Teleport
Profiling
Version preview- Older Versions
Teleport leverages Go's diagnostic capabilities to collect and export profiling data. Profiles can help identify the cause of spikes in CPU, the source of memory leaks, or the reason for a deadlock.
Using the Debug Service
The Teleport Debug Service enables administrators to collect diagnostic profiles without enabling pprof endpoints at startup. The service, enabled by default, ensures local-only access and must be consumed from inside the same instance.
teleport debug profile
collects a list of pprof profiles. It outputs a
compressed tarball (.tar.gz
) to STDOUT. You decompress it using tar
or
direct the result to a file.
By default, it collects goroutine
, heap
and profile
profiles.
Each profile collected will have a correspondent file inside the tarball. For
example, collecting goroutine,trace,heap
will result in goroutine.pprof
,
trace.pprof
, and heap.pprof
files.
Collect default profiles and save to a file.
teleport debug profile > pprof.tar.gztar xvf pprof.tar.gzCollect default profiles and decompress it.
teleport debug profile | tar xzv -C ./Collect "trace" and "mutex" profiles and save to a file.
teleport debug profile trace,mutex > pprof.tar.gzCollect profiles setting the profiling time in seconds
teleport debug profile -s 20 trace > pprof.tar.gz
If your Teleport configuration is not placed on the default path
(/etc/teleport.yaml
), you must specify its location to the CLI command
using the -c/--config
flag.
If you're running Teleport on a Kubernetes cluster you can directly collect profiles to a local directory without an interactive session:
kubectl -n teleport exec my-pod -- teleport debug profile > pprof.tar.gz
After extracting the contents, you can use go tool
commands to explore and
visualize them:
Opens the terminal interactive explorer
go tool pprof heap.pprofOpens the web visualizer
go tool pprof -http : heap.pprofVisualize trace profiles
go tool trace trace.pprof
Using diagnostics endpoints
The profiling endpoint is only enabled if the --debug
flag is supplied.
Teleport's diagnostic HTTP endpoints are disabled by default. You can enable them via:
Start a teleport
instance with the --diag-addr
flag set to the local
address where the diagnostic endpoint will listen:
sudo teleport start --debug --diag-addr=127.0.0.1:3000
Edit a teleport
instance's configuration file (/etc/teleport.yaml
by
default) to include the following:
teleport:
diag_addr: 127.0.0.1:3000
To enable debug logs:
log:
severity: DEBUG
Verify that Teleport is now serving the diagnostics endpoint:
curl http://127.0.0.1:3000/healthz
Collecting profiles
Go's standard profiling endpoints are served at http://127.0.0.1:3000/debug/pprof/
.
Retrieving a profile requires sending a request to the endpoint corresponding
to the desired profile type. When debugging an issue it is helpful to collect
a series of profiles over a period of time.
CPU
CPU profile shows execution statistics gathered over a user specified period:
Download the profile into a file:
curl -o cpu.profile http://127.0.0.1:3000/debug/pprof/profile?seconds=30Visualize the profile
go tool pprof -http : cpu.profile
Goroutine
Goroutine profiles show the stack traces for all running goroutines in the system:
Download the profile into a file:
curl -o goroutine.profile http://127.0.0.1:3000/debug/pprof/goroutineVisualize the profile
go tool pprof -http : goroutine.profile
Heap
Heap profiles show allocated objects in the system:
Download the profile into a file:
curl -o heap.profile http://127.0.0.1:3000/debug/pprof/heapVisualize the profile
go tool pprof -http : heap.profile
Trace
Trace profiles capture scheduling, system calls, garbage collections, heap size, and other events that are collected by the Go runtime over a user specified period of time:
Download the profile into a file:
curl -o trace.out http://127.0.0.1:3000/debug/pprof/trace?seconds=5Visualize the profile
go tool trace trace.out
Further Reading
- More information about diagnostics in the Go ecosystem: https://go.dev/doc/diagnostics
- Go's profiling endpoints: https://golang.org/pkg/net/http/pprof/
- A deep dive on profiling Go programs: https://go.dev/blog/pprof