Audit-Mode for Chef Client

Audit mode is an new phase in Chef which allows you to evaluate custom rules, defined in your recipes, on every node during each chef-client run. Use audits to ensure nodes fall into existing "known states" categories even before Chef converges, and to validate your infrastructure after Chef converges.


As an inheritor of a non-Chef-managed infrastructure
I want to run chef-client and collect data on each node without converging
so that I can determine the existing state of the inherited infrastructure.

As a maintainer of a Chef-managed infrastructure
I want to write custom rules defining expected state
so that I can validate my infrastructure.


Audit mode phase

Audits are evaluated in their own phase. During a full chef-client run, auditing occurs independently after client converges.


By default, client will converge the node and executed audits. Chef can be configured to skip audit mode via the command line flag --no-audit-mode or the configuration file option audit_mode :disabled.

Alternatively, converge can be skipped via the command line flag --audit-mode or the configuration file option audit_mode :audit_only.


As controls are evaluated during the audit phase, results will be streamed to Chef::Config[:log_location]` in an easy to read format using RSpec's documentation formatter.

Event dispatch

The Chef::EventDispatch::Base will be updated to support the following events

Event Name | Context -----------|--------- converge_failed(error) | client did not converge successfully with error audit_phase_start(run_status) | audit phase started audit_phase_complete | audit phase finished audit_phase_failed(exception) | an uncaught exception occurred during the audit phase control_group_started(name) | signifies the start of a controls group with a defined name control_example_success(control_group_name, example_data) | an example in a control_group_name group completed successfully control_example_failure(control_group_name, example_data, error) | an example in a control_group_name group failed with error

The example_data hash contains the informational fields * the name of the evaluated audit rule * the full desc of the evaluated audit rule (includes name) * the resource_type evaluated, if any * the name of the evaluated resource, resource_name * any containing scope is saved in context * the line_number of the failed audit


Audits are written inside recipe files. Audits can be written in a separate recipe or can be added into recipes defining resources. Audits are collected within a named controls block, which does not get evaluated until the audit phase begins.

Audit rules are defined within a controls group using RSpec's it syntax. Rules can be grouped together using the control method, or any other RSpec example group method (e.g., describe or context). RSpec's built-in matchers are available, as well as Serverspec types and matchers. The use of :should is explicitly disabled, as this is deprecated in RSpec 3.

Example: Nobody is listening

Audits can be written to help ensure compliance requirements, such as asserting nothing is listening on port 111. Depending on your distribution and its version, your portmap service may be named "portmap" or "rpcbind", and could be renamed after a version bump. Your recipe may use the correct service provider but the init script may have been removed, preventing any service resource :stop action from completing successfully.

The ports::audit recipe ensures nothing is listening on port 111: ```ruby

cookbook: ports

recipe: audit

controls "port compliance" do control port(111) do it "has nothing listening" expect(port(111)).tonot belistening end end ```

When ports::audit is added to the run-list and chef-client is run with audit mode enabled, you would expect the log output to contain ```sh port compliance Port "111" has nothing listening

Finished in 0.08615 seconds (files took 0.67889 seconds to load) 1 example, 0 failures ```


When an audit fails, the failed example is marked in the log output for debugging. At the end of the client run, Chef will exit with exit status 1.

Example: Somebody is listening

Suppose port 111 was not shut down correctly and someone is listening on it. When the ports::audit recipe is run, the log output would contain something similar to ```sh port compliance Port "111" has nothing listening (FAILED - 1)


1) port compliance Port "111" has nothing listening Failure/Error: expect(port(111)).tonot belistening expected Port "111" not to be listening # cookbooks/ports/recipes/audit.rb:7:in `block (3 levels) in from_file'

Finished in 0.12515 seconds (files took 0.70174 seconds to load) 1 example, 1 failure

Failed examples:

rspec cookbooks/ports/recipes/audit.rb:6 # port compliance Port "111" has nothing listening ```


These exceptions can be raised during a client run due to errors in the audits included in recipes in the run list:

Error handling

Errors occurring in the converge phase do not affect the execution of the audit phase. Similarly, errors occurring in the audit phase do not affect later phases. Errors are collected to be provided to the appropriate error handlers once each phase completes.


Why distribute audits in recipes?

Cookbooks support versioning and are an effective medium for distributing code. Including audits in recipes help to maintain a flat directory structure, and don't require the addition of a new server segment.

Why implement this in core chef?

Even though it's possible to build this logic as external libraries (see minitest-chef-handler) building it as a first class citizen with config options, CLI options and hooks for event handlers and maintaining it overtime will be a challenge.

Also to achieve usability, any TDI (test driven infrastructure) related logic should be available out of the box inside the Client omnibus packages. As long as functionality is available out of the box, building it into core as an alpha feature vs implementing it as an external gem is only an implementation choice. This doesn't change any compatibility commitments.

Why use RSpec and Serverspec in audits?

In the future we can definitely come up with a better DSL than RSpec. But we would like to reuse the awesome tool Serverspec and its practices as well as we would like to provide a generic interface for the power users.

Why support an audit-only phase?

The use case for auditing without converging is to support an existing Chef customer absorbing a non-Chef managed infrastructure. In this instance, they can only run audits until cookbooks have been prepared for the new infrastructure. Similarly, an audit-only phase can help new users convert their unmanaged infrastructure to a Chef-managed infrastructure.

Why run audits after a failed converge?

So that you can validate your infrastructure is still in a state consistent with your expectations. Ideally, when converge fails your audits should still pass. It's reassuring to have that sanity check.


This work is in the public domain. In jurisdictions that do not allow for this, this work is available under CC0. To the extent possible under law, the person who associated CC0 with this work has waived all copyright and related or neighboring rights to this work.