Performance Monitoring 2026-2027 — What Actually Shipped

Forecasts in this space age badly, so this is split into two parts: what has already shipped and changed how monitoring works, and what is visibly in progress. Treat the second half as expectations, not facts.

What already changed: INP replaced FID

Interaction to Next Paint became a Core Web Vital on 12 March 2024, retiring First Input Delay. The difference is substantive rather than cosmetic: FID measured only the delay before the first interaction's handler started, so a page could score perfectly while every click after the first took a second to render. INP measures the full duration — input delay, processing, and the next paint — across all interactions in a visit, and reports near the worst one. Dashboards still tracking FID are measuring something Google stopped using.

What already changed: attribution got possible

Knowing INP is 600 ms is useless without knowing why. The Long Animation Frames API, available in Chromium from version 123, reports frames that took too long along with the scripts responsible — source URL, character position, and the invoker that triggered them. Combined with the web-vitals library's attribution build, you get "this interaction was slow because of this third-party script" in real user data rather than in a lab reproduction. This is the biggest practical change in front-end monitoring in years.

What already changed: OpenTelemetry became the default plumbing

Tracing, metrics and now logs share one instrumentation layer and one wire format, with vendor-specific agents relegated to the export step. The practical effect is that swapping backends stopped being a rewrite, which in turn made pricing negotiable. Semantic conventions for HTTP and database spans have stabilised, so dashboards built on attribute names no longer break on collector upgrades.

What already changed: continuous profiling went mainstream

Always-on profilers based on eBPF — Parca, Pyroscope and the offerings built on them — sample production CPU and memory at low enough overhead to leave running permanently. That closes the gap between "this endpoint is slow" and "this function is why", without reproducing the load in staging.

What is in progress

• Soft navigation measurement. Single-page apps have never had comparable LCP or CLS numbers because the metrics are tied to hard navigations. The soft navigations heuristic in Chromium aims to fix that; it is still experimental and the numbers it produces are not yet comparable across frameworks.
• Cross-browser field data. CrUX only covers Chrome. Safari and Firefox expose progressively more of the Performance Timeline, and the more of it that lands, the less your field data is a Chrome-shaped sample.
• Cost as a monitoring axis. Observability bills grew fast enough that sampling policy, cardinality limits and tiered retention are now design decisions rather than afterthoughts. Expect more tooling that reports spend per service alongside latency.

What to do with this now

Move off FID, adopt the attribution build of web-vitals, and send LoAF data with your INP reports so slow interactions arrive with a cause attached. Instrument with OpenTelemetry rather than a vendor SDK so the backend decision stays reversible. And set your sampling and retention policy deliberately before the invoice sets it for you.