Skip to main content

Profiling

Profiling samples your process's call stacks while it works, so you can see where the time actually goes — inside the transaction that tracing already told you was slow.

Errora accepts both profile shapes:

  • Transaction profiles (v1) — bounded by one transaction.
  • Continuous profiling chunks (v2) — a rolling stream, not tied to a transaction.

Setup​

Profiling rides on tracing, so traces_sample_rate must be non-zero.

sentry_sdk.init(
    dsn="…",
    traces_sample_rate=0.2,
    profiles_sample_rate=1.0,   # fraction *of sampled transactions*
)
// Node.js
import { nodeProfilingIntegration } from "@sentry/profiling-node";

Sentry.init({
  dsn: "…",
  tracesSampleRate: 0.2,
  profilesSampleRate: 1.0,
  integrations: [nodeProfilingIntegration()],
});
# Laravel .env
SENTRY_TRACES_SAMPLE_RATE=0.2
SENTRY_PROFILES_SAMPLE_RATE=1.0

profiles_sample_rate multiplies with traces_sample_rate: 0.2 × 1.0 means 20% of requests are traced and all of those are profiled.

Reading a profile​

Profiling lists profiles with their transaction, duration, sample count, environment and release. Open one for:

  • The flamegraph — width is time, not sample count. Each sample is weighted by the gap to the next sample on the same thread, so a thread that happens to be sampled twice as often does not look twice as slow. Gaps are clamped, so a paused profiler can't dominate the graph.
  • Self time per function — the time spent in the function itself, excluding callees.
  • Threads — switch between them.

Slowest functions for the whole project is its own view. That ranking is computed at ingest and stored per function, so it is an indexed query rather than a decompress-everything scan.

API​

GET /api/v1/projects/{project_id}/profiles?stats_period=24h&limit=&offset=
GET /api/v1/projects/{project_id}/profiles/overview
GET /api/v1/projects/{project_id}/profiles/functions
GET /api/v1/projects/{project_id}/profiles/{profile_id}
GET /api/v1/projects/{project_id}/profiles/{profile_id}/flamegraph

Cost​

Profiles are metered as the Profiling service, per profile. They are also the bulkiest payload the SDK sends, so keep profiles_sample_rate low in production once you have found your hot paths — or turn it up temporarily during an investigation.

Troubleshooting​

SymptomCause / fix
No profiles at alltraces_sample_rate is 0, the profiling package isn't installed (Node needs @sentry/profiling-node), or the platform has no profiler
Profiles but empty flamegraphThe transaction was too short to collect samples — profile a heavier endpoint
Everything is <unknown>Native frames without symbols; ship a build with symbols, or profile the interpreted layer
Costs climbingLower profiles_sample_rate; it multiplies with the trace sample rate