Writing plans in Puppet Language

Plans allow you to run more than one task with a single command, compute values for the input to a task, process the results of tasks, or make decisions based on the result of running a task.

Write plans in the Puppet language, giving them a .pp extension, and place them in the module's /plans directory.

Plans can use any combination of Bolt functions or built in Puppetfunctions.

Naming plans

It is important to name plans correctly according to the module name, file name, and path to ensure easy code readability.

Place plan files in your module's ./plans directory, using these file extensions:
  • Puppet plans — .pp

  • YAML plans — .yaml, not .yml

Plan names are composed of two or more name segments, indicating:

  • The name of the module the plan is located in.

  • The name of the plan file, without the extension.

  • The path within the module, if the plan is in a subdirectory of ./plans.

For example, given a module called mymodule with a plan defined in ./mymodule/plans/myplan.pp, the plan name is mymodule::myplan.

A plan defined in ./mymodule/plans/service/myplan.pp would be mymodule::service::myplan. This name is how you refer to the plan when you run commands.

The plan filename init is special: the plan it defines is referenced using the module name only. For example, in a module called mymodule, the plan defined in init.pp is the mymodule plan.

Avoid giving plans the same names as constructs in the Puppet language. Although plans do not share their namespace with other language constructs, giving plans these names makes your code difficult to read.

Each plan name segment must begin with a lowercase letter and:

  • May include lowercase letters.

  • May include digits.

  • May include underscores.

  • Must not be a reserved word.

  • Must not have the same name as any Puppet data types.

  • Namespace segments must match the following regular expression \A[a-z][a-z0-9_]*\Z

RBAC considerations for writing plans

Take user permissions into consideration when writing plans by understanding how RBAC for plans works.

RBAC for plans is separate from RBAC for individual tasks. This means that a user can be excluded from running a certain task, but still have permission to run a plan that contains that task. Setting one permission does not affect the other.

This structure allows you to write plans with more robust, custom control over task permissions. Instead of allowing a user free rein to run a task that can potentially damage your infrastructure under the right conditions, you can wrap a task in a plan and only allow them to run it under circumstances you can control.

For example, say you are configuring permissions for a new user and allow them to run the plan plan infra::upgrade_git:

plan infra::upgrade_git (
  TargetSpec $nodes,
  Integer $version,
) {
  run_task(‘package’, $nodes, name => ’git’, action => ‘upgrade’, version => $version)
}

Within this plan, they can run the package task, but can only interact with the git package. The plan does not allow them to use any other parameters for the package task.

Even though they can run this plan, they do not have access to individually run the package task outside of this plan unless you grant them permission to do so. In that case, they would have the option to add any parameters they want to the task.

Use parameter types to fine-tune access

Writing parameter types into plan code provides even more control. In the upgrade_git example above, the plan only provides access to the git package, but the user can choose whatever version of git they want.

Let's say there are known vulnerabilities in some versions of the git package and you are concerned with your new user having the ability to use the versions you deem unsafe. You can use parameter types like Enum to restrict the version parameter to versions that are safe enough for deployment.

In this example, the Enum restricts the $version parameter to versions 1:2.17.0-1ubuntu1 and 1:2.17.1-1ubuntu0.4 only:
plan infra::upgrade_git (
  TargetSpec $nodes,
  Enum['1:2.17.0-1ubuntu1', '1:2.17.1-1ubuntu0.4'] $version,
) {
  run_task(‘package’, $nodes, name => ‘git’, action => ‘upgrade’, version => $version)
}

Any user attempting to run this plan must choose one of these versions for the plan to run.

You can also use PuppetDB queries to select parameter types. Using the same example, let's say you need to restrict the nodes that infra::upgrade_git can run on. Use a PuppetDB query to identify which nodes get selected for the git upgrade. It should look something like this:
plan infra::upgrade_git (
   Enum['1:2.17.0-1ubuntu1', '1:2.17.1-1ubuntu0.4'] $version,
) {
  # Use puppetdb to find the nodes from the “other” team's web cluster
  $query = [from, nodes, ['=', [fact, cluster], "other_team"]]
  $selected_nodes = puppetdb_query($query).map() |$target| {
    $target[certname]
  }
  run_task(‘package’, $selected_nodes, name => ‘git’, action => ‘upgrade’, version => $version)
}

Using these ideas, you can write powerful plans that give users exactly what they need without giving them the keys to the kingdom.

Defining plan parameters

Specify parameters within your plan.

Specify each parameter in your plan with its data type. For example, you might want parameters to specify which nodes to run different parts of your plan on.

The following example shows node parameters specified as data type TargetSpec. This allows this parameter to be passed as a single URL, comma-separated URL list, Target data type, or Array of either. For more information about these data types, see the common data types table in the related metadata type topic.

This allows the user to pass, for each parameter, either a node name or a URI that describes the protocol to use, the hostname, username, and password.

The plan then calls the run_task function, specifying which nodes to run the tasks on. The Target names are collected and stored in $webserver_names by iterating over the list of Target objects returned by get_targets. Task parameters are serialized to JSON format; therefore, extracting the names into an array of strings ensures that the webservers parameter is in a format that can be converted to JSON.

plan mymodule::my_plan(
  TargetSpec $load_balancer,
  TargetSpec  $webservers,
) {

  # Extract the Target name from $webservers
  $webserver_names = get_targets($webservers).map |$n| { $n.name }
  
  # process webservers
  run_task('mymodule::lb_remove', $load_balancer, webservers => $webserver_names)
  run_task('mymodule::update_frontend_app', $webservers, version => '1.2.3')
  run_task('mymodule::lb_add', $load_balancer, webservers => $webserver_names)
 }

To execute this plan from the command line, pass the parameters as parameter=value. The Targetspec accepts either an array as json or a comma separated string of target names.

bolt plan run mymodule::myplan --modulepath ./PATH/TO/MODULES load_balancer=lb.myorg.com webservers='["kermit.myorg.com","gonzo.myorg.com"]'
        

Parameters that are passed to the run_* plan functions are serialized to JSON.

To illustrate this concept, consider this plan:
plan test::parameter_passing (
  TargetSpec $nodes,
  Optional[String[1]] $example_nul = undef,
) {
  return run_task('test::demo_undef_bash', $nodes, example_nul => $example_nul)
     }
The default value of $example_nul is undef. The plan calls the test::demo_undef_bash with the example_nul parameter. The implementation of the demo_undef_bash.sh task is:
#!/bin/bash
example_env=$PT_example_nul
echo "Environment: $PT_example_nul"
echo "Stdin:" 
     cat -

By default, the task expects parameters passed as a JSON string on stdin to be accessible in prefixed environment variables.

Consider the output of running the plan against localhost:
bolt@bolt: bolt plan run test::parameter_passing -n localhost
Starting: plan test::parameter_passing
Starting: task test::demo_undef_bash on localhost
Finished: task test::demo_undef_bash with 0 failures in 0.0 sec
Finished: plan test::parameter_passing in 0.01 sec
Finished on localhost:
  Environment: null
  Stdin:
  {"example_nul":null,"_task":"test::demo_undef_bash"}
  {
  }
Successful on 1 node: localhost
     Ran on 1 node

The parameters example_nul and _task metadata are passed to the task as a JSON string over stdin.

Similarly, parameters are made available to the task as environment variables where the name of the parameter is converted to an environment variable prefixed with PT_. The prefixed environment variable points to the String representation in JSON format of the parameter value. So, the PT_example_nul environment variable has the value of null of type String.

Returning results from plans

Use plans to return results that you can use in other plans or save for use outside of Bolt

Plans, unlike functions, are primarily run for side effects but they can optionally return a result. To return a result from a plan use the return function. Any plan that does not call the return function returns undef.

plan return_result(
  $nodes
) {
  return run_task('mytask', $nodes)
}

The result of a plan must match the PlanResult type alias. This roughly includes JSON types as well as the Plan language types which have well defined JSON representations in Bolt.

  • Undef
  • String
  • Numeric
  • Boolean
  • Target
  • Result
  • ResultSet
  • Error
  • Array with only PlanResult
  • Hash with String keys and PlanResult values

or

Variant[Data, String, Numeric, Boolean, Error, Result, ResultSet, Target, Array[Boltlib::PlanResult], Hash[String, Boltlib::PlanResult]]

Returning errors in plans

To return an error if your plan fails, call the fail_plan function.

Specify parameters to provide details about the failure.

For example, if called with run_plan('mymodule::myplan'), this would return an error to the caller.
plan mymodule::myplan {
  Error(
    message    => "Sorry, this plan does not work yet.",
    kind       => 'mymodule/error',
    issue_code => 'NOT_IMPLEMENTED'
    )
  }
  fail_plan("Sorry, this plan does not work yet.", 'mymodule/error')
}

Plan success and failure

There are indicators that a plan has run successfully or failed.

Any plan that completes execution without an error is considered successful. The bolt command exits 0 and any calling plans continue execution. If any calls to run_ functions fail without _catch_errors then the plan halts execution and is considered a failure. Any calling plans also halt until a run_plan call with _catch_errors or a catch_errors block is reached. If one isn't reached, the bolt command performs an exit 2. When writing a plan if you have reason to believe it has failed, you can fail the plan with the fail_plan function. This causes the bolt command to exit 2 and prevents calling plans executing any further, unless run_plan was called with _catch_errors or in a catch_errors block.

Failing plans

If upload_filerun_commandrun_script, or run_task are called without the _catch_errors option and they fail on any nodes, the plan itself fails. To fail a plan directly call the fail_plan function. Create a new error with a message and include the kind, details, or issue code, or pass an existing error to it.
fail_plan('The plan is failing', 'mymodules/pear-shaped', {'failednodes' => $result.error_set.names})
# or
fail_plan($errorobject)

Catching errors in plans

Bolt includes a catch_errors function that executes a block of code and returns the error if an error is raised, or returns the result of the block if no errors are raised. You might get an Error object returned if you call run_plan with _catch_errors, use a catch_errors block, or call the Error method on a result.

The Error data type includes:

  • msg: The error message string.

  • kind: A string that defines the kind of error similar to an error class.

  • details: A hash with details about the error from a task or from information about the state of a plan when it fails, for example, exit_code or stack_trace.

  • issue_code: A unique code for the message that can be used for translation.

Use the Error data type in a case expression to match against different kind of errors. To recover from certain errors, while failing on or ignoring others, set up your plan to include conditionals based on errors that occur while your plan runs. For example, you can set up a plan to retry a task when a timeout error occurs, but to fail when there is an authentication error.

Below, the first plan continues whether it succeeds or fails with a mymodule/not-serious error. Other errors cause the plan to fail.
plan mymodule::handle_errors {
  $result = run_plan('mymodule::myplan', '_catch_errors' => true)
  case $result {
    Error['mymodule/not-serious'] : {
      notice("${result.message}")
    }
    Error : { fail_plan($result) } }
  run_plan('mymodule::plan2')
}

Using the catch errors function:

plan test (String[1] $role) {
  $result_or_error = catch_errors(['bolt/puppetdb-error']) || {
    puppetdb_query("inventory[certname] { app_role == ${role} }")
  }
  $targets = if $result_or_error =~ Error {
    # If the PuppetDB query fails
    warning("Could not fetch from puppet. Using defaults instead")
    # TargetSpec string
    "all"
  } else {
    $result_or_error
  }
}

Puppet and Ruby functions in plans

You can define and call Puppet language and Ruby functions in plans.

This is useful for packaging common general logic in your plan. You can also call the plan functions, such as run_task or run_plan, from within a function.

Not all Puppet language constructs are allowed in plans. The following constructs are not allowed:

  • Defined types.

  • Classes.

  • Resource expressions, such as file { title: mode => '0777' }

  • Resource default expressions, such as File { mode => '0666' }

  • Resource overrides, such as File['/tmp/foo'] { mode => '0444' }

  • Relationship operators: -> <- ~> <~

  • Functions that operate on a catalog: include, require, contain, create_resources.

  • Collector expressions, such as SomeType <| |>, SomeType <<| |>>

  • ERB templates are not supported. Use EPP instead.

Be aware of a few other Puppet behaviors in plans:

  • The --strict_variables option is on, so if you reference a variable that is not set, you get an error.

  • --strict=error is always on, so minor language issues generate errors. For example { a => 10, a => 20 } is an error because there is a duplicate key in the hash.

  • Most Puppet settings are empty and not-configurable when using Bolt.

  • Logs include "source location" (file, line) instead of resource type or name.

Handling plan function results

Each execution function returns an object type ResultSet. For each node that the execution takes place on, this object contains a Result object. The apply action returns a ResultSet containing ApplyResult objects.

A ResultSet has the following methods:

  • names(): The String names (node URIs) of all nodes in the set as an Array.

  • empty(): Returns Boolean if the execution result set is empty.

  • count(): Returns an Integer count of nodes.

  • first(): The first Result object, useful to unwrap single results.

  • find(String $target_name): Look up the Result for a specific target.

  • error_set(): A ResultSet containing only the results of failed nodes.

  • ok_set(): A ResultSet containing only the successful results.

  • filter_set(block): Filters a ResultSet with the given block and returns a ResultSet object (where the filter function returns an array or hash).

  • targets(): An array of all the Target objects from every Result in the set.

  • ok(): Boolean that is the same as error_nodes.empty.

  • to_data(): An array of hashes representing either Result or ApplyResults.

A Result has the following methods:

  • value(): The hash containing the value of the Result.

  • target(): The Target object that the Result is from.

  • error(): An Error object constructed from the _error in the value.

  • message(): The _output key from the value.

  • ok(): Returns true if the Result was successful.

  • []: Accesses the value hash directly.

  • to_data(): Hash representation of Result.

  • action(): String representation of result type (task, command, etc.).

An ApplyResult has the following methods:

  • report(): The hash containing the Puppet report from the application.

  • target(): The Target object that the Result is from.

  • error(): An Error object constructed from the _error in the value.

  • ok(): Returns true if the Result was successful.

  • to_data(): Hash representation of ApplyResult.

  • action(): String representation of result type (apply).

An instance of ResultSet is Iterable as if it were an Array[Variant[Result, ApplyResult]] so that iterative functions such as each, map, reduce, or filter work directly on the ResultSet returning each result.

This example checks if a task ran correctly on all nodes. If it did not, the check fails:

$r = run_task('sometask', ..., '_catch_errors' => true)
unless $r.ok {
  fail("Running sometask failed on the nodes ${r.error_nodes.names}")
}

You can do iteration and checking if the result is an Error. This example outputs feedback about the result of a task:

$r = run_task('sometask', ..., '_catch_errors' => true)
$r.each |$result| {
  $node = $result.target.name
  if $result.ok {
    notice("${node} returned a value: ${result.value}")
  } else {
    notice("${node} errored with a message: ${result.error.message}")
  }
}
Similarly, you can iterate over the array of hashes returned by calling to_data on aResultSet and access hash values. For example:
$r = run_command('whoami', 'localhost,local://0.0.0.0')
$r.to_data.each |$result_hash| { notice($result_hash['result']['stdout']) } 
You can also use filter_set to filter a ResultSet and apply a ResultSet function such as targets to the output:
$filtered = $result.filter_set |$r| {
  $r['tag'] == "you're it"
}.targets

Passing sensitive data to tasks

Task parameters defined as sensitive are masked when they appear in plans.

You define a task parameter as sensitive with the metadata property "sensitive": true. When a task runs, the values for these sensitive parameters are masked.

run_task('task_with_secrets', ..., password => '$ecret!')

Working with the sensitive function

In Puppet you use the Sensitive function to mask data in output logs. Because plans are written in Puppet DSL, you can use this type freely. The run_task() function does not allow parameters of Sensitive function to be passed. When you need to pass a sensitive value to a task, you must unwrap it prior to calling run_task().
$pass = Sensitive('$ecret!')
run_task('task_with_secrets', ..., password => $pass.unwrap)

Target objects

The Target object represents a node and its specific connection options.

The state of a target is stored in the inventory for the duration of a plan allowing you to collect facts or set vars for a target and retrieve them later. You can get a printable representation via the name function, as well as access components of the target: protocol, host, port, user, password.

TargetSpec

The execution function takes a parameter with the type alias TargetSpec. This alias accepts the pattern strings allowed by --nodes, a single Target object, or an Array of Targets and node patterns. Generally, use this type for plans that accept a set of targets as a parameter, to ensure clean interaction with the CLI and other plans. To operate on individual nodes, resolve it to a list via get_targets. For example, to loop over each node in a plan accept a TargetSpec argument, but call get_targets on it before looping.

plan loop(TargetSpec $nodes) {
  get_targets($nodes).each |$target| {
    run_task('my_task', $target)
  }
}

If your plan accepts a single TargetSpec parameter you can call that parameter nodes so that it can be specified with the --nodes flag from the command line.

Variables and facts on targets

When Bolt runs, it loads transport config values, variables, and facts from the inventory. These can be accessed with the $target.facts() and $target.vars() functions. During the course of a plan, you can update the facts or variables for any target. Facts usually come from running facter or another fact collection application on the target or from a fact store like PuppetDB. Variables are computed externally or assigned directly.

Set variables in a plan using $target.set_var:

plan vars(String $host) {
	$target = get_targets($host)[0]
	$target.set_var('newly_provisioned', true)
	$targetvars = $target.vars
	run_command("echo 'Vars for ${host}: ${$targetvars}'", $host)
}

Or set variables in the inventory file using the vars key at the group level.

groups:
  - name: my_nodes
    nodes:
      - localhost
    vars:
      operatingsystem: windows
    config:
      transport: ssh

Collect facts from the targets

The facts plan connects to the target and discovers facts. It then stores these facts on the targets in the inventory for later use.

The methods used to collect facts:

  • On ssh targets, it runs a Bash script.
  • On winrm targets, it runs a PowerShell script.
  • On pcp or targets where the Puppet agent is present, it runs Facter.
This example collects facts with the facts plan and then uses those facts to decide which task to run on the targets.
plan run_with_facts(TargetSpec $nodes) {
  # This collects facts on nodes and update the inventory
  run_plan(facts, nodes => $nodes)

  $centos_nodes = get_targets($nodes).filter |$n| { $n.facts['os']['name'] == 'CentOS' }
  $ubuntu_nodes = get_targets($nodes).filter |$n| { $n.facts['os']['name'] == 'Ubuntu' }
  run_task(centos_task, $centos_nodes)
  run_task(ubuntu_task, $ubuntu_nodes)
}

Collect facts from PuppetDB

When targets are running a Puppet agent and sending facts to PuppetDB, you can use the puppetdb_fact plan to collect facts for them. This example collects facts with the puppetdb_fact plan, and then uses those facts to decide which task to run on the targets. You must configure the PuppetDB client before you run it.

plan run_with_facts(TargetSpec $nodes) {
  # This collects facts on nodes and update the inventory
  run_plan(puppetdb_fact, nodes => $nodes)

  $centos_nodes = get_targets($nodes).filter |$n| { $n.facts['os']['name'] == 'CentOS' }
  $ubuntu_nodes = get_targets($nodes).filter |$n| { $n.facts['os']['name'] == 'Ubuntu' }
  run_task(centos_task, $centos_nodes)
  run_task(ubuntu_task, $ubuntu_nodes)
}

Collect general data from PuppetDB

You can use the puppetdb_query function in plans to make direct queries to PuppetDB. For example you can discover nodes from PuppetDB and then run tasks on them. You'll have to configure the PuppetDB client before running it. You can learn how to structure pql queries here, and find pql reference and examples here
plan pdb_discover {
  $result = puppetdb_query("inventory[certname] { app_role == 'web_server' }")
  # extract the certnames into an array
  $names = $result.map |$r| { $r["certname"] }
  # wrap in url. You can skip this if the default transport is pcp
  $nodes = $names.map |$n| { "pcp://${n}" }
  run_task('my_task', $nodes)
}

Plan logging

Plan run information can be captured in log files or printed to a terminal session using the following methods.

Outputting section to the terminal

Print message strings to STDOUT using the plan function out::message. This function always prints messages regardless of the log level and doesn't log them to the log file.

Puppet log functions

To generate log messages from a plan, use the Puppet log function that corresponds to the level you want to track: error, warn, notice, info, or debug. Configure the log level for both log files and console logging in bolt.yaml. The default log level for the console is warn and for log files is notice. Use the --debug flag to set the console log level to debug for a single run.

Default action logging

Bolt logs actions that a plan takes on targets through the  upload_file,  run_command, run_script, or run_task  functions. By default it logs a notice level message when an action starts and another when it completes. If you pass a description to the function, that is used in place of the generic log message.

run_task(my_task, $targets, "Better description", param1 => "val")

If your plan contains many small actions you may want to suppress these messages and use explicit calls to the Puppet log functions instead. This can be accomplished by wrapping actions in a without_default_logging block which causes the action messages to be logged at info level instead of notice. For example to loop over a series of nodes without logging each action.

plan deploy( TargetSpec $nodes) {
  without_default_logging() || {
    get_targets($nodes).each |$node| {
      run_task(deploy, $node)
    }
  }
}

To avoid complications with parser ambiguity, always call without_default_logging with () and empty block args ||.

without_default_logging() || { run_command('echo hi', $nodes) }

not

without_default_logging { run_command('echo hi', $nodes) }

Example plans

Check out some example plans for inspiration writing your own.

Resource Description Level
facts module Contains tasks and plans to discover facts about target systems. Getting started
facts plan Gathers facts using the facts task and sets the facts in inventory. Getting started
facts::info plan Uses the facts task to discover facts and map relevant fact values to targets. Getting started
reboot module Contains tasks and plans for managing system reboots. Intermediate
reboot plan Restarts a target system and waits for it to become available again. Intermediate
Introducing Masterless Puppet with Bolt Blog post explaining how plans can be used to deploy a load-balanced web server. Advanced
profiles::nginx_install plan Shows an example plan for deploying Nginx and HAProxy. Advanced
  • Getting started resources show simple use cases such as running a task and manipulating the results.
  • Intermediate resources show more advanced features in the plan language.
  • Advanced resources show more complex use cases such as applying puppet code blocks and using external modules.