Profiling Python applications

This page describes how to modify your Python application to capture profiling data and have that data sent to your Google Cloud project. For general information about profiling, see Profiling concepts.

Profile types for Python:

• CPU time
• Wall time (main thread)

Supported Python language versions:

• Python 3.6 or higher.

Supported profiling agent versions:

• The most recent release of the agent is supported. In general, releases older than one year aren't supported. We recommend that you use the most recently released version of the agent.

Supported operating systems:

• Linux. Profiling Python applications is supported for Linux kernels whose standard C library is implemented with glibc or with musl. For configuration information specific to Linux Alpine kernels, see Running on Linux Alpine.

Supported environments:

• Compute Engine
• Google Kubernetes Engine (GKE)
• App Engine flexible environment
• App Engine standard environment (requires Python 3 runtime environment)
• Outside of Google Cloud (For information on the additional configuration requirements, see Profiling applications running outside of Google Cloud.)

Enabling the Profiler API

Before you use the profiling agent, ensure that the underlying Profiler API is enabled. You can check the status of the API and enable it if necessary by using either the Google Cloud CLI or the Google Cloud console:

gcloud CLI

1. If you have not already installed the Google Cloud CLI on your workstation, see the Google Cloud CLI documentation.

2. Run the following command:

   gcloud services enable cloudprofiler.googleapis.com
   

For more information, see gcloud services.

Google Cloud console

1. Enable the required API.

   Enable the API

2. If API enabled is displayed, then the API is already enabled. If not, click the Enable button.

Using Cloud Profiler

For best practices using Python, go to Setting up a Python development environment.

Compute Engine

For Compute Engine, do the following:

1. Install the C/C++ compiler and development tools:

   sudo apt-get install -y build-essential
   

2. Install pip:

   sudo apt-get install -y python3-pip
   

3. Install the Profiler package:

   pip3 install google-cloud-profiler
   

4. Import the googlecloudprofiler module and call the googlecloudprofiler.start function as early as possible in your initialization code:

   import googlecloudprofiler
   
   
   def main():
       # Profiler initialization. It starts a daemon thread which continuously
       # collects and uploads profiles. Best done as early as possible.
       try:
           googlecloudprofiler.start(
               service="hello-profiler",
               service_version="1.0.1",
               # verbose is the logging level. 0-error, 1-warning, 2-info,
               # 3-debug. It defaults to 0 (error) if not set.
               verbose=3,
               # project_id must be set if not running on GCP.
               # project_id='my-project-id',
           )
       except (ValueError, NotImplementedError) as exc:
           print(exc)  # Handle errors here

   You must specify the service parameter in your start function. To filter by the application version in the Profiler interface, specify the service_version parameter. For troubleshooting and exception information, see Troubleshooting.

GKE

Note: If you enable Workload Identity Federation for GKE on a cluster, then ensure that you add the role Cloud Profiler Agent (roles/cloudprofiler.agent) to the Kubernetes service account. For more information, see Use Workload Identity Federation for GKE.

For GKE, do the following:

1. Modify your Dockerfile to install the Profiler package:

   FROM python:3
   ...
   RUN apt-get update && apt-get install -y build-essential python3-pip
   RUN pip3 install google-cloud-profiler
   

2. Import the googlecloudprofiler module and call the googlecloudprofiler.start function as early as possible in your initialization code:

   import googlecloudprofiler
   
   
   def main():
       # Profiler initialization. It starts a daemon thread which continuously
       # collects and uploads profiles. Best done as early as possible.
       try:
           googlecloudprofiler.start(
               service="hello-profiler",
               service_version="1.0.1",
               # verbose is the logging level. 0-error, 1-warning, 2-info,
               # 3-debug. It defaults to 0 (error) if not set.
               verbose=3,
               # project_id must be set if not running on GCP.
               # project_id='my-project-id',
           )
       except (ValueError, NotImplementedError) as exc:
           print(exc)  # Handle errors here

   You must specify the service parameter in your start function. To filter by the application version in the Profiler interface, specify the service_version parameter. For troubleshooting and exception information, see Troubleshooting.

Flexible Environment

For App Engine flexible environment, do the following:

1. Add google-cloud-profiler to your requirements.txt file.

2. Import the googlecloudprofiler module and call the googlecloudprofiler.start function as early as possible in your initialization code.

For App Engine, the service and service_version are derived from your operating environment. For troubleshooting and exception information, see Troubleshooting.

Standard Environment

For App Engine standard environment, which requires that you use the Python 3 runtime environment, do the following:

1. Add google-cloud-profiler to your requirements.txt file.

2. Import the googlecloudprofiler module and call the googlecloudprofiler.start function as early as possible in your initialization code.

For App Engine, the service and service_version are derived from your operating environment. For troubleshooting and exception information, see Troubleshooting.

start function

The googlecloudprofiler.start function creates a daemon thread which continuously collects and uploads profiles. You should call start one time, and as early as possible in your application.

Parameter	Description
service^1	(Required) The name for the service being profiled. For restrictions on the service name, see Service name and version arguments.
service_version^1	(Optional) The version of the service being profiled. For restrictions on the service version, see Service name and version arguments.
verbose	(Optional) The logging level. For details on the logging levels, see Agent logging. Default value is 0 (Error).
project_id^2	(Optional) Your Google Cloud project ID.
disable_cpu_profiling	(Optional) To disable CPU time profiling, set disable_cpu_profiling=True. This parameter is supported only for Python version 3.2 and higher. For all other Python versions, CPU time profiling isn't supported and this parameter is ignored. Default value is False.
disable_wall_profiling	(Optional) To disable Wall profiling, set disable_wall_profiling=True. This parameter is supported for Python 3.6 and higher. For all other Python versions, Wall profiling isn't supported and this parameter is ignored. For restrictions on the start function when Wall profiling is enabled, see Limitations. Default value is False.

¹ For only Compute Engine and GKE. For App Engine, the value is derived from the environment.
² For Google Cloud, the value is derived from the environment. For non-Google Cloud environments, you must provide a value. For information, see Profiling applications running outside Google Cloud.

Analyzing data

After Profiler has collected data, you can view and analyze this data using the Profiler interface.

In the Google Cloud console, go to the Profiler page:

Go to Profiler

You can also find this page by using the search bar.

Service name and version arguments

When you load the Profiler agent, you specify a service-name argument and an optional service-version argument to configure it.

Note: In the App Engine flexible environment, you do not have to specify these arguments; they are derived from the environment.

The service name lets Profiler collect profiling data for all replicas of that service. The profiler service ensures a collection rate of one profile per minute, on average, for each service name across each combination service versions and zones.

For example, if you have a service with two versions running across replicas in three zones, the profiler will create an average of 6 profiles per minute for that service.

If you use different service names for your replicas, then your service will be profiled more often than necessary, with a correspondingly higher overhead.

When selecting a service name:

• Choose a name that clearly represents the service in your application architecture. The choice of service name is less important if you only run a single service or application. It is more important if your application runs as a set of micro-services, for example.

• Make sure to not use any process-specific values, like a process ID, in the service-name string.

• The service-name string must match this regular expression:

  ^[a-z0-9]([-a-z0-9_.]{0,253}[a-z0-9])?$

A good guideline is to use a static string like imageproc-service as the service name.

The service version is optional. If you specify the service version, Profiler can aggregate profiling information from multiple instances and display it correctly. It can be used to mark different versions of your services as they get deployed. The Profiler UI lets you filter the data by service version; this way, you can compare the performance of older and newer versions of the code.

The value of the service-version argument is a free-form string, but values for this argument typically look like version numbers, for example, 1.0.0 or 2.1.2.

Agent logging

By default, the profiling agent logs messages with a severity level of error. To configure the agent to log messages with lower severity levels, specify the verbose parameter when starting the agent. There are four supported values for verbose:

• 0 : Error
• 1 : Warning
• 2 : Informational
• 3 : Debug

If you set the verbose parameter to 1 in your call to start, then messages with a severity level of Warning or Error are logged, while Informational and Debug messages are ignored.

To log all messages, set verbose to 3 when starting the agent:

googlecloudprofiler.start(service='service_name', verbose=3)

Troubleshooting

This section lists limitations, exceptions, and known issues that are specific to profiling Python applications. See Troubleshooting for help with common problems.

Limitations

Profile type

Restrictions and limitations

Wall time

Main-thread profiling only. The profile start function must be called from the main thread. The Profiler signal handler executes only on the main thread. If the main thread cannot execute, no profiling data is captured. Note: Don't enable Wall profiling if your application contains a large number of threads. Use of Wall profiles in this configuration results in slow performance and no profile data.

Exceptions

Error	Cause	Solution
NotImplementedError thrown during start	Application executed in a non-Linux environment.	Execute your application in a Linux environment.
ValueError thrown during start	The start function arguments are invalid, necessary information can't be determined from the environment variables and arguments, or profiling if both CPU time profiling and Wall time profiling are disabled.	Ensure the service name and version meet the requirements defined in Service name and version arguments. If Wall profiling is enabled, make sure start is called from the main thread. Ensure that you are using a supported version of Python and that CPU time or Wall time profiling is enabled. For more information, see start function. Check that you have specified the project_id parameter to start if you are running outside of Google Cloud. For more information, see start function.

Known issues

Behavior	Cause	Solution
You don't have any profile data or you enabled a new profile type and are missing profile data.	Common causes are related to configuration.	See Troubleshooting.
You are using uWSGI and you don't have CPU time and Wall profile data for all processes.	When uWSGI uses multiple workers to handle requests, the default behavior is to perform application initialization only in the primary ("master") process. The forked processes don't perform the initialization sequence. If you configure the profiling agent in your application's initialization sequence, for example, in AppConfig.ready() in a Django application, then the effect is that the profiling agent isn't configured for the forked processes.	To perform application initialization in all worker processes, set the flag lazy-apps to true. See the next topic in this table for a related issue.
You are using uWSGI and you don't have Wall profile data, but you do have CPU time profile data.	The Wall profiler depends on the Python signal module. When the Python interpreter is compiled with thread support, the default configuration disables custom signal handling for forked processes.	For uWSGI applications, enable the custom signal handling by setting the flag py-call-osafterfork to true. See the previous topic in this table for a related issue.
After enabling the profiler, the error log contains new entries: BlockingIOError: [Errno 11] Resource temporarily unavailable Exception ignored when trying to write to the signal wakeup fd GitHub issue	Your application registered with the signal wakeup file descriptor, signal.set_wakeup_fd. By default, if the file descriptor's buffer fills then a warning is logged to stderr. When Cloud Profiler collects profiles it triggers signals with high frequency. This behavior can cause the file descriptor's buffer to become full.	If your application can run safely when signals are lost, then you can use Cloud Profiler. If you are using Python 3.7 or later and want to disable the warning messages, then pass warn_on_full_buffer=False as a parameter to signal.set_wakeup_fd. If your application can't safely run when signals are lost, then we recommend that you stop using Cloud Profiler. Continued use might cause loss of signal numbers and excessive entries in the error log.

Running with Linux Alpine

Alpha

This product or feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA products and features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

The Python profiling agent for Linux Alpine is supported only for Google Kubernetes Engine configurations.

To build the Python profiling agent, you must install the package build-base. To use the Python profiling agent on Alpine without installing additional dependencies on to the final Alpine image, you can use a two-stage build and compile the Python profiling agent in the first stage. For example, the following Docker image uses a multi-stage build to compile and install the Python profiling agent:

FROM python:3.7-alpine as builder

# Install build-base to allow for compilation of the profiling agent.
RUN apk add --update --no-cache build-base

# Compile the profiling agent, generating wheels for it.
RUN pip3 wheel --wheel-dir=/tmp/wheels google-cloud-profiler


FROM python:3.7-alpine

# Copy over the directory containing wheels for the profiling agent.
COPY --from=builder /tmp/wheels /tmp/wheels

# Install the profiling agent.
RUN pip3 install --no-index --find-links=/tmp/wheels google-cloud-profiler

# Install any other required modules or dependencies, and copy an app which
# enables the profiler as described in "Enable the profiler in your
# application".
COPY ./bench.py .

# Run the application when the docker image is run, using either CMD (as is done
# here) or ENTRYPOINT.
CMD python3 -u bench.py

Authentication error

If you use Docker images that run with Linux Alpine (such as golang:alpine or just alpine), you might see the following authentication error:

connection error: desc = "transport: authentication handshake failed: x509: failed to load system roots and no roots provided"

Note that to see the error you must have agent logging enabled.

The error indicates that the Docker images with Linux Alpine don't have the root SSL certificates installed by default. Those certificates are necessary for the profiling agent to communicate with the profiler API. To resolve this error, add the following apk command to your Dockerfile:

FROM alpine
...
RUN apk add --no-cache ca-certificates

You then need to rebuild and redeploy your application.

What's next

• Select the profiles to analyze
• Interact with the flame graph
• Filter the flame graph
• Focus the flame graph
• Compare profiles