Configure profiling and tracing to troubleshoot Grafana

You can set up the grafana-server process to enable certain diagnostics when it starts. This can be useful when investigating certain performance problems. It’s not recommended to have these enabled by default.

Turn on profiling and collect profiles

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 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 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. Adding a small fraction of block and mutex profiling, such as 10-5 (10%-20%) should in general be fine.

Enable profiling:

./grafana server -profile -profile-addr=0.0.0.0 -profile-port=8080

Enable profiling with block and mutex profiling enabled with a fraction of 20%:

./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 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.

You can configure or override profiling settings using environment variables:

export GF_DIAGNOSTICS_PROFILING_ENABLED=true
export GF_DIAGNOSTICS_PROFILING_ADDR=0.0.0.0
export GF_DIAGNOSTICS_PROFILING_PORT=8080
export GF_DIAGNOSTICS_PROFILING_BLOCK_RATE=5
export GF_DIAGNOSTICS_PROFILING_MUTEX_RATE=5

In general, you use the Go command pprof to both collect and analyze profiling data. You can also use curl 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.

curl http://<profile-addr>:<profile-port>/debug/pprof/heap > heap1.pprof
sleep 30
curl http://<profile-addr>:<profile-port>/debug/pprof/heap > heap2.pprof

You can then use pprof tool to compare two heap profiles:

go tool pprof -http=localhost:8081 --base heap1.pprof heap2.pprof

Analyzing high CPU usage:

When experiencing high CPU usage it’s suggested to collect CPU profiles over a period of time, e.g. 30 seconds.

curl 'http://<profile-addr>:<profile-port>/debug/pprof/profile?seconds=30' > profile.pprof

You can then use pprof tool to analyze the collected CPU profile:

go tool pprof -http=localhost:8081 profile.pprof

Use tracing

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:

./grafana server -tracing -tracing-file=/tmp/trace.out

You can configure or override profiling settings using environment variables:

export GF_DIAGNOSTICS_TRACING_ENABLED=true
export GF_DIAGNOSTICS_TRACING_FILE=/tmp/trace.out

View the trace in a web browser (Go required to be installed):

go tool trace <trace file>
2019/11/24 22:20:42 Parsing trace...
2019/11/24 22:20:42 Splitting trace...
2019/11/24 22:20:42 Opening browser. Trace viewer is listening on http://127.0.0.1:39735

For more information about how to analyze trace files, refer to Go command trace.