Celery Flower Security in Production

Question

I am looking to use Flower (https://github.com/mher/flower) to monitor my Celery tasks in place of the django-admin as reccomended in their docs (http://docs.celeryproject.org/en/latest/userguide/monitoring.html#flower-real-time-celery-web-monitor). However, because I am new to this I am a little confused about the way Flower's page is only based on HTTP, and not HTTPS. How can I enable security for my Celery tasks such that any old user can't just visit the no-login-needed website http://flowerserver.com:5555 and change something?

I have considered Celery's own documentation on this, but they unfortunately there is no mention of how to secure Flower's api or web ui. All it says: [Need more text here]

Thanks!

Update: My question is in part a duplicate of here: How do I add authentication and endpoint to Django Celery Flower Monitoring?

However, I clarify his question here by asking how to run it using an environment that includes nginx, gunicorn, and celery all on the same remote machine. I too am wondering about how to set up Flower's outside accessible url, but also would prefer something like https instead of http if possible (or some way of securing the webui and accessing it remotely). I also need to know if leaving Flower running is a considerable security risk for anyone who may gain access to Flower's internal API and what the best way for securing this could be, or if it should just be disabled altogether and used just on an as-needed basis.

Solution 1

You can run flower with --auth flag, which will authenticate using a particular google email:

celery flower [email protected]

Edit 1:

New version of Flower requires couple more flags and a registered OAuth2 Client with Google Developer Console:

celery flower \
    [email protected] \
    --oauth2_key="client_id" \
    --oauth2_secret="client_secret" \
    --oauth2_redirect_uri="http://example.com:5555/login"

oauth2_redirect_uri has to be the actual flower login url, and it also has to be added to authorized redirect url's in Google Development Console.

Unfortunately this feature doesn't work properly in current stable version 0.7.2, but it is now fixed in development version 0.8.0-dev with this commit.

Edit 2:

You can configure Flower using basic authentication:

celery flower --basic_auth=user1:password1,user2:password2

Then block 5555 port for all but localhost and configure reverse proxy for nginx or for apache:

ProxyRequests off
ProxyPreserveHost On
ProxyPass / http://localhost:5555

Then make sure proxy mod is on:

sudo a2enmod proxy
sudo a2enmod proxy_http

In case you can't set it up on a separate subdomain, ex: flower.example.com (config above), you can set it up for example.com/flower:

run flower with url_prefix:

celery flower --url_prefix=flower --basic_auth=user1:password1,user2:password2

in apache config:

ProxyPass /flower http://localhost:5555

Of course, make sure SSL is configured, otherwise there is no point :)

Solution 2

I have figured out it using proxy on Django side https://pypi.org/project/django-revproxy/. So Flower is hidden behind Django auth which is more flexible than basic auth. And you don't need rewrite rule in NGINX.

Flower 0.9.5 and higher

URL prefix must be moved into proxy path: https://github.com/mher/flower/pull/766

urls.py

urlpatterns = [
    FlowerProxyView.as_url(),
    ...
]

views.py

class FlowerProxyView(UserPassesTestMixin, ProxyView):
    # `flower` is Docker container, you can use `localhost` instead
    upstream = 'http://{}:{}'.format('flower', 5555)
    url_prefix = 'flower'
    rewrite = (
        (r'^/{}$'.format(url_prefix), r'/{}/'.format(url_prefix)),
     )

    def test_func(self):
        return self.request.user.is_superuser

    @classmethod
    def as_url(cls):
        return re_path(r'^(?P<path>{}.*)$'.format(cls.url_prefix), cls.as_view())

Flower 0.9.4 and lower

urls.py

urlpatterns = [
    re_path(r'^flower/?(?P<path>.*)$', FlowerProxyView.as_view()),
    ...
]

views.py

from django.contrib.auth.mixins import UserPassesTestMixin
from revproxy.views import ProxyView


class FlowerProxyView(UserPassesTestMixin, ProxyView):
    # `flower` is Docker container, you can use `localhost` instead
    upstream = 'http://flower:5555'

    def test_func(self):
        return self.request.user.is_superuser

Solution 3

I wanted flower on a subdirectory of my webserver, so my nginx reverse proxy configuration looked like this:

location /flower/ {
    proxy_pass http://localhost:5555/;
    proxy_redirect off;
    proxy_set_header Host $host;
    proxy_set_header X-Real-IP $remote_addr;
    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header X-Forwarded-Protocol $scheme;
    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";
    proxy_http_version 1.1;

    auth_basic  "Restricted";
    auth_basic_user_file  /etc/nginx/.htpasswd;
}

Now I can get to flower (password-protected) via www.example.com/flower

Most of this is derived from the Flower documentation page about configuring an nginx reverse proxy:

http://flower.readthedocs.org/en/latest/reverse-proxy.html

Solution 4

I followed @petr-přikryl's approach using a proxy view. However I couldn't get it to verify authentication (I don't think test_func is ever called). Instead I chose to embed this in the Django Admin views and use AdminSite.admin_view() (as described here) to wrap the view with Django Admin authentication.

Specifically, I made the following changes:

# Pipfile
[packages]
...
django-revproxy="*"

# admin.py
class MyAdminSite(admin.AdminSite):
    # ...
    def get_urls(self):
        from django.urls import re_path

        # Because this is hosted in the root `urls.py` under `/admin` this 
        # makes the total prefix /admin/flower
        urls = super().get_urls()
        urls += [
            re_path(
                r"^(?P<path>flower.*)$",
                self.admin_view(FlowerProxyView.as_view()),
            )
        ]
        return urls

# views.py
from __future__ import annotations

from django.urls import re_path

from revproxy.views import ProxyView


class FlowerProxyView(ProxyView):
    # Need `/admin/` here because the embedded view in the admin app drops the
    # `/admin` prefix before sending the URL to the ProxyView
    upstream = "http://{}:{}/admin/".format("localhost", 5555)

Lastly, we need to make sure that --url_prefix is set when running flower, so I set it to run like this in our production and dev environments:

celery flower --app=my_app.celery:app --url_prefix=admin/flower

Solution 5

Yep there's not auth on flower, since it's just talking to the broker, but if you run it over SSL then basic auth should be good enough.

View more solutions

Comments

[mh00h]

I am referring to the webui that Flower provides. I was not under the impression that a user name and password are required to connect to the web interface that Flower provides. Else, even if there is one, the web ui talks over HTTP, not HTTPS, so I question the security aspect of controlling Celery through a non-secure web interface.

[adam]

Why would a login be necessary to access Flower UI? In order for Flower to work you already have access to the Celery backend broker, no? If you're afraid of snooping your network packets, set up access control so only internal machines can access the server runs Celery. Or in our practice, only a handful of people who can log into the production servers can run Flower.

[mh00h]

Ah, this is where I'm new to this. I am using celery to run time-intensive tasks on the same (remote) machine as my django server. To clarify, you are suggesting I set up iptables to allow only the localhost access to server:5555, and then use ssh port forwarding to access Flower? I had thought there would be a more practical https option w/ a login, but no? Else as I understand it, once Flower is set up to log into the broker w/ credentials, anyone can access the ui itself, no?

[mh00h]

Also, are you saying that Flower is not something that one would never leave running all the time on a production server, even given the resources to do so?

[adam]

The HTTPS option might be technically possible but there is much manual work to set it up (registering certificates, changing Flower source code to use that certificates if that's possible at all). In my opinion, setting up iptables and port blocking is much easier alternative. And I'd rather lock down the machine runs Flower and ssh into the machine to run it. Essentially you don't want to allow access to Flower except the request comes from localhost.

[adam]

You can leave Flower running on the same server as Celery or Web Server. But I wouldn't recommend it. In my practice we have separate AWS instances hosting different applications (web server, celery, etc). So in case one of them fails, we can easily spin up another instance without affecting one another.

[JiminyCricket]

my AWS machine's permissions have locked down port 6379 only to other production infrastructure. I have access to the actual machine Redis is running on (ssh key). Is it possible to somehow port forward through that? (I'm new to port-forwarding)

[Mikle]

This is an amazingly well written answer, thanks. Now whether flower is worth all the above mess - that's a different question.

[Raj]

Just wanted to point out that flower seems to have a localhost option now: celery flower -A proj --address=127.0.0.1 --port=5555

[luvwinnie]

just wondering is that have a way to clear the basic authentication with timeout method or something? i do like to clear the authentication after a time..

[Scott Stafford]

@Leow see stackoverflow.com/questions/8034899/…

[GKeps]

This is great. The one change I made to work on the server was upstream = 'http://{}:{}'.format('localhost', 5555)

[Denys Halenok]

Worth mentioning that there's a security vulnerability in OAuth authentication. github.com/mher/flower/issues/1217