Use nginx as centralized reverse proxy for API services

date:

2017-12-02 00:00

tags:

nginx, load balancing, wsgi, API

In the previous cycle, most OpenStack API services provided a WSGI application which is being serviced through uWSGI. The aim of this spec is to outline a plan for taking advantage of the greater flexibility uWSGI provides and making use of nginx as a centralized reverse proxy.

https://blueprints.launchpad.net/openstack-ansible/+spec/centralized-nginx

Problem description

Within most roles, OpenStack API services are now deployed and served through uWSGI. Some however, also include the installation of a web server proxy in front of that. The web server is currently assumed to be installed on the host in isolation and managed through the installing role. This causes issues with converged installations, such as an all bare metal scenario. SSL encryption of service requests is also currently difficult because of this inconsistency between individual service deployments. For the most part, SSL termination in OpenStack-Ansible deployments is expected to be handled at the load balancer but deployers may require additional controls over the handling of encrypted traffic.

Proposed change

Remove the installation of nginx from any OpenStack role that includes it today and instead deploy it separately as a shared reverse proxy with individual sites configured for each OpenStack API service. The OpenStack roles will only need to provide site configurations for an existing nginx installation.

For management of the uWSGI backends that nginx will be proxying, install a uWSGI FastRouter alongside nginx. The FastRouter will create a shared socket that nginx can connect to, and a subscription server for the applications to subscribe to and automatically provide load balancing for.

Alternatives

nginx could instead be deployed alongside each OpenStack service, but that could have an effect on performance due to the additional processes running on the host and wouldn’t address the duplication of tasks and conflicts caused when multiple roles are attempting to manage the same web server on the same physical host or container.

The load balancer backends could use uWSGI http/https directly, most projects recommended a dedicated web server however. With the proposed change, there is a clear separation between the handling of http requests and running of Python code. Load balancer configurations can also be simplified since all OpenStack API backends would use the same set of nginx hosts, instead of requiring knowledge of individual containers. And for scaling purposes, a new or replaced API service installation will only need to subscribe to the FastRouter to be included in the pool.

Playbook/Role impact

A new playbook will need to be created to install nginx and a uWSGI FastRouter. An existing nginx role from galaxy would be preferred, but one may need to be created or contributed to if there is not one that supports all of the same operating systems that OpenStack-Ansible does or that does not provide sufficient configuration options for our needs.

Each existing OpenStack role that deploys a web server will need to have those tasks removed and replaced with a task to configure and enable an nginx site, delegated to any hosts running nginx. The uWSGI configuration within those roles will also need to be updated with options to subscribe to an existing FastRouter.

The HAProxy server role may require more flexibility, such as allowing SSL passthrough and just-in-time configuration of services and backends.

Upgrade impact

Since load balancer backends will likely be changing, the load balancer will need to initially contain both the existing backends and the centralized nginx backends to limit downtime during upgrades.

Security impact

The proposed change should increase the security posture of OpenStack-Ansible since it will allow deployers to have more flexibility over where SSL is terminated, including preventing traffic from leaving a host once it is unencrypted.

Performance impact

nginx is a high performance web server and reverse proxy. There may be some host performance increase with less web servers running on a single controller host. Deployment times may also be improved by removing the duplicated tasks of installing and configuring a web server within multiple roles.

End user impact

N/A

Deployer impact

Deployers will need to be aware of the deployment architecture changes this change includes. Additional options will be available to deployers to configure a shared nginx instance, uWSGI FastRouter, and the distribution of SSL certificates. Deployers must also be made aware of any upgrade impact this change may have.

Developer impact

Future OpenStack roles will need to include the patterns established by this change. Tests will require additional Ansible roles, playbooks, and variables.

Dependencies

The HAProxy server role may require some changes and increased configuration flexibility.

The keystone role currently conditionally installs Apache when used as an identity provider or service provider for an external identity provider. This will need to be investigated further to ensure that nginx can provide the same level of support for those roles, including gated test scenarios.

The horizon role also currently installs Apache and uses it to serve static content. When running on a container, that static content may need to move to a mount to remain accessible to an external web server.

Implementation

Assignee(s)

Primary assignee:

jimmy-mccrory (jmccrory)

Other contributors:

<launchpad-id or None>

Work items

  • Evaluate existing nginx roles in galaxy

  • Develop new nginx role if necessary

  • Develop playbook for deploying nginx and uWSGI FastRouter

  • Adapt HAProxy role

  • Evaluate bind mounting of files statically served by web server

  • Update OpenStack roles to create nginx site configuration and subscribe to FastRouter for API services

Testing

Individual roles and the integrated repo will test the changes involved as they are implemented.

Documentation impact

The changes to the deployment architecture and any additional options for configuring nginx, uWSGI FastRouter, HAProxy, and SSL will need to be documented.

The impacts to upgrades and steps to minimize, and hopefully avoid, API downtime will also need to be documented.

References

https://uwsgi-docs.readthedocs.io/en/latest/Fastrouter.html