Fetching contributors…
Cannot retrieve contributors at this time
809 lines (618 sloc) 30.1 KB

Resource guide


The mgmt tool has built-in resource primitives which make up the building blocks of any configuration. Each instance of a resource is mapped to a single vertex in the resource graph. This guide is meant to instruct developers on how to write a brand new resource. Since mgmt and the core resources are written in golang, some prior golang knowledge is assumed.


Resources in mgmt are similar to resources in other systems in that they are idempotent. Our resources are uniquely different in that they can detect when their state has changed, and as a result can run to revert or repair this change instantly. For some background on this design, please read the original article on the subject.

Resource Prerequisites


You'll need to import a few packages to make writing your resource easier. Here is the list:


The engine package contains most of the interfaces and helper functions that you'll need to use. The traits package contains some base functionality which you can use to easily add functionality to your resource without needing to implement it from scratch.

Resource struct

Each resource will implement methods as pointer receivers on a resource struct. The naming convention for resources is that they end with a Res suffix.

The resource struct should include an anonymous reference to the Base trait. Other traits can be added to the resource to add additional functionality. They are discussed below.

You'll most likely want to store a reference to the *Init struct type as defined by the engine. This is data that the engine will provide to your resource on Init.

Lastly you should define the public fields that make up your resource API, as well as any private fields that you might want to use throughout your resource. Do not depend on global variables, since multiple copies of your resource could get instantiated.

You'll want to add struct tags based on the different frontends that you want your resources to be able to use. Some frontends can infer this information if it is not specified, but others cannot, and some might poorly infer if the struct name is ambiguous.

If you'd like your resource to be accessible by the YAML graph API (GAPI), then you'll need to include the appropriate YAML fields as shown below. This is used by the Puppet compiler as well, so make sure you include these struct tags if you want existing Puppet code to be able to run using the mgmt engine.


type FooRes struct {
	traits.Base // add the base methods without re-implementation

	init *engine.Init

	Whatever string `lang:"whatever" yaml:"whatever"` // you pick!
	Baz      bool   `lang:"baz" yaml:"baz"`           // something else

	something string // some private field

Resource API

To implement a resource in mgmt it must satisfy the Res interface. What follows are each of the method signatures and a description of each.


Default() engine.Res

This returns a populated resource struct as a Res. It shouldn't populate any values which already have the correct default as the golang zero value. In general it is preferable if the zero values make for the correct defaults.


// Default returns some sensible defaults for this resource.
func (obj *FooRes) Default() Res {
	return &FooRes{
		Answer: 42, // sometimes, defaults shouldn't be the zero value


Validate() error

This method is used to validate if the populated resource struct is a valid representation of the resource kind. If it does not conform to the resource specifications, it should return an error. If you notice that this method is quite large, it might be an indication that you should reconsider the parameter list and interface to this resource. This method is called by the engine before Init. It can also be called occasionally after a Send/Recv operation to verify that the newly populated parameters are valid. Remember not to expect access to the outside world when using this.


// Validate reports any problems with the struct definition.
func (obj *FooRes) Validate() error {
	if obj.Answer != 42 { // validate whatever you want
		return fmt.Errorf("expected an answer of 42")
	return nil


Init() error

This is called to initialize the resource. If something goes wrong, it should return an error. It should do any resource specific work such as initializing channels, sync primitives, or anything else that is relevant to your resource. If it is not need throughout, it might be preferable to do some initialization and tear down locally in either the Watch method or CheckApply method. The choice depends on your particular resource and making the best decision requires some experience with mgmt. If you are unsure, feel free to ask an existing mgmt contributor. During Init, the engine will pass your resource a struct containing some useful data and pointers. You should save a copy of this pointer since you will need to use it in other parts of your resource.


// Init initializes the Foo resource.
func (obj *FooRes) Init(init *engine.Init) error
	obj.init = init // save for later

	// run the resource specific initialization, and error if anything fails
	if some_error {
		return err // something went wrong!
	return nil

This method is always called after Validate has run successfully, with the exception that we can't prevent a malicious or buggy libmgmt user to not run this. In other words, you should expect Validate to have run first, but you shouldn't allow Init to dangerously rm -rf /$the_world if your code only checks $the_world in Validate. Remember to always program safely!


Close() error

This is called to cleanup after the resource. It is usually not necessary, but can be useful if you'd like to properly close a persistent connection that you opened in the Init method and were using throughout the resource. It is not the shutdown signal that tells the resource to exit. That happens in the Watch loop.


// Close runs some cleanup code for this resource.
func (obj *FooRes) Close() error {
	err := obj.conn.Close() // close some internal connection
	obj.someMap = nil       // free up some large data structure from memory
	return err

You should probably check the return errors of your internal methods, and pass on an error if something went wrong.


CheckApply(apply bool) (checkOK bool, err error)

CheckApply is where the real work is done. Under normal circumstances, this function should check if the state of this resource is correct, and if so, it should return: (true, nil). If the apply variable is set to true, then this means that we should then proceed to run the changes required to bring the resource into the correct state. If the apply variable is set to false, then the resource is operating in noop mode and no operational changes should be made!

After having executed the necessary operations to bring the resource back into the desired state, or after having detected that the state was incorrect, but that changes can't be made because apply is false, you should then return (false, nil).

You must cause the resource to converge during a single execution of this function. If you cannot, then you must return an error! The exception to this rule is that if an external force changes the state of the resource while it is being remedied, it is possible to return from this function even though the resource isn't now converged. This is not a bug, as the resources Watch facility will detect the new change, ultimately resulting in a subsequent call to CheckApply.


// CheckApply does the idempotent work of checking and applying resource state.
func (obj *FooRes) CheckApply(apply bool) (bool, error) {
	// check the state
	if state_is_okay { return true, nil } // done early! :)

	// state was bad

	if !apply { return false, nil } // don't apply, we're in noop mode

	if any_error { return false, err } // anytime there's an err!

	// do the apply!
	return false, nil // after success applying

The CheckApply function is called by the mgmt engine when it believes a call is necessary. Under certain conditions when a Watch call does not invalidate the state of the resource, and no refresh call was sent, its execution might be skipped. This is an engine optimization, and not a bug. It is mentioned here in the documentation in case you are confused as to why a debug message you've added to the code isn't always printed.

Paired execution

For many resources it is not uncommon to see CheckApply run twice in rapid succession. This is usually not a pathological occurrence, but rather a healthy pattern which is a consequence of the event system. When the state of the resource is incorrect, CheckApply will run to remedy the state. In response to having just changed the state, it is usually the case that this repair will trigger the Watch code! In response, a second CheckApply is triggered, which will likely find the state to now be correct.


  • Anytime an error occurs during CheckApply, you should return (false, err).
  • If the state is correct and no changes are needed, return (true, nil).
  • You should only make changes to the system if apply is set to true.
  • After checking the state and possibly applying the fix, return (false, nil).
  • Returning (true, err) is a programming error and can have a negative effect.


Watch() error

Watch is a main loop that runs and sends messages when it detects that the state of the resource might have changed. To send a message you should write to the input event channel using the Event helper method. The Watch function should run continuously until a shutdown message is received. If at any time something goes wrong, you should return an error, and the mgmt engine will handle possibly restarting the main loop based on the retry meta parameter.

It is better to send an event notification which turns out to be spurious, than to miss a possible event. Resources which can miss events are incorrect and need to be re-engineered so that this isn't the case. If you have an idea for a resource which would fit this criteria, but you can't find a solution, please contact the mgmt maintainers so that this problem can be investigated and a possible system level engineering fix can be found.

You may have trouble deciding how much resource state checking should happen in the Watch loop versus deferring it all to the CheckApply method. You may want to put some simple fast path checking in Watch to avoid generating obviously spurious events, but in general it's best to keep the Watch method as simple as possible. Contact the mgmt maintainers if you're not sure.

If the resource is activated in polling mode, the Watch method will not get executed. As a result, the resource must still work even if the main loop is not running.


The lifetime of most resources Watch method should be spent in an infinite loop that is bounded by a select call. The select call is the point where our method hands back control to the engine (and the kernel) so that we can sleep until something of interest wakes us up. In this loop we must process events from the engine via the <-obj.init.Events channel, and receive events for our resource itself!


If we receive an internal event from the <-obj.init.Events channel, we should read it with the obj.init.Read helper function. This function tells us if we should shutdown our resource. It also handles pause functionality which blocks our resource temporarily in this method. If this channel shuts down, then we should treat that as an exit signal.

When we want to send an event, we use the Event helper function. It is also important to mark the resource state as dirty if we believe it might have changed. We do this by calling the obj.init.Dirty function.


Once the Watch function has finished starting up successfully, it is important to generate one event to notify the mgmt engine that we're now listening successfully, so that it can run an initial CheckApply to ensure we're safely tracking a healthy state and that we didn't miss anything when Watch was down or from before mgmt was running. You must do this by calling the obj.init.Running method. If it returns an error, you must exit and return that error.


The engine might be asked to shutdown when the entire state of the system has not seen any changes for some duration of time. The engine can determine this automatically, but each resource can block this if it is absolutely necessary. If you need this functionality, please contact one of the maintainers and ask about adding this feature and improving these docs right here.

This particular facility is most likely not required for most resources. It may prove to be useful if a resource wants to start off a long operation, but avoid sending out erroneous Event messages to keep things alive until it finishes.


// Watch is the listener and main loop for this resource.
func (obj *FooRes) Watch() error {
	// setup the Foo resource
	var err error
	if err, = OpenFoo(); err != nil {
		return err // we couldn't startup
	defer obj.whatever.CloseFoo() // shutdown our Foo

	// notify engine that we're running
	if err := obj.init.Running(); err != nil {
		return err // exit if requested

	var send = false // send event?
	for {
		select {
		case event, ok := <-obj.init.Events:
			if !ok {
				// shutdown engine
				// (it is okay if some `defer` code runs first)
				return nil
			if err := obj.init.Read(event); err != nil {
				return err

		// the actual events!
		case event := <
			if is_an_event {
				send = true
				obj.init.Dirty() // dirty

		// event errors
		case err := <
			return err // will cause a retry or permanent failure

		// do all our event sending all together to avoid duplicate msgs
		if send {
			send = false
			if err := obj.init.Event(); err != nil {
				return err // exit if requested


  • Remember to call Running when the Watch is running successfully.
  • Remember to process internal events and shutdown promptly if asked to.
  • Ensure the design of your resource is well thought out.
  • Have a look at the existing resources for a rough idea of how this all works.


Cmp(engine.Res) error

Each resource must have a Cmp method. It is an abbreviation for Compare. It takes as input another resource and must return whether they are identical or not. This is used for identifying if an existing resource can be used in place of a new one with a similar set of parameters. In particular, when switching from one graph to a new (possibly identical) graph, this avoids recomputing the state for resources which don't change or that are sufficiently similar that they don't need to be swapped out.

In general if all the resource properties are identical, then they usually don't need to be changed. On occasion, not all of them need to be compared, in particular if they store some generated state, or if they aren't significant in some way.

If the resource is identical, then you should return nil. If it is not, then you should return a short error message which gives the reason it differs.


// Cmp compares two resources and returns if they are equivalent.
func (obj *FooRes) Cmp(r engine.Res) error {
	// we can only compare FooRes to others of the same resource kind
	res, ok := r.(*FooRes)
	if !ok {
		return fmt.Errorf("not a %s", obj.Kind())

	if obj.Whatever != res.Whatever {
		return fmt.Errorf("the Whatever param differs")
	if obj.Flag != res.Flag {
		return fmt.Errorf("the Flag param differs")

	return nil // they must match!


Resources can have different traits, which means they can be extended to have additional functionality or special properties. Those special properties are usually added by extending your resource so that it is compatible with additional interface that contain the Res interface. Each of these interfaces represents the additional functionality. Since in most cases this requires some common boilerplate, you can usually get some or most of the functionality by embedding the correct trait struct anonymously in your struct. This is shown in the struct example above. You'll always want to include the Base trait in all resources. This provides some basics which you'll always need.

What follows are a list of available traits.


Some resources may choose to support receiving refresh notifications. In general these should be avoided if possible, but nevertheless, they do make sense in certain situations. Resources that support these need to verify if one was sent during the CheckApply phase of execution. This is accomplished by calling the obj.init.Refresh() bool method, and inspecting the return value. This is only necessary if you plan to perform a refresh action. Refresh actions should still respect the apply variable, and no system changes should be made if it is false. Refresh notifications are generated by any resource when an action is applied by that resource and are transmitted through graph edges which have enabled their propagation. Resources that currently perform some refresh action include svc, timer, and password.

It is very important that you include the traits.Refreshable struct in your resource. If you do not include this, then calling obj.init.Refresh may trigger a panic. This is programmer error.


Edgeable is a trait that allows your resource to automatically connect itself to other resources that use this trait to add edge dependencies between the two. An older blog post on this topic is available.

After you've included this trait, you'll need to implement two methods on your resource.


UIDs() []engine.ResUID

The UIDs method returns a list of ResUID interfaces that represent the particular resource uniquely. This is used with the AutoEdges API to determine if another resource can match a dependency to this one.


AutoEdges() (engine.AutoEdge, error)

This returns a struct that implements the AutoEdge interface. This struct is used to match other resources that might be relevant dependencies for this resource.


Groupable is a trait that can allow your resource automatically group itself to other resources. Doing so can reduce the resource or runtime burden on the engine, and improve performance in some scenarios. An older blog post on this topic is available.


Sendable is a trait that allows your resource to send values through the graph edges to another resource. These values are produced during CheckApply. They can be sent to any resource that has an appropriate parameter and that has the Recvable trait. You can read more about this in the Send/Recv section below.


Recvable is a trait that allows your resource to receive values through the graph edges from another resource. These values are consumed during the CheckApply phase, and can be detected there as well. They can be received from any resource that has an appropriate value and that has the Sendable trait. You can read more about this in the Send/Recv section below.


This is currently a stub and will be updated once the DSL is further along.

Resource Initialization

During the resource initialization in Init, the engine will pass in a struct containing a bunch of data and methods. What follows is a description of each one and how it is used.


Program is a string containing the name of the program. Very few resources need this.


Hostname is the uuid for the host. It will be occasionally useful in some resources. It is preferable if you can avoid depending on this. It is possible that in the future this will be a channel which changes if the local hostname changes.


Running must be called after your watches are all started and ready. It is only called from within Watch. It is used to notify the engine that you're now ready to detect changes.


Event sends an event notifying the engine of a possible state change. It is only called from within Watch.


Events is a channel that we must watch for messages from the engine. When it closes, this is a signal to shutdown. It is only called from within Watch.


Read processes messages that come in from the Events channel. It is a helper method that knows how to handle the pause mechanism correctly. It is only called from within Watch.


Dirty marks the resource state as dirty. This signals to the engine that CheckApply will have some work to do in order to converge it. It is only called from within Watch.


Refresh returns whether the resource received a notification. This flag can be used to tell a svc to reload, or to perform some state change that wouldn't otherwise be noticed by inspection alone. You must implement the Refreshable trait for this to work. It is only called from within CheckApply.


Send exposes some variables you wish to send via the Send/Recv mechanism. You must implement the Sendable trait for this to work. It is only called from within CheckApply.


Recv provides a map of variables which were sent to this resource via the Send/Recv mechanism. You must implement the Recvable trait for this to work. It is only called from within CheckApply.


World provides a connection to the outside world. This is most often used for communicating with the distributed database. It can be used in Init, CheckApply and Watch. Use with discretion and understanding of the internals if needed in Close.


VarDir is a facility for local storage. It is used to return a path to a directory which may be used for temporary storage. It should be cleaned up on resource Close if the resource would like to delete the contents. The resource should not assume that the initial directory is empty, and it should be cleaned on Init if that is a requirement.


Debug signals whether we are running in debugging mode. In this case, we might want to log additional messages.


Logf is a logging facility which will correctly namespace any messages which you wish to pass on. You should use this instead of the log package directly for production quality resources.

Further considerations

There is some additional information that any resource writer will need to know. Each issue is listed separately below!

Resource registration

All resources must be registered with the engine so that they can be found. This also ensures they can be encoded and decoded. Make sure to include the following code snippet for this to work.

func init() { // special golang method that runs once
	// set your resource kind and struct here (the kind must be lower case)
	engine.RegisterResource("foo", func() engine.Res { return &FooRes{} })

YAML Unmarshalling

To support YAML unmarshalling for your resource, you must implement an additional method. It is recommended if you want to use your resource with the Puppet compiler.

UnmarshalYAML(unmarshal func(interface{}) error) error // optional

This is optional, but recommended for any resource that will have a YAML accessible struct. It is not required because to do so would mean that third-party or custom resources (such as those someone writes to use with libmgmt) would have to implement this needlessly.

The signature intentionally matches what is required to satisfy the go-yaml Unmarshaler interface.


// UnmarshalYAML is the custom unmarshal handler for this struct.
// It is primarily useful for setting the defaults.
func (obj *FooRes) UnmarshalYAML(unmarshal func(interface{}) error) error {
	type rawRes FooRes // indirection to avoid infinite recursion

	def := obj.Default()     // get the default
	res, ok := def.(*FooRes) // put in the right format
	if !ok {
		return fmt.Errorf("could not convert to FooRes")
	raw := rawRes(*res) // convert; the defaults go here

	if err := unmarshal(&raw); err != nil {
		return err

	*obj = FooRes(raw) // restore from indirection with type conversion!
	return nil


In mgmt there is a novel concept called Send/Recv. For some background, please read the introductory article. When using this feature, the engine will automatically send the user specified value to the intended destination without requiring much resource specific code. Any time that one of the destination values is changed, the engine automatically marks the resource state as dirty. To detect if a particular value was received, and if it changed (during this invocation of CheckApply) from the previous value, you can query the obj.init.Recv() method. It will contain a map of all the keys which can be received on, and the value has a Changed property which will indicate whether the value was updated on this particular CheckApply invocation. The type of the sending key must match that of the receiving one. This can only be done inside of the CheckApply function!

// inside CheckApply, probably near the top
if val, exists := obj.init.Recv()["SomeKey"]; exists {
	obj.init.Logf("the SomeKey param was sent to us from: %s.%s", val.Res, val.Key)
	if val.Changed {
		obj.init.Logf("the SomeKey param was just updated!")
		// you may want to invalidate some local cache

The specifics of resource sending are not currently documented. Please send a patch here!

Composite resources

Composite resources are resources which embed one or more existing resources. This is useful to prevent code duplication in higher level resource scenarios. The best example of this technique can be seen in the nspawn resource which can be seen to partially embed a svc resource, but without its Watch. Unfortunately no further documentation about this subject has been written. To expand this section, please send a patch! Please contact us if you'd like to work on a resource that uses this feature, or to add it to an existing one!

Frequently asked questions

(Send your questions as a patch to this FAQ! I'll review it, merge it, and respond by commit with the answer.)

Can I write resources in a different language?

Currently golang is the only supported language for built-in resources. We might consider allowing external resources to be imported in the future. This will likely require a language that can expose a C-like API, such as python or ruby. Custom golang resources are already possible when using mgmt as a lib. Higher level resource collections will be possible once the mgmt DSL is ready.

Why does the resource API have CheckApply instead of two separate methods?

In an early version we actually had both "parts" as separate methods, namely: StateOK (Check) and Apply, but the decision was made to merge the two into a single method. There are two reasons for this:

  1. Many situations would involve the engine running both Check and Apply. If the resource needed to share some state (for efficiency purposes) between the two calls, this is much more difficult. A common example is that a resource might want to open a connection to dbus or http to do resource state testing and applying. If the methods are combined, there's no need to open and close them twice. A counter argument might be that you could open the connection in Init, and close it in Close, however you might not want that open for the full lifetime of the resource if you only change state occasionally.
  2. Suppose you came up with a really good reason why you wanted the two methods to be separate. It turns out that the current CheckApply can wrap this easily. It would look approximately like this:
func (obj *FooRes) CheckApply(apply bool) (bool, error) {
	// my private split implementation of check and apply
	if c, err := obj.check(); err != nil {
		return false, err // we errored
	} else if c {
		return true, nil // state was good!

	if !apply {
		return false, nil // state needs fixing, but apply is false

	err := obj.apply() // errors if failure or unable to apply

	return false, err // always return false, with an optional error

Feel free to use this pattern if you're convinced it's necessary. Alternatively, if you think I got the Res API wrong and you have an improvement, please let us know!

What new resource primitives need writing?

There are still many ideas for new resources that haven't been written yet. If you'd like to contribute one, please contact us and tell us about your idea!

Is the resource API stable? Does it ever change?

Since we are pre 1.0, the resource API is not guaranteed to be stable, however it is not expected to change significantly. The last major change kept the core functionality nearly identical, simplified the implementation of all the resources, and took about five to ten minutes to port each resource to the new API. The fundamental logic and behaviour behind the resource API has not changed since it was initially introduced.

Where can I find more information about mgmt?

Additional blog posts, videos and other material is available!.


If you have any ideas for API changes or other improvements to resource writing, please let us know! We're still pre 1.0 and pre 0.1 and happy to break API in order to get it right!