Continuous Profiling

The third pillar after metrics and tracing. By 2026 it's affordable and ubiquitous.

What it is

Sample CPU / memory / IO / off-CPU from production processes continuously. Get flame graphs filtered by service / version / time / container — without restarting anything.

Open-source backends

• ★ Pyroscope (Grafana Labs) — Phlare-merged; multi-format (pprof, JFR, .NET); pairs with Grafana. The default.
• Parca (Polar Signals) — eBPF-based whole-system profiling; minimal app instrumentation.
• Apache SkyWalking eBPF — bundled with SkyWalking.

Hosted

• ★ Grafana Cloud Profiles — managed Pyroscope; generous free tier.
• Polar Signals Cloud — managed Parca.
• Datadog Continuous Profiler — paid; great UI.
• New Relic CodeStream — Code-level perf; paid.
• Sentry Continuous Profiling — bundled with Sentry; paid tier.

Agents / instrumentation

• ★ Pyroscope agent — Go / Java / Python / Ruby / .NET / Node.js (eBPF + manual SDKs).
• ★ Parca agent — eBPF; sees every process on a host with no app changes.
• OpenTelemetry profiling signal — newer standard; SDKs adding it; Pyroscope and Parca both can ingest.
• pprof endpoints in Go / Rust services — Pyroscope scrapes these directly.
• Async-profiler (Java) — bundled or scraped by Pyroscope.

Profiling types

• CPU — where time is spent.
• Allocations — where memory is being allocated.
• Inuse memory — where memory currently lives.
• Goroutines / threads — concurrency view.
• Off-CPU / blocked — where threads are waiting.
• Lock contention — synchronization.

Patterns to adopt

• ★ Default to eBPF where possible. Parca / Pyroscope eBPF agents avoid the "instrument every service" problem.
• Tag profiles with service / version / region. Filter in the UI.
• Compare versions — the killer feature. "Did my change make this slower?"
• Sample, don't always-record at full rate. 100Hz is usually plenty.
• Look at off-CPU when latency is the question — CPU profiles miss waiting time.

When profiling matters

• A service is slow and metrics / traces don't explain why.
• Memory growth that smells like a leak.
• A specific request type is slow consistently — flame graph the spans involved.
• Cost optimization: figure out what's burning CPU you're paying for.
• Pre-launch perf testing.

When it's overkill

• Tiny services, low traffic, occasional perf issue. node --prof or pprof ad-hoc is enough.
• You don't have basic metrics yet — add those first.

Pick this if…

• Default OSS: Pyroscope.
• Whole-system / eBPF: Parca.
• Hosted, free tier: Grafana Cloud Profiles.
• Already on Datadog: their continuous profiler.
• Java-heavy: async-profiler integration via Pyroscope.
• Go service: Pyroscope scraping /debug/pprof.