PROBLEM DESCRIPTION
I am trying to create a custom managed VM for Google App Engine that behaves identically to the standard python27 managed VM provided by Google. (I'm doing this as a first step to adding a C++ library to the runtime).
From google documentation, the following Dockerfile specifies the standard python27 runtime:
FROM gcr.io/google_appengine/python-compat
ADD . /app
I have verified that this is the right Dockerfile by examining the one generated by gcloud preview app run
when using the standard python27 runtime. It is identical to this.
But when I run my application with this Dockerfile using dev_appserver.py
or with gcloud preview app run
I get an error saying:
The --custom_entrypoint flag must be set for custom runtimes
I am using the latest versions of gcloud (1.9.86, with app-engine-python component version 1.9.28) and the standalone python app engine SDK (1.9.28). I had the same problem with earlier versions, so I updated to the latest.
THINGS I HAVE TRIED:
gcloud preview app run --help
has the following to say about --custom-entrypoint
:
--custom-entrypoint CUSTOM_ENTRYPOINT
Specify an entrypoint for custom runtime modules. This is required when
such modules are present. Include "{port}" in the string (without
quotes) to pass the port number in as an argument. For instance:
--custom_entrypoint="gunicorn -b localhost:{port} mymodule:application"
I am not sure what to make of this. Should the docker image not already contain an ENTRYPOINT? Why am I being required to provide one in addition? Also, what should the entrypoint be for the gcr.io/google_appengine/python-compat
image be? Google provides no documentation for this.
I have tried a meaningless --custom-entrypoint="echo"
, which silences the error, but the application does not response to any HTTP requests.
The two other relevant stackoverflow questions I have found have not helped. The accepted answers seem to suggest that this is a bug in the SDK that was resolved. But I have tried it in two versions of the SDK, including the latest, and I still have the problem.
- How to fix “
The --custom_entrypoint flag must be set for custom runtimes
”? - Google Managed VM error - custom entry point
STEPS TO REPRORDUCE:
To highlight my problem, I have created a trivial application that generates the error. It consists of just three files:
app.yaml
:
module: default
version: 1
runtime: custom
api_version: 1
threadsafe: true
vm: true
handlers:
- url: /.*
script: wsgi.app
Dockerfile
:
FROM gcr.io/google_appengine/python-compat
ADD . /app
This Dockerfile
is the same one that is used for the python27 runtime (and in fact literally copy-pasted from the Dockerfile generated by gcloud preview app run
when using the python27 runtime), so this should be identical to setting runtime: python27
.
wsgi.py
:
import webapp2
class Hello(webapp2.RequestHandler):
def get(self):
self.response.write(u'Hello')
app = webapp2.WSGIApplication([('/Hello', Hello)], debug=True)
When I run dev_appserver.py app.yaml
in the directory containing these three files however, I get the following error:
Traceback (most recent call last):
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/dev_appserver.py", line 83, in <module>
_run_file(__file__, globals())
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/dev_appserver.py", line 79, in _run_file
execfile(_PATHS.script_file(script_name), globals_)
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/devappserver2.py", line 1033, in <module>
main()
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/devappserver2.py", line 1026, in main
dev_server.start(options)
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/devappserver2.py", line 818, in start
self._dispatcher.start(options.api_host, apis.port, request_data)
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/dispatcher.py", line 194, in start
_module.start()
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/module.py", line 1555, in start
self._add_instance()
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/module.py", line 1707, in _add_instance
expect_ready_request=True)
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/custom_runtime.py", line 73, in new_instance
assert self._runtime_config_getter().custom_config.custom_entrypoint
File "/home/vagrant/google-cloud-sdk/platform/google_appengine/google/appengine/tools/devappserver2/module.py", line 383, in _get_runtime_config
raise ValueError('The --custom_entrypoint flag must be set for '
ValueError: The --custom_entrypoint flag must be set for custom runtimes
After trying to get my custom VM working with dev_appserver for the better part of a day, the accepted answer to this post came as a pretty unpleasant surprise. But I figured deploying a dev server couldn't be that much of a hassle because, after all, the VM is a standard Docker image.
Well there did turn out the be a few issues that prevent a straight deployment from working. I've provided a summary of these issues below as well as how I solved them. I may have missed some incompatibilities between the Docker and App Engine environments (especially with the many aspects of App Engine that my project didn't use) but hopefully it's enough to get people up and running.
Sources of trouble
First, I found the python environment run in the Compute Engine VMs to be a bit more lenient than a normal VM environment (e.g. packages like
webapp2
are always available). Consequently, deploying to the less forgiving Docker container environment surfaced some latent errors in my project.That being said, there are some differences in the environments that necessitate some tweaks even if your project is flawless:
Problem: gunicorn (or the server of your choice) must be installed on the Docker container's path.
gunicorn
in my project'srequirements.txt
file. Unfortunately, I was installing all of these dependencies usingpip install -t ...
which is only able to install source. As a result, there was nogunicorn
binary on the image let alone on thePATH
.Solution: Explicitly
pip install gunicorn
google.appengine.*
packages are not available from the App Engine base Docker image nor are they available via pip (AFAICT).google.appengine.ext.vendor
is the recommended interface for importing third-party libraries into your App Engine app.How to get the script
The script to build and deploy the VM docker image to a docker container running locally is available here.
For an working example, check out my project.
Let me know if you have a comment, a feature request, or if you write prettier bash than me (I feel I've set the bar comfortably low on that count).
EDIT 1: The solution posted by user862857 makes use of Docker itself to build images from the Dockerfile and run them in containers. This is also a good approach to running Managed VMs and Custom Runtimes in development contexts.
The accepted answer doesn't seem correct. A github README should not trump the official docs for authoritativeness when dealing with a rapidly-evolving Beta product. It's perfectly possible to a
runtime: custom
app in the dev environment, using the Dockerfile mentioned in OP's post,using the
--runtime=python-compat
flag. They would need to catch requests to/_ah/start
and/_ah/health
, though. Attempt to run the following command, given the following files, and see for yourself:devserver command
app.yaml
Dockerfile
main.py
Send a request to
/?domain=stackoverflow.com
to see this in action.N.B.
If they wished to entirely decouple from the python-compat runtime and simply deploy/test a python WSGI app via, they could also use the
--custom_entrypoint
flag, so long as they had a command which would start running the appropriate WSGI app on a suitable port (such a command would be uwsgi or gunicorn).UPDATE
THIS MAY NO LONGER BE ACCURATE. SEE NICK'S ANSWER.
(Though I could not get that working. But I did not try very hard)
There is a completely undocumented but absolutely essential piece of information w.r.t. custom managed VMs:
THEY CANNOT BE RUN ON THE DEVELOPMENT SERVER!
If you think this crucial fact would be mentioned anywhere sane, like say, the documentation page for custom managed VMs, or for
dev_appserver.py
, or even as an error message when runningdev_appserver.py
, then you are giving Google far too much credit.The only place where I can find any kind of statement about this is in the Readme file of the appengine-java-vm-guestbook-extras demo on github (seriously):
Google does not care to:
I hope this answer saves some sorry developer from the days of torment I suffered because of this.