September 2026 Summaries

Grafana Tempo 3.1 expands the distributed tracing platform with Kafka ingestion enhancements, including TLS, additional SASL authentication methods, rack-aware fetching to reduce cross-zone transfer costs, and configurable producer compression for broader backend compatibility. The release adds query-based trace redaction through TraceQL, enabling users to locate and permanently remove sensitive data across traces while providing dry-run, tenant metrics, and time-range controls. TraceQL metrics gains sampling-aware extrapolation, arithmetic expressions for calculations such as error rates, and faster span-only query processing for vParquet5 blocks. Other improvements include clearer service-graph diagnostics, support for updated OpenTelemetry database semantic conventions, lower metrics-generator resource use, experimental span pruning and filtering for trace-by-ID requests, and experimental trace-diff tools designed to simplify manual and AI-assisted investigations. Tempo 3.1 also makes vParquet5 the default storage format, improving query performance and adding expanded dedicated-column configuration options, while users are advised to review release notes and upgrade guidance before deployment.

AI agent reliability cannot be judged solely through conventional operational metrics such as latency, token use, or error rates, because agents can produce fluent but incorrect responses. The proposed approach uses evaluations—either deterministic checks such as regexes or model-based judges—to measure distinct behaviors including groundedness, task fulfillment, toxicity, privacy leakage, prompt-injection resistance, latency, and cost. Each behavior can then be assigned a service level objective, such as 95% fulfillment over 30 days, with the remaining allowance treated as an error budget that supports controlled experimentation while signaling when quality needs attention. Tracking budget consumption helps distinguish isolated incidents from gradual quality drift, which may otherwise remain hidden in plausible-looking responses, and budget-based alerts can reduce noisy monitoring. Target values should be informed by accumulated evaluation data and adjusted for an agent’s risk profile, with Grafana Cloud’s Agent Observability and Grafana Assistant positioned as tools to create evaluations, recommend targets, and establish SLOs from real agent conversations.

Grafana 13.2 generally introduces multiple notification policies for Grafana-managed alerting, allowing organizations to replace reliance on a single shared routing tree with independently managed, named policy trees. Alert rules can select a policy associated with a team, service, domain, or environment, while unassigned rules continue using the default policy, enabling gradual adoption without disrupting existing configurations. Each policy preserves Grafana’s familiar label-based routing, grouping, timing, and contact-point behavior but can be provisioned, updated, secured, and automated separately through the UI, API, or Terraform. The feature addresses scaling challenges in large shared alerting setups by reducing configuration complexity, isolating changes and notification state between policies, and supporting policy-level role-based access controls so teams can manage their own routing without access to unrelated or sensitive configurations.

Grafana Cloud’s Digital Experience Monitoring combines Frontend Observability and Synthetic Monitoring to help engineering teams understand real user impact, proactively detect problems in critical user journeys, correlate frontend signals with backend traces, and reduce investigation time. Its Session Replay capability, powered by the open-source Faro JavaScript instrumentation library, records user interactions with privacy-first masking and lets teams review synchronized replays alongside performance metrics, errors, user journeys, and traces. New integration between synthetic browser checks and Frontend Observability automatically creates a corresponding frontend session for each check, allowing responders to move directly from an alert and failed execution to a replay of the test, its user-experience metrics, and related traces. Grafana Cloud offers documentation, a Digital Experience Monitoring solutions page, and a free tier that includes up to 100,000 test executions per month.

Grafana Cloud is updating Synthetic Monitoring custom labels to align them with labeling practices across the platform by attaching labels directly to every check metric and log and removing the former `label_` prefix. Previously, labels existed only on the `sm_check_info` metric, requiring joins and specialized naming conventions for dashboards, alerts, cost attribution, and other workflows. Existing users are encouraged to complete a three-stage migration—prefixed, dual-write, and un-prefixed—by March 1, 2027, after which legacy behavior will be retired and unmigrated stacks will be automatically migrated. Administrators must verify labels against the reserved-label list before migrating, while dual-write allows old and new labels to coexist so teams can update queries, dashboards, SLOs, alerts, and reports without disruption. Grafana Labs states that the migration will not increase cardinality or billing costs and that historical data will remain queryable under its original labels throughout the retention period.

Cypress test suites can be monitored over time by using plugin lifecycle hooks to convert per-spec and per-test results, including pass and failure counts, durations, and execution states, into Prometheus metrics. Because Cypress jobs are short-lived, the approach sends accumulated metrics from each completed spec to a Prometheus Pushgateway, which preserves them for Grafana Alloy to scrape and remote-write to Grafana Cloud Metrics. A shared suite-level run ID and optional GitHub Actions run ID enable dashboards to group results by test execution, spec, and CI workflow, helping teams identify flaky tests, failing specs, and performance regressions. The setup keeps monitoring separate from test outcomes by catching metric-push errors, escapes free-text labels to preserve valid Prometheus formatting, and can be enabled through environment variables for either local development or CI environments.