Feature/use gunicorn in dev mode #19
Conversation
The meinheld worker seems to be unmaintained and it doesn't work with gunicorn's reload mechanism.
sergiofenoll
force-pushed
the
feature/use-gunicorn-in-dev-mode
branch
from
c1bc262
to
5981fc6
Compare
Having inotify installed enables gunicorn's reload mechanism to survive through syntax errors and such in the service code.
Using gunicorn's reload mechanism, importing the code in a try-except meant that the reloader would not pick up changes if the initial code import threw an error. The reloader checks which files it needs to listen to notifications for based on its imports. It seems that the combination of a throwing import inside a try block makes it so the reloader doesn't register the service entrypoint as a file it needs to listen to.
We no longer use Flask in dev mode, so this if block is never entered.
|
I support this, since it only seems to make things better! The Gunicorn/Meinheld setup was introduced here, but was no more than a choice for a WSGI server rather than a specific commitment to Meinheld workers. Using this PR image in some of our frequently used dev setups will be the best way to test this further if we want to build more confidence before merging. |
|
@sergiofenoll (or someone with access to this repo): could you set up a build? that will make it easier to test 👼 |
|
@nvdk I don't think I can enable the build under the semtech namespace (I think someone with write rights for the organization/repo would have to recreate the PR and use that to push). I have enabled the Woodpecker on my fork though, you should be able to use this image: |
|
This is something I've noticed while using this locally: while We could circumvent this by setting |
This PR changes the template so that the same mechanism (gunicorn) that is used for running in production is also used in development mode. This should ensure that the behaviour of services using this template is consistent between development mode and production mode.
Why
With the previous setup, in development mode the template would use Flask to serve the code, which meant that route handlers can take as long as you want, whereas in production mode there's a timeout for workers set by Gunicorn, so a route handler can't take more than a certain amount of time before it gets aborted.
This disparity is quite annoying because you might make something that works perfectly locally in dev-mode and then as soon as you deploy it you notice it breaks with errors you can't seem to reproduce locally.
How
Gunicorn provides a
--reloadoption so that we don't lose the live code reload functionality of the template, but to get this working it was necessary to replace the worker class. It seems that the workers we usedmeinhelddon't work well alongside gunicorn's reload mechanism: notice that in this issue the reporter is using meinheld workers. Meinheld itself also seems quite unmaintained, as the last commit for it was over 3 years ago and its official website has been squatted, so it might make sense to replace the worker class regardless.Warning
I know next to nothing about Gunicorn's architecture and the working of WSGI servers. I'm not really sure if the suggested worker is "better" than what we're currently using. So thorough testing with existing services is needed before merging this.
Nice extras
A lot of our services make use of
apschedulerto build cronjobs. Due to how Flask's debug mode interacted with this library, in development mode usage of this library would result in 2 schedulers being instantiated. Every cron job would run twice. Although this didn't happen in production mode, it could be quite annoying/confusing when developing using this template. Switching to gunicorn for development mode has inadvertently resolved this. For some extra context see this StackOverflow post.Encountered issues
Gunicorn's
--reloadmechanism seems to be broken for certain errors: benoitc/gunicorn#2244This would mean that the reloader stops working if there's e.g. a syntax error in your file while developing a service, which is quite likely to happen. To be determined if this gets fixed if we use a newer version of gunicorn (the currently used version is 2 years old).
Addinginotifyenables that mechanism for the reloader, in which case it can survive the mentioned errors.By changing the service-code import mechanism we can circumvent this issue. The reloader tracks the files it needs to listen to based on imports. Because we used to set the import inside a try-except block, the reloader wouldn't correctly pick up file changes if importing the file throws. By moving this import outside of the try-except block, we keep the reloader happy and also "fix" #12: a missing dependency or a syntax error in service-code will not silently print a single log line and set up a web server with no endpoints, instead it crashes the service.