces don't appear in search or aggregations, so it's a lookup, not a replacement for sampling the traces you want to search." This is what lets teams "sample aggressively with confidence."
  • Fleet-managed Adaptive Logs: drop at the source to save egress + network. The same drop rules/exemptions are pushed directly to the collectors you already run, via Fleet Management. "Matching logs are filtered out before they ever leave your environment, so you save on egress and network overhead on top of your storage savings." Grafana keeps the collector pipelines in sync for you; pattern-based recommendations still run in Grafana Cloud for everything that remains.
  • Archive and replay: keep a way back instead of permanently dropping. Because "a dropped line can't be recovered," teams hold back on dropping. Archive routes logs to your own low-cost object storage instead of dropping them; Replay brings them back on demand for an investigation, audit, or compliance request. You keep the savings from not indexing low-value logs without permanently losing data you might need — a cold-tier/object-store move applied to log economics.
  • Adaptive Metrics Auto-Apply (recap): aggregation recommendations are applied continuously with no human in the loop; coverage-first keeps pace with usage automatically, cost-first guarantees the series count never increases. (Full mechanics in sources/2026-10-15-grafana-coverage-first-or-cost-first-auto-apply-adaptive-metrics.)
  • All of this is signal-agnostic. The post repeatedly stresses the pattern applies to "metrics, logs, traces, or profiles" — keep what's valuable as the data grows, framed explicitly against the AI-era telemetry explosion ("telemetry is generated by every model call, tool invocation, and agent step").

Systems / concepts extracted

  • Adaptive Telemetry — the umbrella suite (a Grafana Cloud feature family, not a standalone system) spanning Adaptive Traces / Adaptive Logs / Adaptive Metrics. Recorded as a ## section on Grafana Cloud per the Systems gate (feature/SKU of an existing system; no independent architecture disclosed).
  • Adaptive Traces — tail-sampling engine + the three new refinements (volumetric sampling, diversity sampling, query dropped traces). Grafana Cloud feature; ## section, not a page.
  • Adaptive Logs — pattern-based log trimming + Fleet-managed drop + Archive/replay. Grafana Cloud feature; ## section, not a page.
  • Adaptive Metrics — usage-driven aggregation + Auto-Apply (already documented on the Grafana Cloud page from the 2026-10-15 source).
  • Volumetric / diversity sampling, query dropped traces, Fleet-managed Adaptive Logs, Archive & replay — single-source, product-specific mechanisms. Recorded as prose + tags, NOT minted as pages (Taxonomy / Systems gate: one source, vendor-specific implementations; map into the existing concepts below). Lint-promotion candidates if a second source (e.g. OpenTelemetry tail-sampling, a non-Grafana sampling post) cites the same ideas.
  • Observability — the home concept; the cost-vs-volume thesis is a direct instance of its build-vs-buy "pricing-model misalignment / more-data-≠-better-insights" framing.
  • Metric cardinality — the "store raw data elsewhere / aggregate what nobody queries" response family; Adaptive Metrics Auto-Apply is the usage-driven aggregation shield.
  • Egress cost — Fleet-managed drop-at-source is an egress-reduction move: filter before data crosses the network boundary, the same "where in the pipeline you pay the cross-boundary byte matters" lesson.
  • Tiered storage to object store — Archive & replay routes cold/low-value logs to customer-owned object storage with on-demand replay; a log-economics instance of the hot-index / cold-object-store split.
  • Rapid7 — named adopter (plain text, no page): "using Adaptive Telemetry cut our metrics costs enough to fund logs and traces as well."

Operational numbers / specifics

  • Cut trace volume by up to 90% (headline claim for Adaptive Traces' combined refinements) while keeping high-value traces.
  • Volumetric sampling: ~25% higher information density than probabilistic sampling at the same volume (Grafana Labs' internal testing).
  • Diversity sampling cadence: every 15 minutes stores a unique attribute combination (service / route / status codes).
  • Query dropped traces: 24-hour by-trace-ID retrievability window after ingest; dropped traces excluded from search/aggregations (lookup only).
  • GA / preview status (as of post): volumetric + diversity sampling and query-dropped-traces are GA; Fleet-managed Adaptive Logs is public preview; Archive is public preview, Replay coming soon; Adaptive Metrics Auto-Apply is GA.

Caveats

  • Borderline-scope product-launch/roundup post (Tier-2 Grafana). Ingested because the selective-retention mechanisms (budget-fair sampling, diversity guarantee, by-ID lookup of dropped data, source-side drop for egress, archive-to-object-store-with-replay) and the cost-vs-volume decoupling thesis are genuine, reusable observability-economics design content that maps cleanly onto existing concepts. Marketing (CTAs, "sign up for the private preview," customer quotes) is confined to intro/outro.
  • No fleet-scale production numbers beyond the 90% / 25% / 15-min / 24h figures above; the 90% and 25% are vendor claims, not independently reproduced. No internals of how volumetric/diversity sampling compute their decisions, no tail-sampling pipeline architecture, no Fleet Management wire details.
  • Vendor-specific to Grafana Cloud / Adaptive Telemetry.

Source

Last updated · 854 distilled / 2,448 read