Scaling Ansible and AWX/Tower with Network and Cloud Inventories

This topic is covered more in-depth in my Red Hat Summit talk on Managing 15,000 Network Devices.

Quick primer: Ansible is a CLI orchestration application that is written in Python and that operates over SSH and HTTPS. AWX (downstream, unsupported) and Tower (upstream, supported) are the suite of UI/API, job scheduler, and security gateway functionalities around Ansible.

Ansible and AWX/Tower operate and function somewhat differently when configuring network, cloud, and generic platform endpoints, versus when performing traditional OS management or targeting APIs. The differentiator between Ansible’s connectivity is, quite frankly, OS and applications — things that can run Python — versus everything else that cannot run Python.

Ansible vs. Operating Systems

When Ansible runs against an OS like Linux and Windows, the remote hosts receive a tarball of python programs/plugins, Operating System, or API commands via SSH or HTTPS. The remote hosts unpack and runs these playbooks, while APIs receive a sequence of URLs. In either case, both types of OS and API configurations returns the results to Ansible/Tower. In the case of OS’ like Linux and Windows, these hosts process their own data and state changes, and then return the results to Ansible/Tower.

As an example with a Linux host, a standard playbook to enable and configure the host logging service would be initiated by Ansible/Tower, and would then run entirely on the remote host. Upon completion, only task results and state changes are sent back to Ansible. With OS automation, Tower orchestrates changes and processes data.

Ansible vs. Network and Cloud Devices

Network and cloud devices, on the other hand,  don’t perform their own data processing, and are often sending nonstop command output back to Ansible. In this case, all data processing is performed locally on Ansible or AWX/Tower nodes.

Rather than being able to rely on remote devices to do their own work, Ansible handles all data processing as it’s received from network cloud devices. This will have drastic, and potentially catastrophic, implications when running playbooks at scale against network/cloud inventories.

Ansible Networking at Scale — Things to Consider

In the pursuit of scaling Ansible and AWX/Tower to manage network and cloud devices, we must consider a number of factors that will directly impact playbook and job performance:

  • Frequency/extent of orchestrating/scheduling device changes
    With any large inventory, there comes a balancing act between scheduling frequent or large-scale configuration changes, while avoiding physical resource contention. At a high level, this can be as simple as benchmarking job run times with Tower resource loads, and setting job template forks accordingly. This will become critical in future development. More on that later.
  • Device configuration size
    Most network automation roles will be utilizing Ansible Facts derived from inventory vars and device configs. By looking at the raw device config sizes, such as the text output from show run all, we can establish a rough estimate of per-host memory usage during large jobs.
  • Inventory sizes and devices families, e.g. IOS, NXOS, XR
    Depending on overall inventory size, and the likelihood of significant inventory metadata, it’s critical to ensure that inventories are broken into multiple smaller groups — group sizes of 500 or less are preferable, while it’s highly recommended to limit max group sizes to 5,000 or less.

    It’s important to note that device types/families perform noticeably faster/slower than others. IOS, for instance, is often 3-4 faster than NXOS.
  • Making Use of Ansible Facts
    Ansible can collect device “facts” — useful variables about remote hosts — that can be used in playbooks. These facts can be cached in Tower, as well. The combination of using network facts and fact caching can allow you to poll existing data rather than parsing real-time commands.

    Effectively using facts, and the fact cache, will significantly increase Ansible/Tower job speed, while reducing overall processing loads.
  • Development methodology
    This one goes without saying, but bear with me.

    When creating new automation roles, it’s imperative that you establish solid development practices. Ideally, you want to outright avoid potentially significant processing and execution times that plague novice developers.

    Start with simple, stepping through your automation workflow task-by-task, and understand the logical progression of tasks/changes.. Ansible is a wonderfully simple tool, but it’s easy to overcomplicate code with faulty or overly-complex logic.

    And be careful with numerous role dependencies, and dependency recursion using layer-upon-layer of ’include’ and ’import’. If you’re traversing more than 3-4 levels per role, then it’s time to break out that automation/logic into smaller chunks. Otherwise, a large role running against a large inventory can run OOM simply from attempting to load the same million dependencies per host.

Leave a Reply

Your email address will not be published. Required fields are marked *