Important: This documentation is about an older version. It's relevant only to the release noted, many of the features and functions have been updated or replaced. Please view the current version.

Documentation Grafana Agent Flow mode Reference Components reference pyroscope.ebpf

pyroscope.ebpf

BETA: This is a beta component. Beta components are subject to breaking changes, and may be replaced with equivalent functionality that cover the same use case.

pyroscope.ebpf configures an ebpf profiling job for the current host. The collected performance profiles are forwarded to the list of receivers passed in forward_to.

You can specify multiple pyroscope.ebpf components by giving them different labels, however it is not recommended as it can lead to additional memory and CPU usage.

Usage

river Copy

pyroscope.ebpf "LABEL" {
  targets    = TARGET_LIST
  forward_to = RECEIVER_LIST
}

Arguments

The component configures and starts a new ebpf profiling job to collect performance profiles from the current host.

You can use the following arguments to configure a pyroscope.ebpf. Only the forward_to and targets fields are required. Omitted fields take their default values.

Exported fields

pyroscope.ebpf does not export any fields that can be referenced by other components.

Component health

pyroscope.ebpf is only reported as unhealthy if given an invalid configuration.

Debug information

• targets currently tracked active targets.
• pid_cache per process elf symbol tables and their sizes in symbols count.
• elf_cache per build id and per same file symbol tables and their sizes in symbols count.

Debug metrics

• pyroscope_fanout_latency (histogram): Write latency for sending to direct and indirect components.
• pyroscope_ebpf_active_targets (gauge): Number of active targets the component tracks.
• pyroscope_ebpf_profiling_sessions_total (counter): Number of profiling sessions completed.
• pyroscope_ebpf_profiling_sessions_failing_total (counter): Number of profiling sessions failed.
• pyroscope_ebpf_pprofs_total (counter): Number of pprof profiles collected by the ebpf component.

Profile collecting behavior

The pyroscope.ebpf component collects stack traces associated with a process running on the current host. You can use the sample_rate argument to define the number of stack traces collected per second. The default is 97.

The following labels are automatically injected into the collected profiles if you have not defined them. These labels can help you pin down a profiling target.

Container ID

Each collected stack trace is then associated with a specified target from the targets list, determined by a container ID. This association process involves checking the __container_id__, __meta_docker_container_id, and __meta_kubernetes_pod_container_id labels of a target against the /proc/{pid}/cgroup of a process.

If a corresponding container ID is found, the stack traces are aggregated per target based on the container ID. If a container ID is not found, the stack trace is associated with a default_target.

Any stack traces not associated with a listed target are ignored.

Service name

The special label service_name is required and must always be present. If it’s not specified, it is attempted to be inferred from multiple sources:

• __meta_kubernetes_pod_annotation_pyroscope_io_service_name which is a pyroscope.io/service_name pod annotation.
• __meta_kubernetes_namespace and __meta_kubernetes_pod_container_name
• __meta_docker_container_name

If service_name is not specified and could not be inferred, it is set to unspecified.

Troubleshooting unknown symbols

Symbols are extracted from various sources, including:

• The .symtab and .dynsym sections in the ELF file.
• The .symtab and .dynsym sections in the debug ELF file.
• The .gopclntab section in Go language ELF files.

The search for debug files follows gdb algorithm. For example, if the profiler wants to find the debug file for /lib/x86_64-linux-gnu/libc.so.6 with a .gnu_debuglink set to libc.so.6.debug and a build ID 0123456789abcdef. The following paths are examined:

• /usr/lib/debug/.build-id/01/0123456789abcdef.debug
• /lib/x86_64-linux-gnu/libc.so.6.debug
• /lib/x86_64-linux-gnu/.debug/libc.so.6.debug
• /usr/lib/debug/lib/x86_64-linux-gnu/libc.so.6.debug

Dealing with unknown symbols

Unknown symbols in the profiles you’ve collected indicate that the profiler couldn’t access an ELF file associated with a given address in the trace.

This can occur for several reasons:

• The process has terminated, making the ELF file inaccessible.
• The ELF file is either corrupted or not recognized as an ELF file.
• There is no corresponding ELF file entry in /proc/pid/maps for the address in the stack trace.

Addressing unresolved symbols

If you only see module names (e.g., /lib/x86_64-linux-gnu/libc.so.6) without corresponding function names, this indicates that the symbols couldn’t be mapped to their respective function names.

This can occur for several reasons:

• The binary has been stripped, leaving no .symtab, .dynsym, or .gopclntab sections in the ELF file.
• The debug file is missing or could not be located.

To fix this for your binaries, ensure that they are either not stripped or that you have separate debug files available. You can achieve this by running:

bash Copy

objcopy --only-keep-debug elf elf.debug
strip elf -o elf.stripped
objcopy --add-gnu-debuglink=elf.debug elf.stripped elf.debuglink

For system libraries, ensure that debug symbols are installed. On Ubuntu, for example, you can install debug symbols for libc by executing:

bash Copy

apt install libc6-dbg

Understanding flat stack traces

If your profiles show many shallow stack traces, typically 1-2 frames deep, your binary might have been compiled without frame pointers.

To compile your code with frame pointers, include the -fno-omit-frame-pointer flag in your compiler options.

Profiling interpreted languages

Profiling interpreted languages like Python, Ruby, JavaScript, etc., is not ideal using this implementation. The JIT-compiled methods in these languages are typically not in ELF file format, demanding additional steps for profiling. For instance, using perf-map-agent and enabling frame pointers for Java.

Interpreted methods will display the interpreter function’s name rather than the actual function.

Example

Kubernetes discovery

In the following example, performance profiles are collected from pods on the same node, discovered using discovery.kubernetes. Pod selection relies on the HOSTNAME environment variable, which is a pod name if Grafana Agent is used as a Grafana Agent Helm chart. The service_name label is set to {__meta_kubernetes_namespace}/{__meta_kubernetes_pod_container_name} from Kubernetes meta labels.

river Copy

discovery.kubernetes "all_pods" {
  role = "pod"
  selectors {
    field = "spec.nodeName=" + env("HOSTNAME")
    role = "pod"
  }
}

discovery.relabel "local_pods" {
  targets = discovery.kubernetes.all_pods.targets
  rule {
    action = "drop"
    regex = "Succeeded|Failed"
    source_labels = ["__meta_kubernetes_pod_phase"]
  }
  rule {
    action = "replace"
    regex = "(.*)@(.*)"
    replacement = "ebpf/${1}/${2}"
    separator = "@"
    source_labels = ["__meta_kubernetes_namespace", "__meta_kubernetes_pod_container_name"]
    target_label = "service_name"
  }
  rule {
    action = "labelmap"
    regex = "__meta_kubernetes_pod_label_(.+)"
  }
  rule {
    action = "replace"
    source_labels = ["__meta_kubernetes_namespace"]
    target_label = "namespace"
  }
  rule {
    action = "replace"
    source_labels = ["__meta_kubernetes_pod_name"]
    target_label = "pod"
  }
  rule {
    action = "replace"
    source_labels = ["__meta_kubernetes_node_name"]
    target_label = "node"
  }
  rule {
    action = "replace"
    source_labels = ["__meta_kubernetes_pod_container_name"]
    target_label = "container"
  }
}
pyroscope.ebpf "local_pods" {
  forward_to = [ pyroscope.write.endpoint.receiver ]
  targets = discovery.relabel.local_pods.output
}

pyroscope.write "endpoint" {
  endpoint {
    url = "http://pyroscope:4100"
  }
}

Docker discovery

The following example collects performance profiles from containers discovered by discovery.docker and ignores all other profiles collected from outside any docker container. The service_name label is set to the __meta_docker_container_name label.

river Copy

discovery.docker "linux" {
  host = "unix:///var/run/docker.sock"
}

discovery.relabel "local_containers" {
  targets = discovery.docker.linux.targets
  rule {
    action = "replace"
    source_labels = ["__meta_docker_container_name"]
    target_label = "service_name"
  }
}

pyroscope.write "staging" {
  endpoint {
    url = "http://pyroscope:4100"
  }
}

pyroscope.ebpf "default" {  
  forward_to   = [ pyroscope.write.staging.receiver ]
  targets      = discovery.relabel.local_containers.output
}