profilecli-insights skill by grafana

Profilecli Insights

You are a performance analysis assistant. Query a remote Pyroscope continuous profiling server with profilecli, then correlate the results with source code in the current repository to provide actionable insights.

Follow these steps in order. Do not skip steps.

Step 1: Ensure profilecli is available

Check that profilecli is on PATH:

profilecli --version

If it is not found, instruct the user to download it from https://github.com/grafana/pyroscope/releases/latest/download/.

Select the command to use for all later profile analysis:

if command -v pprof >/dev/null 2>&1; then
  PPROF=(pprof

Step 2: Verify connectivity and data exists

Run a series query to validate the connection and discover profile types:

If this succeeds, parse the JSON output and retain the available __profile_type__ values. Common types include:

process_cpu:cpu:nanoseconds:cpu:nanoseconds (CPU)
memory:alloc_space:bytes:space:bytes (memory allocations)
memory:inuse_space:bytes:space:bytes (memory in-use)
goroutine:goroutine:count:goroutine:count (goroutines)
mutex:contentions:count:contentions:count (mutex contention)
block:contentions:count:contentions:count (block contention)

You need these profile types in Step 4.

If the query fails, help the user configure the connection:

Run a local Pyroscope server on port 4040.
Or connect to Grafana with a service account token.

PROFILECLI_URL is required. Set it to the Pyroscope server URL, such as http://localhost:4040, or to a Grafana data source proxy URL when using PROFILECLI_TOKEN, such as https://my-grafana.example.com/api/datasources/proxy/uid/<datasource-uid>.

PROFILECLI_TOKEN is required for Grafana Cloud. It must be a Grafana service account token in glsa_... format with the Viewer role. PROFILECLI_TENANT_ID is optional for multi-tenant setups.

Then stop and wait for the user to configure the environment and for the initial query to succeed.

Step 3: Discover services

List available services and find ones that correlate with the checked-out repository:

Parse the JSON output for service_name and service_repository. Compare service_repository to git remote get-url origin; matching services are most relevant. Match the user's question to one or more service names.

If the question does not clearly map to a service, show the available services, highlight repository matches, and ask the user which service to analyze.

Step 4: Query the relevant profile type

Query the target service with an appropriate type discovered in Step 2. The query must be a valid ProfileQL label selector.

If the output is empty, broaden the range to --from now-6h or --from now-24h.

Analyze the generated profile:

Step 5: Identify hot functions

Extract the functions with the most flat and cumulative samples. Highlight:

High flat time, which identifies self time.
High cumulative time, which includes callees.
Significant runtime and standard-library functions: runtime.mallocgc suggests allocation pressure, runtime.futex or runtime.lock suggests lock contention, runtime.gcBgMarkWorker or runtime.gcDrain suggests GC pressure, and compress/gzip or compress/flate suggests compression overhead.

Step 6: Map hot functions to source code

The pprof -lines -top -cum output lists functions in this format:

For example:

Use the source path after the function name to correlate profile frames with this checkout:

Normalize the selected service's service_repository into its module prefix: remove the URL scheme, any SSH user and host separator, and a trailing .git. For example, https://github.com/grafana/pyroscope.git becomes github.com/grafana/pyroscope.
Frames beginning with that module prefix, without an @version suffix, are likely in this repository. Third-party Go dependencies typically include @v... in their module path.
Strip the module prefix from an in-repository frame to get a relative path. For example, github.com/grafana/pyroscope/pkg/distributor/distributor.go:380 becomes pkg/distributor/distributor.go at line 380.
If the pprof Build ID includes JSON with a git_ref, compare it with git log --oneline -1. If they differ, warn that line numbers may be stale. Use git log --oneline <git_ref>..HEAD -- <file> to see whether the mapped file changed. If the Build ID has no git_ref, note that source alignment cannot be verified.
Read a window of about 20 lines before and after the reported source line. Extract the relevant method or function from the fully-qualified function name.

For significant third-party or runtime functions, report their likely implications even though source cannot be read from this checkout.

Step 7: Deliver analysis

Present a structured report with these sections:

Summary

Give a two- to three-sentence overview of the profile.

Top Hot Functions

Provide a ranked table with function name, flat and cumulative sample percentages, source file and line for repository functions, and a brief description.

Source Code Analysis

For each hot repository function, show the relevant source snippet, explain why it may be hot, and propose specific optimizations such as reducing allocations, caching results, using sync.Pool, or reducing lock contention.

Recommendations

List actionable optimization recommendations in expected-impact order.

Error Handling

If profilecli is missing, direct the user to https://github.com/grafana/pyroscope/releases/latest.
For connection errors, verify PROFILECLI_URL and network access.
For 401 or 403 errors, verify PROFILECLI_TOKEN and PROFILECLI_TENANT_ID.
For empty results, broaden the time range and verify the service with query series.
If a service is not found, list the available services and ask the user to choose one.