This is documentation for the next version of Pyroscope. For the latest stable release, go to the latest version.

Open source

Pyroscope server HTTP API

Pyroscope server exposes an HTTP API for querying profiling data and ingesting profiling data from other sources.


Grafana Pyroscope doesn’t include an authentication layer. Operators should use an authenticating reverse proxy for security.

In multi-tenant mode, Pyroscope requires the X-Scope-OrgID HTTP header set to a string identifying the tenant. This responsibility should be handled by the authenticating reverse proxy. For more information, refer to the multi-tenancy documentation.


There is one primary endpoint: POST /ingest. It accepts profile data in the request body and metadata as query parameters.

The following query parameters are accepted:

nameapplication namerequired
fromUNIX time of when the profiling startedrequired
untilUNIX time of when the profiling stoppedrequired
formatformat of the profiling dataoptional (default is folded)
sampleRatesample rate used in Hzoptional (default is 100 Hz)
spyNamename of the spy usedoptional
unitsname of the profiling data unitoptional (default is samples
aggregrationTypetype of aggregation to merge profilesoptional (default is sum)

name specifies application name. For example:{env=staging,region=us-west-1}

The request body contains profiling data, and the Content-Type header may be used alongside format to determine the data format.

Some of the query parameters depend on the format of profiling data. Pyroscope currently supports three major ingestion formats.

Text formats

These formats handle simple ingestion of profiling data, such as cpu samples, and typically don’t support metadata (e.g., labels) within the format. All necessary metadata is derived from query parameters, and the format is specified by the format query parameter.

Supported formats:

  • Folded: Also known as collapsed, this is the default format. Each line contains a stacktrace followed by the sample count for that stacktrace. For example:
foo;bar 100
foo;baz 200
  • Lines: Similar to folded, but it represents each sample as a separate line rather than aggregating samples per stacktrace. For example:

The pprof format

The pprof format is a widely used binary profiling data format, particularly prevalent in the Go ecosystem.

When using this format, certain query parameters have specific behaviors:

  • format: This should be set to pprof.
  • name: This parameter contains the prefix of the application name. Since a single request might include multiple profile types, the complete application name is formed by concatenating this prefix with the profile type. For instance, if you send CPU profiling data and set name to my-app{}, it will be displayed in pyroscope as my-app.cpu{}.
  • units, aggregationType, and sampleRate: These parameters are ignored. The actual values are determined based on the profile types present in the data (refer to the “Sample Type Configuration” section for more details).

Sample type configuration

Pyroscope server inherently supports standard Go profile types such as cpu, inuse_objects, inuse_space, alloc_objects, and alloc_space. When dealing with software that generates data in pprof format, you may need to supply a custom sample type configuration for Pyroscope to interpret the data correctly.

For an example Python script to ingest a pprof file with a custom sample type configuration, see this Python script.

To ingest pprof data with custom sample type configuration, modify your requests as follows:

  • Set Content-Type to multipart/form-data.
  • Upload the profile data in a form file field named profile.
  • Include the sample type configuration in a form file field named sample_type_config.

A sample type configuration is a JSON object formatted like this:

  "inuse_space": {
    "units": "bytes",
    "aggregation": "average",
    "display-name": "inuse_space_bytes",
    "sampled": false
  "alloc_objects": {
    "units": "objects",
    "aggregation": "sum",
    "display-name": "alloc_objects_count",
    "sampled": true
  "cpu": {
    "units": "samples",
    "aggregation": "sum",
    "display-name": "cpu_samples",
    "sampled": true
  // pprof supports multiple profiles types in one file,
  //   so there can be multiple of these objects

Explanation of sample type configuration fields:

  • units
    • Supported values: samples, objects, bytes
    • Description: Changes the units displayed in the frontend. samples = CPU samples, objects = objects in RAM, bytes = bytes in RAM.
  • display-name
    • Supported values: Any string.
    • Description: This becomes a suffix of the app name, e.g., my-app.inuse_space_bytes.
  • aggregation
    • Supported values: sum, average.
    • Description: Alters how data is aggregated on the frontend. Use sum for data to be summed over time (e.g., CPU samples, memory allocations), and average for data to be averaged over time (e.g., memory in-use objects).
  • sampled
    • Supported values: true, false.
    • Description: Determines if the sample rate (specified in the pprof file) is considered. Set to true for sampled events (e.g., CPU samples), and false for memory profiles.

This configuration allows for customized visualization and analysis of various profile types within Pyroscope.

JFR format

This is the Java Flight Recorder format, typically used by JVM-based profilers, also supported by our Java integration.

When this format is used, some of the query parameters behave slightly different:

  • format should be set to jfr.
  • name contains the prefix of the application name. Since a single request may contain multiple profile types, the final application name is created concatenating this prefix and the profile type. For example, if you send cpu profiling data and set name to my-app{}, it will appear in pyroscope as my-app.cpu{}.
  • units is ignored, and the actual units depends on the profile types available in the data.
  • aggregationType is ignored, and the actual aggregation type depends on the profile types available in the data.

JFR ingestion support uses the profile metadata to determine which profile types are included, which depend on the kind of profiling being done. Currently supported profile types include:

  • cpu samples, which includes only profiling data from runnable threads.
  • itimer samples, similar to cpu profiling.
  • wall samples, which includes samples from any threads independently of their state.
  • alloc_in_new_tlab_objects, which indicates the number of new TLAB objects created.
  • alloc_in_new_tlab_bytes, which indicates the size in bytes of new TLAB objects created.
  • alloc_outside_tlab_objects, which indicates the number of new allocated objects outside any TLAB.
  • alloc_in_new_tlab_bytes, which indicates the size in bytes of new allocated objects outside any TLAB.

JFR with labels

In order to ingest JFR data with dynamic labels, you have to make the following changes to your requests:

  • use an HTTP form (multipart/form-data) Content-Type.
  • send the JFR data in a form file field called jfr.
  • send LabelsSnapshot protobuf message in a form file field called labels.
message Context {
    // string_id -> string_id
    map<int64, int64> labels = 1;
message LabelsSnapshot {
  // context_id -> Context
  map<int64, Context> contexts = 1;
  // string_id -> string
  map<int64, string> strings = 2;

Where context_id is a parameter set in async-profiler


Here’s a sample code that uploads a very simple profile to pyroscope:

printf "foo;bar 100\n foo;baz 200" | curl \
--data-binary @- \
import requests
import urllib.parse
from datetime import datetime

now = round( / 10 * 10
params = {'from': f'{now - 10}', 'name': 'python.example{foo=bar}'}

url = f'http://localhost:4040/ingest?{urllib.parse.urlencode(params)}'
data = "foo;bar 100\n" \
"foo;baz 200", data = data)

Here’s a sample code that uploads a JFR profile with labels to pyroscope:

curl -X POST \
  -F jfr=@profile.jfr \
  -F labels=@labels.pb  \

Querying profile data

There is one primary endpoint for querying profile data: GET /pyroscope/render.

The search input is provided via query parameters. The output is typically a JSON object containing one or more time series and a flame graph.

Query parameters

Here is an overview of the accepted query parameters:

querycontains the profile type and label selectorsrequired
fromUNIX time for the start of the search windowrequired
untilUNIX time for the end of the search windowoptional (default is now)
formatformat of the profiling dataoptional (default is json)
maxNodesthe maximum number of nodes the resulting flame graph will containoptional (default is max_flamegraph_nodes_default)
groupByone or more label names to group the time series by (doesn’t apply to the flame graph)optional (default is no grouping)


The query parameter is the only required search input. It carries the profile type and any labels we want to use to narrow down the output. The format for this parameter is similar to that of a PromQL query and can be defined as:

<profile_type>{<label_name>="<label_value>", <label_name>="<label_value>", ...}

Here is a specific example:


In a Kubernetes environment, a query could also look like:

process_cpu:cpu:nanoseconds:cpu:nanoseconds{namespace="dev", container="my_application_name"}


Refer to the profiling types documentation for more information and profile-metrics.json for a list of valid profile types.

from and until

The from and until parameters determine the start and end of the time period for the query. They can be provided in absolute and relative form.

Absolute time

This table details the options for passing absolute values.

Date20231223Format: YYYYMMDD
Unix Time seconds1577836800
Unix Time milliseconds1577836800000
Unix Time microseconds1577836800000000
Unix Time nanoseconds1577836800000000000

Relative time

Relative values are always expressed as offsets from now.

3 hours agonow-3h
30 minutes agonow-30m
2 days agonow-2d
1 week agonow-7d or now-1w

Note that a single offset has to be provided, values such as now-3h30m will not work.


The from and until parameters are subject to validation rules related to max_query_lookback and max_query_length server parameters. You can find more details on these parameters in the limits section of the server configuration docs.

  • If max_query_lookback is configured andfrom is before now - max_query_lookback, from will be set to now - max_query_lookback.
  • If max_query_lookback is configured and until is before now - max_query_lookback the query will not be executed.
  • If max_query_length is configured and the query interval is longer than this configuration, the query will no tbe executed.


The format can either be:

  • json, in which case the response will contain a JSON object
  • dot, in which case the response will be text containing a DOT representation of the profile

See the Query output section for more information on the response structure.


The maxNodes parameter truncates the number of elements in the profile response, to allow tools (for example, a frontend) to render large profiles efficiently. This is typically used for profiles that are known to have large stack traces.

When no value is provided, the default is taken from the max_flamegraph_nodes_default configuration parameter. When a value is provided, it is capped to the max_flamegraph_nodes_max configuration parameter.


The groupBy parameter impacts the output for the time series portion of the response. When a valid label is provided, the response contains as many series as there are label values for the given label.


Pyroscope supports a single label for the group by functionality.

Query output

The output of the /pyroscope/render endpoint is a JSON object based on the following schema:

type FlamebearerProfileV1 struct {
	Flamebearer FlamebearerV1                  `json:"flamebearer"`
	Metadata FlamebearerMetadataV1             `json:"metadata"`
	Timeline *FlamebearerTimelineV1            `json:"timeline"`
	Groups   map[string]*FlamebearerTimelineV1 `json:"groups"`


The flamebearer field contains data in a form suitable for rendering a flame graph. Data within the flamebearer is organized in separate arrays containing the profile symbols and the sample values.


The metadata field contains additional information that is helpful to interpret the flamebearer data such as the unit (nanoseconds, bytes), sample rate and more.


The timeline field represents the time series for the profile. Pyroscope pre-computes the step interval (resolution) of the timeline using the query interval (from and until). The minimum step interval is 10 seconds.

The raw profile sample data is down-sampled to the step interval (resolution) using an aggregation function. Currently only sum is supported.

A timeline contains a start time, a list of sample values and the step interval:

  "timeline": {
    "startTime": 1577836800,
    "samples": [
    "durationDelta": 10


The groups field is only populated when grouping is requested by the groupBy query parameter. When this is the case, the groups field has an entry for every label value found for the query.

This example groups by a cluster:

  "groups": {
    "eu-west-2": { "startTime": 1577836800, "samples": [ 200, 300, 500 ] },
    "us-east-1": { "startTime": 1577836800, "samples": [ 100, 200, 400 ] }

Alternative query output

When the format query parameter is dot, the endpoint responds with a DOT format data representing the queried profile. This can be used to create an alternative visualization of the profile.

Example queries

This example queries a local Pyroscope server for a CPU profile from the pyroscope service for the last hour.

curl \

Here is the same query made more readable:

curl --get \
  --data-urlencode "query=process_cpu:cpu:nanoseconds:cpu:nanoseconds{service_name=\"pyroscope\"}" \
  --data-urlencode "from=now-1h" \

Here is the same example in Python:

import requests

application_name = 'my_application_name'
query = f'process_cpu:cpu:nanoseconds:cpu:nanoseconds{{service_name="{application_name}"}}'
query_from = 'now-1h'
url = f'http://localhost:4040/pyroscope/render?query={query}&from={query_from}'


See this Python script for a complete example.

Profile CLI

The profilecli tool can also be used to interact with the Pyroscope server API. The tool supports operations such as ingesting profiles, querying for existing profiles, and more. Refer to the Profile CLI page for more information.