The `grafana-server` can be started with the command-line option `-profile` to enable profiling, `-profile-addr` to override the default HTTP address (`localhost`), and
`-profile-port` to override the default HTTP port (`6060`) where the `pprof` debugging endpoints are available. Further, [`-profile-block-rate`](https://pkg.go.dev/runtime#SetBlockProfileRate) controls the fraction of goroutine blocking events that are reported in the blocking profile, default `1` (i.e. track every event) for backward compatibility reasons, and [`-profile-mutex-rate`](https://pkg.go.dev/runtime#SetMutexProfileFraction) controls the fraction of mutex contention events that are reported in the mutex profile, default `0` (i.e. track no events). The higher the fraction (that is, the smaller this value) the more overhead it adds to normal operations.
Running Grafana with profiling enabled and without block and mutex profiling enabled should only add a fraction of overhead and is suitable for [continuous profiling](https://grafana.com/oss/pyroscope/). Adding a small fraction of block and mutex profiling, such as 10-5 (10%-20%) should in general be fine.
Enable profiling with block and mutex profiling enabled with a fraction of 20%:
```bash
./grafana server -profile -profile-addr=0.0.0.0 -profile-port=8080 -profile-block-rate=5 -profile-mutex-rate=5
```
Note that `pprof` debugging endpoints are served on a different port than the Grafana HTTP server. Check what debugging endpoints are available by browsing `http://<profile-addr><profile-port>/debug/pprof`.
There are some additional [godeltaprof](https://github.com/grafana/pyroscope-go/tree/main/godeltaprof) endpoints available which are more suitable in a continuous profiling scenario. These endpoints are `/debug/pprof/delta_heap`, `/debug/pprof/delta_block`, `/debug/pprof/delta_mutex`.
In general, you use the [Go command pprof](https://golang.org/cmd/pprof/) to both collect and analyze profiling data. You can also use [curl](https://curl.se/) or similar to collect profiles which could be convenient in environments where you don't have the Go/pprof command available. Next, some usage examples of using curl and pprof to collect and analyze memory and CPU profiles.
**Analyzing high memory usage/memory leaks:**
When experiencing high memory usage or potential memory leaks it's useful to collect several heap profiles and later when analyzing, compare them. It's a good idea to wait some time, e.g. 30 seconds, between collecting each profile to allow memory consumption to increase.
The `grafana-server` can be started with the arguments `-tracing` to enable tracing and `-tracing-file` to override the default trace file (`trace.out`) where trace result is written to. For example:
