ansible/docs/docsite/rst/dev_guide/platforms/aws_guidelines.rst

.. _AWS_module_development:

****************************************************
Guidelines for Ansible Amazon AWS module development
****************************************************

The Ansible AWS modules and these guidelines are maintained by the Ansible AWS Working Group.  For
further information see
the `AWS working group community page <https://github.com/ansible/community/wiki/aws>`_.
If you are planning to contribute AWS modules to Ansible then getting in touch with the working
group will be a good way to start, especially because a similar module may already be under
development.

.. contents::
   :local:

Maintaining existing modules
============================

Fixing bugs
-----------

Bug fixes to code that relies on boto will still be accepted. When possible,
the code should be ported to use boto3.

Adding new features
-------------------

Try to keep backward compatibility with relatively recent versions of boto3. That means that if you
want to implement some functionality that uses a new feature of boto3, it should only fail if that
feature actually needs to be run, with a message stating the missing feature and minimum required
version of boto3.

Use feature testing (e.g. ``hasattr('boto3.module', 'shiny_new_method')``) to check whether boto3
supports a feature rather than version checking. For example, from the ``ec2`` module:

.. code-block:: python

   if boto_supports_profile_name_arg(ec2):
       params['instance_profile_name'] = instance_profile_name
   else:
       if instance_profile_name is not None:
           module.fail_json(msg="instance_profile_name parameter requires boto version 2.5.0 or higher")

Migrating to boto3
------------------

Prior to Ansible 2.0, modules were written in either boto3 or boto. We are
still porting some modules to boto3. Modules that still require boto should be ported to use boto3 rather than using both libraries (boto and boto3). We would like to remove the boto dependency from all modules.

Porting code to AnsibleAWSModule
---------------------------------

Some old AWS modules use the generic ``AnsibleModule`` as a base rather than the more efficient ``AnsibleAWSModule``. To port an old module to ``AnsibleAWSModule``, change:

.. code-block:: python

   from ansible.module_utils.basic import AnsibleModule
   ...
   module = AnsibleModule(...)

to:

.. code-block:: python

   from ansible.module_utils.aws.core import AnsibleAWSModule
   ...
   module = AnsibleAWSModule(...)

Few other changes are required. AnsibleAWSModule
does not inherit methods from AnsibleModule by default, but most useful methods
are included. If you do find an issue, please raise a bug report.

When porting, keep in mind that AnsibleAWSModule also will add the default ec2
argument spec by default. In pre-port modules, you should see common arguments
specified with:

.. code-block:: python

   def main():
       argument_spec = ec2_argument_spec()
       argument_spec.update(dict(
           state=dict(default='present', choices=['present', 'absent', 'enabled', 'disabled']),
           name=dict(default='default'),
           # ... and so on ...
       ))
       module = AnsibleModule(argument_spec=argument_spec, supports_check_mode=True,)

These can be replaced with:

.. code-block:: python

   def main():
       argument_spec = dict(
           state=dict(default='present', choices=['present', 'absent', 'enabled', 'disabled']),
           name=dict(default='default'),
           # ... and so on ...
       )
       module = AnsibleAWSModule(argument_spec=argument_spec, supports_check_mode=True,)

Creating new AWS modules
========================

Use boto3 and AnsibleAWSModule
-------------------------------

All new AWS modules must use boto3 and ``AnsibleAWSModule``.

``AnsibleAWSModule`` greatly simplifies exception handling and library
management, reducing the amount of boilerplate code. If you cannot
use ``AnsibleAWSModule`` as a base, you must document the reason and request an exception to this rule.

Naming your module
------------------

Base the name of the module on the part of AWS that you actually use. (A good rule of thumb is to
take whatever module you use with boto as a starting point).  Don't further abbreviate names - if
something is a well known abbreviation of a major component of AWS (for example, VPC or ELB), that's fine, but
don't create new ones independently.

Unless the name of your service is quite unique, please consider using ``aws_`` as a prefix. For example ``aws_lambda``.

Importing botocore and boto3
----------------------------

The ``ansible.module_utils.ec2`` module and ``ansible.module_utils.core.aws`` modules both
automatically import boto3 and botocore.  If boto3 is missing from the system then the variable
``HAS_BOTO3`` will be set to false.  Normally, this means that modules don't need to import
boto3 directly. There is no need to check ``HAS_BOTO3`` when using AnsibleAWSModule
as the module does that check:

.. code-block:: python

   from ansible.module_utils.aws.core import AnsibleAWSModule
   try:
       import botocore
   except ImportError:
       pass  # handled by AnsibleAWSModule

or:

.. code-block:: python

   from ansible.module_utils.basic import AnsibleModule
   from ansible.module_utils.ec2 import HAS_BOTO3
   try:
       import botocore
   except ImportError:
       pass  # handled by imported HAS_BOTO3

   def main():

       if not HAS_BOTO3:
           module.fail_json(msg='boto3 and botocore are required for this module')

Supporting Module Defaults
--------------------------

The existing AWS modules support using :ref:`module_defaults <module_defaults>` for common 
authentication parameters.  To do the same for your new module, add an entry for it in
``lib/ansible/config/module_defaults.yml``.  These entries take the form of:

.. code-block:: yaml

  aws_module_name:
  - aws

Connecting to AWS
=================

AnsibleAWSModule provides the ``resource`` and ``client`` helper methods for obtaining boto3 connections.
These handle some of the more esoteric connection options, such as security tokens and boto profiles.

If using the basic AnsibleModule then you should use ``get_aws_connection_info`` and then ``boto3_conn``
to connect to AWS as these handle the same range of connection options.

These helpers also for missing profiles or a region not set when it needs to be, so you don't have to.

An example of connecting to ec2 is shown below. Note that unlike boto there is no ``NoAuthHandlerFound``
exception handling like in boto. Instead, an ``AuthFailure`` exception will be thrown when you use the
connection. To ensure that authorization, parameter validation and permissions errors are all caught,
you should catch ``ClientError`` and ``BotoCoreError`` exceptions with every boto3 connection call.
See exception handling:

.. code-block:: python

   module.client('ec2')

or for the higher level ec2 resource:

.. code-block:: python

   module.resource('ec2')


An example of the older style connection used for modules based on AnsibleModule rather than AnsibleAWSModule:

.. code-block:: python

   region, ec2_url, aws_connect_params = get_aws_connection_info(module, boto3=True)
   connection = boto3_conn(module, conn_type='client', resource='ec2', region=region, endpoint=ec2_url, **aws_connect_params)

.. code-block:: python

   region, ec2_url, aws_connect_params = get_aws_connection_info(module, boto3=True)
   connection = boto3_conn(module, conn_type='client', resource='ec2', region=region, endpoint=ec2_url, **aws_connect_params)


Common Documentation Fragments for Connection Parameters
--------------------------------------------------------

There are two :ref:`common documentation fragments <module_docs_fragments>`
that should be included into almost all AWS modules:

* ``aws`` - contains the common boto connection parameters
* ``ec2`` - contains the common region parameter required for many AWS modules

These fragments should be used rather than re-documenting these properties to ensure consistency
and that the more esoteric connection options are documented. For example:

.. code-block:: python

   DOCUMENTATION = '''
   module: my_module
   # some lines omitted here
   requirements: [ 'botocore', 'boto3' ]
   extends_documentation_fragment:
       - aws
       - ec2
   '''

Handling exceptions
===================

You should wrap any boto3 or botocore call in a try block. If an exception is thrown, then there
are a number of possibilities for handling it.

* Catch the general ``ClientError`` or look for a specific error code with
    ``is_boto3_error_code``.
* Use ``aws_module.fail_json_aws()`` to report the module failure in a standard way
* Retry using AWSRetry
* Use ``fail_json()`` to report the failure without using ``ansible.module_utils.aws.core``
* Do something custom in the case where you know how to handle the exception

For more information on botocore exception handling see the `botocore error documentation <https://botocore.readthedocs.io/en/latest/client_upgrades.html#error-handling>`_.

Using is_boto3_error_code
-------------------------

To use ``ansible.module_utils.aws.core.is_boto3_error_code`` to catch a single
AWS error code, call it in place of ``ClientError`` in your except clauses. In
this case, *only* the ``InvalidGroup.NotFound`` error code will be caught here,
and any other error will be raised for handling elsewhere in the program.

.. code-block:: python

   try:
       info = connection.describe_security_groups(**kwargs)
   except is_boto3_error_code('InvalidGroup.NotFound'):
       pass
   do_something(info)  # do something with the info that was successfully returned

Using fail_json_aws()
---------------------

In the AnsibleAWSModule there is a special method, ``module.fail_json_aws()`` for nice reporting of
exceptions.  Call this on your exception and it will report the error together with a traceback for
use in Ansible verbose mode.

You should use the AnsibleAWSModule for all new modules, unless not possible. If adding significant
amounts of exception handling to existing modules, we recommend migrating the module to use AnsibleAWSModule
(there are very few changes required to do this)

.. code-block:: python

   from ansible.module_utils.aws.core import AnsibleAWSModule

   # Set up module parameters
   # module params code here

   # Connect to AWS
   # connection code here

   # Make a call to AWS
   name = module.params.get['name']
   try:
       result = connection.describe_frooble(FroobleName=name)
   except (botocore.exceptions.BotoCoreError, botocore.exceptions.ClientError) as e:
       module.fail_json_aws(e, msg="Couldn't obtain frooble %s" % name)

Note that it should normally be acceptable to catch all normal exceptions here, however if you
expect anything other than botocore exceptions you should test everything works as expected.

If you need to perform an action based on the error boto3 returned, use the error code.

.. code-block:: python

   # Make a call to AWS
   name = module.params.get['name']
   try:
       result = connection.describe_frooble(FroobleName=name)
   except botocore.exceptions.ClientError as e:
       if e.response['Error']['Code'] == 'FroobleNotFound':
           workaround_failure()  # This is an error that we can work around
       else:
           module.fail_json_aws(e, msg="Couldn't obtain frooble %s" % name)
   except botocore.exceptions.BotoCoreError as e:
       module.fail_json_aws(e, msg="Couldn't obtain frooble %s" % name)

using fail_json() and avoiding ansible.module_utils.aws.core
------------------------------------------------------------

Boto3 provides lots of useful information when an exception is thrown so pass this to the user
along with the message.

.. code-block:: python

   from ansible.module_utils.ec2 import HAS_BOTO3
   try:
       import botocore
   except ImportError:
       pass  # caught by imported HAS_BOTO3

   # Connect to AWS
   # connection code here

   # Make a call to AWS
   name = module.params.get['name']
   try:
       result = connection.describe_frooble(FroobleName=name)
   except botocore.exceptions.ClientError as e:
       module.fail_json(msg="Couldn't obtain frooble %s: %s" % (name, str(e)),
                        exception=traceback.format_exc(),
                        **camel_dict_to_snake_dict(e.response))

Note: we use `str(e)` rather than `e.message` as the latter doesn't
work with python3

If you need to perform an action based on the error boto3 returned, use the error code.

.. code-block:: python

   # Make a call to AWS
   name = module.params.get['name']
   try:
       result = connection.describe_frooble(FroobleName=name)
   except botocore.exceptions.ClientError as e:
       if e.response['Error']['Code'] == 'FroobleNotFound':
           workaround_failure()  # This is an error that we can work around
       else:
           module.fail_json(msg="Couldn't obtain frooble %s: %s" % (name, str(e)),
                            exception=traceback.format_exc(),
                            **camel_dict_to_snake_dict(e.response))
   except botocore.exceptions.BotoCoreError as e:
       module.fail_json_aws(e, msg="Couldn't obtain frooble %s" % name)


API throttling (rate limiting) and pagination
=============================================

For methods that return a lot of results, boto3 often provides
`paginators <https://boto3.readthedocs.io/en/latest/guide/paginators.html>`_. If the method
you're calling has ``NextToken`` or ``Marker`` parameters, you should probably
check whether a paginator exists (the top of each boto3 service reference page has a link
to Paginators, if the service has any). To use paginators, obtain a paginator object,
call ``paginator.paginate`` with the appropriate arguments and then call ``build_full_result``.

Any time that you are calling the AWS API a lot, you may experience API throttling,
and there is an ``AWSRetry`` decorator that can be used to ensure backoff. Because
exception handling could interfere with the retry working properly (as AWSRetry needs to
catch throttling exceptions to work correctly), you'd need to provide a backoff function
and then put exception handling around the backoff function.

You can use ``exponential_backoff`` or ``jittered_backoff`` strategies - see
the cloud ``module_utils`` ()/lib/ansible/module_utils/cloud.py)
and `AWS Architecture blog <https://www.awsarchitectureblog.com/2015/03/backoff.html>`_ for more details.

The combination of these two approaches is then:

.. code-block:: python

   @AWSRetry.exponential_backoff(retries=5, delay=5)
   def describe_some_resource_with_backoff(client, **kwargs):
        paginator = client.get_paginator('describe_some_resource')
        return paginator.paginate(**kwargs).build_full_result()['SomeResource']

   def describe_some_resource(client, module):
       filters = ansible_dict_to_boto3_filter_list(module.params['filters'])
       try:
           return describe_some_resource_with_backoff(client, Filters=filters)
       except botocore.exceptions.ClientError as e:
           module.fail_json_aws(e, msg="Could not describe some resource")


If the underlying ``describe_some_resources`` API call throws a ``ResourceNotFound``
exception, ``AWSRetry`` takes this as a cue to retry until it's not thrown (this
is so that when creating a resource, we can just retry until it exists).

To handle authorization failures or parameter validation errors in
``describe_some_resource_with_backoff``, where we just want to return ``None`` if
the resource doesn't exist and not retry, we need:

.. code-block:: python

   @AWSRetry.exponential_backoff(retries=5, delay=5)
   def describe_some_resource_with_backoff(client, **kwargs):
        try:
            return client.describe_some_resource(ResourceName=kwargs['name'])['Resources']
        except botocore.exceptions.ClientError as e:
            if e.response['Error']['Code'] == 'ResourceNotFound':
                return None
            else:
                raise
        except BotoCoreError as e:
            raise

   def describe_some_resource(client, module):
       name = module.params.get['name']
       try:
           return describe_some_resource_with_backoff(client, name=name)
       except (botocore.exceptions.BotoCoreError, botocore.exceptions.ClientError) as e:
           module.fail_json_aws(e, msg="Could not describe resource %s" % name)


To make use of AWSRetry easier, it can now be wrapped around a client returned
by ``AnsibleAWSModule``. any call from a client. To add retries to a client,
create a client:

.. code-block:: python

   module.client('ec2', retry_decorator=AWSRetry.jittered_backoff(retries=10))

Any calls from that client can be made to use the decorator passed at call-time
using the `aws_retry` argument. By default, no retries are used.

.. code-block:: python

   ec2 = module.client('ec2', retry_decorator=AWSRetry.jittered_backoff(retries=10))
   ec2.describe_instances(InstanceIds=['i-123456789'], aws_retry=True)

   # equivalent with normal AWSRetry
   @AWSRetry.jittered_backoff(retries=10)
   def describe_instances(client, **kwargs):
       return ec2.describe_instances(**kwargs)

   describe_instances(module.client('ec2'), InstanceIds=['i-123456789'])

The call will be retried the specified number of times, so the calling functions
don't need to be wrapped in the backoff decorator.

You can also use customization for ``retries``, ``delay`` and ``max_delay`` parameters used by
``AWSRetry.jittered_backoff`` API using module params. You can take a look at
the `cloudformation <cloudformation_module>` module for example.

To make all Amazon modules uniform, prefix the module param with ``backoff_``, so ``retries`` becomes ``backoff_retries``
 and likewise with ``backoff_delay`` and ``backoff_max_delay``.

Returning Values
================

When you make a call using boto3, you will probably get back some useful information that you
should return in the module.  As well as information related to the call itself, you will also have
some response metadata.  It is OK to return this to the user as well as they may find it useful.

Boto3 returns all values CamelCased.  Ansible follows Python standards for variable names and uses
snake_case. There is a helper function in module_utils/ec2.py called `camel_dict_to_snake_dict`
that allows you to easily convert the boto3 response to snake_case.

You should use this helper function and avoid changing the names of values returned by Boto3.
E.g. if boto3 returns a value called 'SecretAccessKey' do not change it to 'AccessKey'.

.. code-block:: python

   # Make a call to AWS
   result = connection.aws_call()

   # Return the result to the user
   module.exit_json(changed=True, **camel_dict_to_snake_dict(result))

Dealing with IAM JSON policy
============================

If your module accepts IAM JSON policies then set the type to 'json' in the module spec. For
example:

.. code-block:: python

   argument_spec.update(
       dict(
           policy=dict(required=False, default=None, type='json'),
       )
   )

Note that AWS is unlikely to return the policy in the same order that is was submitted. Therefore,
use the `compare_policies` helper function which handles this variance.

`compare_policies` takes two dictionaries, recursively sorts and makes them hashable for comparison
and returns True if they are different.

.. code-block:: python

   from ansible.module_utils.ec2 import compare_policies

   import json

   # some lines skipped here

   # Get the policy from AWS
   current_policy = json.loads(aws_object.get_policy())
   user_policy = json.loads(module.params.get('policy'))

   # Compare the user submitted policy to the current policy ignoring order
   if compare_policies(user_policy, current_policy):
       # Update the policy
       aws_object.set_policy(user_policy)
   else:
       # Nothing to do
       pass

Dealing with tags
=================

AWS has a concept of resource tags. Usually the boto3 API has separate calls for tagging and
untagging a resource.  For example, the ec2 API has a create_tags and delete_tags call.

It is common practice in Ansible AWS modules to have a `purge_tags` parameter that defaults to
true.

The `purge_tags` parameter means that existing tags will be deleted if they are not specified by
the Ansible task.

There is a helper function `compare_aws_tags` to ease dealing with tags. It can compare two dicts
and return the tags to set and the tags to delete.  See the Helper function section below for more
detail.

Helper functions
================

Along with the connection functions in Ansible ec2.py module_utils, there are some other useful
functions detailed below.

camel_dict_to_snake_dict
------------------------

boto3 returns results in a dict.  The keys of the dict are in CamelCase format. In keeping with
Ansible format, this function will convert the keys to snake_case.

``camel_dict_to_snake_dict`` takes an optional parameter called ``ignore_list`` which is a list of
keys not to convert (this is usually useful for the ``tags`` dict, whose child keys should remain with
case preserved)

Another optional parameter is ``reversible``. By default, ``HTTPEndpoint`` is converted to ``http_endpoint``,
which would then be converted by ``snake_dict_to_camel_dict`` to ``HttpEndpoint``.
Passing ``reversible=True`` converts HTTPEndpoint to ``h_t_t_p_endpoint`` which converts back to ``HTTPEndpoint``.

snake_dict_to_camel_dict
------------------------

`snake_dict_to_camel_dict` converts snake cased keys to camel case. By default, because it was
first introduced for ECS purposes, this converts to dromedaryCase. An optional
parameter called `capitalize_first`, which defaults to `False`, can be used to convert to CamelCase.

ansible_dict_to_boto3_filter_list
---------------------------------

Converts a an Ansible list of filters to a boto3 friendly list of dicts.  This is useful for any
boto3 `_facts` modules.

boto_exception
--------------

Pass an exception returned from boto or boto3, and this function will consistently get the message from the exception.

Deprecated: use `AnsibleAWSModule`'s `fail_json_aws` instead.


boto3_tag_list_to_ansible_dict
------------------------------

Converts a boto3 tag list to an Ansible dict. Boto3 returns tags as a list of dicts containing keys
called 'Key' and 'Value' by default.  This key names can be overridden when calling the function.
For example, if you have already camel_cased your list of tags you may want to pass lowercase key
names instead i.e. 'key' and 'value'.

This function converts the list in to a single dict where the dict key is the tag key and the dict
value is the tag value.

ansible_dict_to_boto3_tag_list
------------------------------

Opposite of above. Converts an Ansible dict to a boto3 tag list of dicts. You can again override
the key names used if 'Key' and 'Value' is not suitable.

get_ec2_security_group_ids_from_names
-------------------------------------

Pass this function a list of security group names or combination of security group names and IDs
and this function will return a list of IDs.  You should also pass the VPC ID if known because
security group names are not necessarily unique across VPCs.

compare_policies
----------------

Pass two dicts of policies to check if there are any meaningful differences and returns true
if there are. This recursively sorts the dicts and makes them hashable before comparison.

This method should be used any time policies are being compared so that a change in order
doesn't result in unnecessary changes.

compare_aws_tags
----------------

Pass two dicts of tags and an optional purge parameter and this function will return a dict
containing key pairs you need to modify and a list of tag key names that you need to remove.  Purge
is True by default.  If purge is False then any existing tags will not be modified.

This function is useful when using boto3 'add_tags' and 'remove_tags' functions. Be sure to use the
other helper function `boto3_tag_list_to_ansible_dict` to get an appropriate tag dict before
calling this function. Since the AWS APIs are not uniform (e.g. EC2 versus Lambda) this will work
without modification for some (Lambda) and others may need modification before using these values
(such as EC2, with requires the tags to unset to be in the form `[{'Key': key1}, {'Key': key2}]`).

Integration Tests for AWS Modules
=================================

All new AWS modules should include integration tests to ensure that any changes in AWS APIs that
affect the module are detected. At a minimum this should cover the key API calls and check the
documented return values are present in the module result.

For general information on running the integration tests see the :ref:`Integration Tests page of the
Module Development Guide <testing_integration>`, especially the section on configuration for cloud tests.

The integration tests for your module should be added in `test/integration/targets/MODULE_NAME`.

You must also have a aliases file in `test/integration/targets/MODULE_NAME/aliases`. This file serves
two purposes. First indicates it's in an AWS test causing the test framework to make AWS credentials
available during the test run. Second putting the test in a test group causing it to be run in the
continuous integration build.

Tests for new modules should be added to the same group as existing AWS tests. In general just copy
an existing aliases file such as the `aws_s3 tests aliases file <https://github.com/ansible/ansible/blob/devel/test/integration/targets/aws_s3/aliases>`_.

AWS Credentials for Integration Tests
-------------------------------------

The testing framework handles running the test with appropriate AWS credentials, these are made available
to your test in the following variables:

* `aws_region`
* `aws_access_key`
* `aws_secret_key`
* `security_token`

So all invocations of AWS modules in the test should set these parameters. To avoid duplicating these
for every call, it's preferable to use :ref:`module_defaults <module_defaults>`. For example:

.. code-block:: yaml

   - name: set connection information for aws modules and run tasks
     module_defaults:
       group/aws:
         aws_access_key: "{{ aws_access_key }}"
         aws_secret_key: "{{ aws_secret_key }}"
         security_token: "{{ security_token | default(omit) }}"
         region: "{{ aws_region }}"

     block:

     - name: Do Something
       ec2_instance:
         ... params ...

     - name: Do Something Else
       ec2_instance:
         ... params ...

AWS Permissions for Integration Tests
-------------------------------------

As explained in the :ref:`Integration Test guide <testing_integration>`
there are defined IAM policies in ``hacking/aws_config/testing_policies/`` that contain the necessary permissions
to run the AWS integration test. The permissions used by CI are more restrictive than those in ``hacking/aws_config/testing_policies``; for CI we want
the most restrictive policy possible that still allows the given tests to pass.

If your module interacts with a new service or otherwise requires new permissions, tests will fail when you submit a pull request and the
`Ansibullbot <https://github.com/ansible/ansibullbot/blob/master/ISSUE_HELP.md>`_ will tag your PR as needing revision.
We do not automatically grant additional permissions to the roles used by the continuous integration builds. You must provide the minimum IAM permissions required to run your integration test.

If your PR has test failures, check carefully to be certain the failure is only due to the missing permissions. If you've ruled out other sources of failure, add a comment with the `ready_for_review`
tag and explain that it's due to missing permissions.

Your pull request cannot be merged until the tests are passing. If your pull request is failing due to missing permissions,
you must collect the minimum IAM permissions required to
run the tests.

There are two ways to figure out which IAM permissions you need for your PR to pass:

* Start with the most permissive IAM policy, run the tests to collect information about which resources your tests actually use, then construct a policy based on that output. This approach only works on modules that use `AnsibleAWSModule`.
* Start with the least permissive IAM policy, run the tests to discover a failure, add permissions for the resource that addresses that failure, then repeat. If your module uses `AnsibleModule` instead of `AnsibleAWSModule`, you must use this approach.

To start with the most permissive IAM policy:

1) `Create an IAM policy <https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_create.html#access_policies_create-start>`_ that allows all actions (set ``Action`` and ``Resource`` to ``*```).
2) Run your tests locally with this policy. On AnsibleAWSModule-based modules, the ``debug_botocore_endpoint_logs`` option is automatically set to ``yes``, so you should see a list of AWS ACTIONS after the PLAY RECAP showing all the permissions used. If your tests use a boto/AnsibleModule module, you must start with the least permissive policy (see below).
3) Modify your policy to allow only the actions your tests use. Restrict account, region, and prefix where possible. Wait a few minutes for your policy to update.
4) Run the tests again with a user or role that allows only the new policy.
5) If the tests fail, troubleshoot (see tips below), modify the policy, run the tests again, and repeat the process until the tests pass with a restrictive policy.
6) Open a pull request proposing the minimum required policy to the `testing policies <https://github.com/mattclay/aws-terminator/tree/master/aws/policy>`_.

To start from the least permissive IAM policy:

1) Run the integration tests locally with no IAM permissions.
2) Examine the error when the tests reach a failure.
    a) If the error message indicates the action used in the request, add the action to your policy.
    b) If the error message does not indicate the action used in the request:
        - Usually the action is a CamelCase version of the method name - for example, for an ec2 client the method `describe_security_groups` correlates to the action `ec2:DescribeSecurityGroups`.
        - Refer to the documentation to identify the action.
    c) If the error message indicates the resource ARN used in the request, limit the action to that resource.
    d) If the error message does not indicate the resource ARN used:
        - Determine if the action can be restricted to a resource by examining the documentation.
        - If the action can be restricted, use the documentation to construct the ARN and add it to the policy.
3) Add the action or resource that caused the failure to `an IAM policy <https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies_create.html#access_policies_create-start>`_. Wait a few minutes for your policy to update.
4) Run the tests again with this policy attached to your user or role.
5) If the tests still fail at the same place with the same error you will need to troubleshoot (see tips below). If the first test passes, repeat steps 2 and 3 for the next error. Repeat the process until the tests pass with a restrictive policy.
6) Open a pull request proposing the minimum required policy to the `testing policies <https://github.com/mattclay/aws-terminator/tree/master/aws/policy>`_.

Troubleshooting IAM policies
^^^^^^^^^^^^^^^^^^^^^^^^^^^^

- When you make changes to a policy, wait a few minutes for the policy to update before re-running the tests.
- Use the `policy simulator <https://policysim.aws.amazon.com/>`_ to verify that each action (limited by resource when applicable) in your policy is allowed.
- If you're restricting actions to certain resources, replace resources temporarily with `*`. If the tests pass with wildcard resources, there is a problem with the resource definition in your policy.
- If the initial troubleshooting above doesn't provide any more insight, AWS may be using additional undisclosed resources and actions.
- Examine the AWS FullAccess policy for the service for clues.
- Re-read the AWS documentation, especially the list of `Actions, Resources and Condition Keys <https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_policies_actions-resources-contextkeys.html>`_ for the various AWS services.
- Look at the `cloudonaut <https://iam.cloudonaut.io>`_ documentation as a troubleshooting cross-reference.
- Use a search engine.
- Ask in the Ansible IRC channel #ansible-aws (on freenode IRC).

Some cases where tests should be marked as unsupported:
1) The tests take longer than 10 or 15 minutes to complete
2) The tests create expensive resources
3) The tests create inline policies