The mgmt tool has various frontends, each of which may produce a stream of
between zero or more graphs that are passed to the engine for desired state
application. In almost all scenarios, you're going to want to use the language
frontend. This guide describes some of the internals of the language.
The mgmt language is a declarative (immutable) functional, reactive programming
language. It is implemented in golang. A longer introduction to the language
is available as a blog post here!
All expressions must have a type. A composite type such as a list of strings
([]str) is different from a list of integers ([]int).
There is a variant type in the language's type system, but it is only used internally and only appears briefly when needed for type unification hints during static polymorphic function generation. This is an advanced topic which is not required for normal usage of the software.
The implementation of the internal types can be found in lang/types/.
A true or false value.
Any "string!" enclosed in quotes.
A number like 42 or -13. Integers are represented internally as golang's
int64.
A floating point number like: 3.1415926. Float's are represented internally as
golang's float64.
An ordered collection of values of the same type, eg: [6, 7, 8, 9,]. It is
worth mentioning that empty lists have a type, although without type hints it
can be impossible to infer the item's type.
An unordered set of unique keys of the same type and corresponding value pairs
of another type, eg:
{"boiling" => 100, "freezing" => 0, "room" => 25, "house" => 22, "canada" => -30,}.
That is to say, all of the keys must have the same type, and all of the values
must have the same type. You can use any type for either, although it is
probably advisable to avoid using very complex types as map keys.
An ordered set of field names and corresponding values, each of their own type,
eg: struct{answer => "42", james => "awesome", is_mgmt_awesome => true,}.
These are useful for combining more than one type into the same value. Note the
syntactical difference between these and map's: the key's in map's have types,
and as a result, string keys are enclosed in quotes, whereas struct fields are
not string values, and as such are bare and specified without quotes.
An ordered set of optionally named, differently typed input arguments, and a
return type, eg: func(s str) int or:
func(bool, []str, {str: float}) struct{foo str; bar int}.
Expressions, and the Expr interface need to be better documented. For now
please consume
lang/interfaces/ast.go.
These docs will be expanded on when things are more certain to be stable.
There are a very small number of statements in our language. They include:
-
bind: bind's an expression to a variable within that scope without output
- eg:
$x = 42
- eg:
-
if: produces up to one branch of statements based on a conditional expression
if <conditional> { <statements> } else { # the else branch is optional for if statements <statements> } -
resource: produces a resource
file "/tmp/hello" { content => "world", mode => "o=rwx", } -
edge: produces an edge
File["/tmp/hello"] -> Print["alert4"] -
class: bind's a list of statements to a class name in scope without output
class foo { # some statements go here }or
class bar($a, $b) { # a parameterized class # some statements go here } -
include: include a particular class at this location producing output
include foo include bar("hello", 42) include bar("world", 13) # an include can be called multiple times -
import: import a particular scope from this location at a given namespace
# a system module import import "fmt" # a local, single file import (relative path, not a module) import "dir1/file.mcl" # a local, module import (relative path, contents are a module) import "dir2/" # a remote module import (absolute remote path, contents are a module) import "git://github.com/purpleidea/mgmt-example1/"or
import "fmt" as * # contents namespaced into top-level names import "foo.mcl" # namespaced as foo import "dir1/" as bar # namespaced as bar import "git://github.com/purpleidea/mgmt-example1/" # namespaced as example1
All statements produce output. Output consists of between zero and more
edges and resources. A resource statement can produce a resource, whereas an
if statement produces whatever the chosen branch produces. Ultimately the goal
of executing our programs is to produce a list of resources, which along with
the produced edges, is built into a resource graph. This graph is then passed
to the engine for desired state application.
This section needs better documentation.
This section needs better documentation.
Resources express the idempotent workloads that we want to have apply on our
system. They correspond to vertices in a graph
which represent the order in which their declared state is applied. You will
usually want to pass in a number of parameters and associated values to the
resource to control how it behaves. For example, setting the content parameter
of a file resource to the string hello, will cause the contents of that file
to contain the string hello after it has run.
For some parameters, there is a distinction between an unspecified parameter,
and a parameter with a zero value. For example, for the file resource, you
might choose to set the content parameter to be the empty string, which would
ensure that the file has a length of zero. Alternatively you might wish to not
specify the file contents at all, which would leave that property undefined. If
you omit listing a property, then it will be undefined. To control this property
programmatically, you need to specify an is-defined value, as well as the
value to use if that boolean is true. You can do this with the resource-specific
elvis operator.
$b = true # change me to false and then try editing the file manually
file "/tmp/mgmt-elvis" {
content => $b ?: "hello world\n",
state => $const.res.file.state.exists,
}
This example is static, however you can imagine that the $b value might be
chosen in a programmatic way, even one in which that value varies over time. If
it evaluates to true, then the parameter will be used. If no elvis operator
is specified, then the parameter value will also be used. If the parameter is
not specified, then it will obviously not be used.
Resources may specify meta parameters. To do so, you must add them as you would
a regular parameter, except that they start with Meta and are capitalized. Eg:
file "/tmp/f1" {
content => "hello!\n",
Meta:noop => true,
Meta:delay => $b ?: 42,
Meta:autoedge => false,
}
As you can see, they also support the elvis operator, and you can add as many as you like. While it is not recommended to add the same meta parameter more than once, it does not currently cause an error, and even though the result of doing so is officially undefined, it will currently take the last specified value.
You may also specify a single meta parameter struct. This is useful if you'd like to reuse a value, or build a combined value programmatically. For example:
file "/tmp/f1" {
content => "hello!\n",
Meta => $b ?: struct{
noop => false,
retry => -1,
delay => 0,
poll => 5,
limit => 4.2,
burst => 3,
sema => ["foo:1", "bar:3",],
autoedge => true,
autogroup => false,
},
}
Remember that the top-level Meta field supports the elvis operator, while the
individual struct fields in the struct type do not. This is to be expected, but
since they are syntactically similar, it is worth mentioning to avoid confusion.
Please note that at the moment, you must specify a full metaparams struct, since partial struct types are currently not supported in the language. Patches are welcome if you'd like to add this tricky feature!
Each resource must have a unique name of type str that is used to uniquely
identify that resource, and can be used in the functioning of the resource at
that resources discretion. For example, the file resource uses the unique name
value to specify the path.
Alternatively, the name value may be a list of strings []str to build a list
of resources, each with a name from that list. When this is done, each resource
will use the same set of parameters. The list of internal edges specified in the
same resource block is created intelligently to have the appropriate edge for
each separate resource.
Using this construct is a veiled form of looping (iteration). This technique is
one of many ways you can perform iterative tasks that you might have
traditionally used a for loop for instead. This is preferred, because flow
control is error-prone and can make for less readable code.
Resources may also declare edges internally. The edges may point to or from
another resource, and may optionally include a notification. The four properties
are: Before, Depend, Notify and Listen. The first two represent normal
edge dependencies, and the second two are normal edge dependencies that also
send notifications. You may have multiples of these per resource, including
multiple Depend lines if necessary. Each of these properties also supports the
conditional inclusion elvis operator as well.
For example, you may write:
$b = true # for example purposes
if $b {
pkg "drbd" {
state => "installed",
# multiple properties may be used in the same resource
Before => File["/etc/drbd.conf"],
Before => Svc["drbd"],
}
}
file "/etc/drbd.conf" {
content => "some config",
Depend => $b ?: Pkg["drbd"],
Notify => Svc["drbd"],
}
svc "drbd" {
state => "running",
}
There are two unique properties about these edges that is different from what you might expect from other automation software:
- The ability to specify multiples of these properties allows you to avoid having to manage arrays and conditional trees of these different dependencies.
- The keywords all have the same length, which means your code lines up nicely.
Edges express dependencies in the graph of resources which are output. They can be chained as a pair, or in any greater number. For example, you may write:
Pkg["drbd"] -> File["/etc/drbd.conf"] -> Svc["drbd"]
to express a relationship between three resources. The first character in the resource kind must be capitalized so that the parser can't ascertain unambiguously that we are referring to a dependency relationship.
A class is a grouping structure that bind's a list of statements to a name in
the scope where it is defined. It doesn't directly produce any output. To
produce output it must be called via the include statement.
Defining classes follows the same scoping and shadowing rules that is applied to
the bind statement, although they exist in a separate namespace. In other
words you can have a variable named foo and a class named foo in the same
scope without any conflicts.
Classes can be both parameterized or naked. If a parameterized class is defined,
then the argument types must be either specified manually, or inferred with the
type unification algorithm. One interesting property is that the same class
definition can be used with include via two different input signatures,
although in practice this is probably fairly rare. Some usage examples include:
A naked class definition:
class foo {
# some statements go here
}
A parameterized class with both input types being inferred if possible:
class bar($a, $b) {
# some statements go here
}
A parameterized class with one type specified statically and one being inferred:
class baz($a str, $b) {
# some statements go here
}
Classes can also be nested within other classes. Here's a contrived example:
import "fmt"
class c1($a, $b) {
# nested class definition
class c2($c) {
test $a {
stringptr => fmt.printf("%s is %d", $b, $c),
}
}
if $a == "t1" {
include c2(42)
}
}
Defining polymorphic classes was considered but is not currently allowed at this time.
Recursive classes are not currently supported and it is not clear if they will be in the future. Discussion about this topic is welcome on the mailing list.
The include statement causes the previously defined class to produce the
contained output. This statement must be called with parameters if the named
class is defined with those.
The defined class can be called as many times as you'd like either within the same scope or within different scopes. If a class uses inferred type input parameters, then the same class can even be called with different signatures. Whether the output is useful and whether there is a unique type unification solution is dependent on your code.
The import statement imports a scope into the specified namespace. A scope can
contain variable, class, and function definitions. All are statements.
Furthermore, since each of these have different logical uses, you could
theoretically import a scope that contains an int variable named foo, a
class named foo, and a function named foo as well. Keep in mind that
variables can contain functions (they can have a type of function) and are
commonly called lambdas.
There are a few different kinds of imports. They differ by the string contents
that you specify. Short single word, or multiple-word tokens separated by zero
or more slashes are system imports. Eg: math, fmt, or even math/trig.
Local imports are path imports that are relative to the current directory. They
can either import a single mcl file, or an entire well-formed module. Eg:
file1.mcl or dir1/. Lastly, you can have a remote import. This must be an
absolute path to a well-formed module. The common transport is git, and it can
be represented via an FQDN. Eg: git://github.com/purpleidea/mgmt-example1/.
The namespace that any of these are imported into depends on how you use the
import statement. By default, each kind of import will have a logic namespace
identifier associated with it. System imports use the last token in their name.
Eg: fmt would be imported as fmt and math/trig would be imported as
trig. Local imports do the same, except the required .mcl extension, or
trailing slash are removed. Eg: foo/file1.mcl would be imported as file1 and
bar/baz/ would be imported as baz. Remote imports use some more complex
rules. In general, well-named modules that contain a final directory name in the
form: mgmt-whatever/ will be named whatever. Otherwise, the last path token
will be converted to lowercase and the dashes will be converted to underscores.
The rules for remote imports might change, and should not be considered stable.
In any of the import cases, you can change the namespace that you're imported
into. Simply add the as whatever text at the end of the import, and whatever
will be the name of the namespace. Please note that whatever is not surrounded
by quotes, since it is an identifier, and not a string. If you'd like to add
all of the import contents into the top-level scope, you can use the as * text
to dump all of the contents in. This is generally not recommended, as it might
cause a conflict with another identifier.
The mgmt compiler runs in a number of stages. In order of execution they are:
- Lexing
- Parsing
- Interpolation
- Scope propagation
- Type unification
- Function graph generation
- Function engine creation and validation
All of the above needs to be done every time the source code changes. After this point, the function engine runs and produces events. On every event, we "interpret" which produces a resource graph. This series of resource graphs are passed to the engine as they are produced.
What follows are some notes about each step.
Lexing is done using nex. It is a pure-golang
implementation which is similar to Lex or Flex, but which produces golang
code instead of C. It integrates reasonably well with golang's yacc which is
used for parsing. The token definitions are in:
lang/lexer.nex.
Lexing and parsing run together by calling the LexParse method.
The parser used is golang's implementation of
yacc. The documentation is
quite abysmal, so it's helpful to rely on the documentation from standard yacc
and trial and error. One small advantage yacc has over standard yacc is that it
can produce error messages from examples. The best documentation is to examine
the source. There is a short write up available here.
The yacc file exists at:
lang/parser.y.
Lexing and parsing run together by calling the LexParse method.
Interpolation is used to transform the AST (which was produced from lexing and
parsing) into one which is either identical or different. It expands strings
which might contain expressions to be interpolated (eg: "the answer is: ${foo}")
and can be used for other scenarios in which one statement or expression would
be better represented by a larger AST. Most nodes in the AST simply return their
own node address, and do not modify the AST.
Scope propagation passes the parent scope (starting with the top-level, built-in
scope) down through the AST. This is necessary so that children nodes can access
variables in the scope if needed. Most AST node's simply pass on the scope
without making any changes. The ExprVar node naturally consumes scope's and
the StmtProg node cleverly passes the scope through in the order expected for
the out-of-order bind logic to work.
This step typically calls the ordering algorithm to determine the correct order of statements in a program.
Each expression must have a known type. The unpleasant option is to force the
programmer to specify by annotation every type throughout their whole program
so that each Expr node in the AST knows what to expect. Type annotation is
allowed in situations when you want to explicitly specify a type, or when the
compiler cannot deduce it, however, most of it can usually be inferred.
For type inferrence to work, each node in the AST implements a Unify method
which is able to return a list of invariants that must hold true. This starts at
the top most AST node, and gets called through to it's children to assemble a
giant list of invariants. The invariants can take different forms. They can
specify that a particular expression must have a particular type, or they can
specify that two expressions must have the same types. More complex invariants
allow you to specify relationships between different types and expressions.
Furthermore, invariants can allow you to specify that only one invariant out of
a set must hold true.
Once the list of invariants has been collected, they are run through an
invariant solver. The solver can return either return successfully or with an
error. If the solver returns successfully, it means that it has found a trivial
mapping between every expression and it's corresponding type. At this point it
is a simple task to run SetType on every expression so that the types are
known. If the solver returns in error, it is usually due to one of two
possibilities:
-
Ambiguity
The solver does not have enough information to make a definitive or unique determination about the expression to type mappings. The set of invariants is ambiguous, and we cannot continue. An error will be returned to the programmer. In this scenario the user will probably need to add a type annotation, possibly because of a design bug in the user's program.
-
Conflict
The solver has conflicting information that cannot be reconciled. In this situation an explicit conflict has been found. If two invariants are found which both expect a particular expression to have different types, then it is not possible to find a valid solution. This almost always happens if the user has made a type error in their program.
Only one solver currently exists, but it is possible to easily plug in an alternate implementation if someone more skilled in the art of solver design would like to propose a more logical or performant variant.
At this point we have a fully type AST. The AST must now be transformed into a directed, acyclic graph (DAG) data structure that represents the flow of data as necessary for everything to be reactive. Note that this graph is different from the resource graph which is produced and sent to the engine. It is just a coincidence that both happen to be DAG's. (You don't freak out when you see a list data structure show up in more than one place, do you?)
To produce this graph, each node has a Graph method which it can call. This
starts at the top most node, and is called down through the AST. The edges in
the graphs must represent the individual expression values which are passed
from node to node. The names of the edges must match the function type argument
names which are used in the definition of the corresponding function. These
corresponding functions must exist for each expression node and are produced by
calling that expression's Func method. These are usually called by the
function engine during function creation and validation.
Finally we have a graph of the data flows. The function engine must first
initialize which creates references to each of the necessary function
implementations, and gets information about each one. It then needs to be type
checked to ensure that the data flows all correctly match what is expected. If
you were to pass an int to a function expecting a bool, this would be a
problem. If all goes well, the program should get run shortly.
At this point the function engine runs. It produces a stream of events which
cause the Output() method of the top-level program to run, which produces the
list of resources and edges. These are then transformed into the resource graph
which is passed to the engine.
If you'd like to create a built-in, core function, you'll need to implement the
function API interface named Func. It can be found in
lang/interfaces/func.go.
Your function must have a specific type. For example, a simple math function
might have a signature of func(x int, y int) int. As you can see, all the
types are known before compile time.
A separate discussion on this matter can be found in the function guide.
What follows are each of the method signatures and a description of each. Failure to implement the API correctly can cause the function graph engine to block, or the program to panic.
Info() *InfoThe Info method must return a struct containing some information about your function. The struct has the following type:
type Info struct {
Sig *types.Type // the signature of the function, must be KindFunc
}You must implement this correctly. Other fields in the Info struct may be
added in the future. This method is usually called before any other, and should
not depend on any other method being called first. Other methods must not depend
on this method being called first.
func (obj *FooFunc) Info() *interfaces.Info {
return &interfaces.Info{
Sig: types.NewType("func(a str, b int) float"),
}
}Init(*Init) errorInit is called by the function graph engine to create an implementation of this function. It is passed in a struct of the following form:
type Init struct {
Hostname string // uuid for the host
Input chan types.Value // Engine will close `input` chan
Output chan types.Value // Stream must close `output` chan
World resources.World
Debug bool
Logf func(format string, v ...interface{})
}These values and references may be used (wisely) inside your function. Input
will contain a channel of input structs matching the expected input signature
for your function. Output will be the channel which you must send values to
whenever a new value should be produced. This must be done in the Stream()
function. You may carefully use World to access functionality provided by the
engine. You may use Logf to log informational messages, however there is no
guarantee that they will be displayed to the user. Debug specifies whether the
function is running in a user-requested debug mode. This might cause you to want
to print more log messages for example. You will need to save references to any
or all of these info fields that you wish to use in the struct implementing this
Func interface. At a minimum you will need to save Output as a minimum of
one value must be produced.
Please see the example functions in
[lang/funcs/core/](https://github.com/purpleidea/mgmt/tree/master/lang/funcs/core/).Stream() errorStream is called by the function engine when it is ready for your function to
start accepting input and producing output. You must always produce at least one
value. Failure to produce at least one value will probably cause the function
engine to hang waiting for your output. This function must close the Output
channel when it has no more values to send. The engine will close the Input
channel when it has no more values to send. This may or may not influence
whether or not you close the Output channel.
Please see the example functions in
[lang/funcs/core/](https://github.com/purpleidea/mgmt/tree/master/lang/funcs/core/).Close() errorClose asks the particular function to shutdown its Stream() function and
return.
Please see the example functions in
[lang/funcs/core/](https://github.com/purpleidea/mgmt/tree/master/lang/funcs/core/).For some functions, it might be helpful to be able to implement a function once,
but to have multiple polymorphic variants that can be chosen at compile time.
For this more advanced topic, you will need to use the
Polymorphic Function API. This will help with code
reuse when you have a small, finite number of possible type signatures, and also
for more complicated cases where you might have an infinite number of possible
type signatures. (eg: []str, or [][]str, or [][][]str, etc...)
Suppose you want to implement a function which can assume different type
signatures. The mgmt language does not support polymorphic types-- you must use
static types throughout the language, however, it is legal to implement a
function which can take different specific type signatures based on how it is
used. For example, you might wish to add a math function which could take the
form of func(x int, x int) int or func(x float, x float) float depending on
the input values. You might also want to implement a function which takes an
arbitrary number of input arguments (the number must be statically fixed at the
compile time of your program though) and which returns a string.
The PolyFunc interface adds additional methods which you must implement to
satisfy such a function implementation. If you'd like to implement such a
function, then please notify the project authors, and they will expand this
section with a longer description of the process.
What follows are a few examples that might help you understand some of the language details.
TODO: please add an example here!
TODO: please add an example here!
(Send your questions as a patch to this FAQ! I'll review it, merge it, and respond by commit with the answer.)
The language contains both an if expression, and and if statement. An if
expression takes a boolean conditional and it must contain exactly two
branches (a then and an else branch) which each contain one expression. The
if expression will return the value of one of the two branches based on the
conditional.
# this is an if expression, and both branches must exist
$b = true
$x = if $b {
42
} else {
-13
}
The if statement also takes a boolean conditional, but it may have either one
or two branches. Branches must only directly contain statements. The if
statement does not return any value, but it does produce output when it is
evaluated. The output consists primarily of resources (vertices) and edges.
# this is an if statement, and in this scenario the else branch was omitted
$b = true
if $b {
file "/tmp/hello" {
content => "world",
}
}
In the lang/types library, there is a types.Value interface. Every value in
our type system must implement this interface. One of the methods in this
interface is the String() string method. This lets you print a representation
of the value. You will probably never need to use this method.
In addition, the types.Value interface implements a number of helper functions
which return the value as an equivalent golang type. If you know that the value
is a bool, you can call x.Bool() on it. If it's a string you can call
x.Str(). Make sure not to call one of those type methods unless you know the
value is of that type, or you will trigger a panic!
If you create a base type like bool, str, int, or float, all you need to
do is build the &BoolValue and set the V field. Eg:
someBool := &types.BoolValue{V: true}If you are building a container type like list, map, struct, or func,
then you also need to specify the type of the contained values. This is
because a list has a type of []str, or []int, or even [][]foo. Eg:
someListOfStrings := &types.ListValue{
T: types.NewType("[]str"), # must match the contents!
V: []types.Value{
&types.StrValue{V: "a"},
&types.StrValue{V: "bb"},
&types.StrValue{V: "ccc"},
},
}If you don't build these properly, then you will cause a panic! Even empty lists have a type.
Not really, but practically it can be used as such. The class statement is not
a singleton since it can be called multiple times in different locations, and it
can also be parameterized and called multiple times (with include) using
different input parameters. The reason it can be used as such is that statement
output (from multple classes) that is compatible (and usually identical) will
be automatically collated and have the duplicates removed. In that way, you can
assume that an unparameterized class is always a singleton, and that
parameterized classes can often be singletons depending on their contents and if
they are called in an identical way or not. In reality the de-duplication
actually happens at the resource output level, so anything that produces
multiple compatible resources is allowed.
Recursive class definitions where the contents of a class contain a
self-referential include, either directly, or with indirection via any other
number of classes is not supported. It's not clear if it ever will be in the
future, unless we decide it's worth the extra complexity. The reason is that our
FRP actually generates a static graph which doesn't change unless the code does.
To support dynamic graphs would require our FRP to be a "higher-order" FRP,
instead of the simpler "first-order" FRP that it is now. You might want to
verify that I got the nomenclature
correct. If it turns out that there's an important advantage to supporting a
higher-order FRP in mgmt, then we can consider that in the future.
I realized that recursion would require a static graph when I considered the structure required for a simple recursive class definition. If some "depth" value wasn't known statically by compile time, then there would be no way to know how large the graph would grow, and furthermore, the graph would need to change if that "depth" value changed.
Yes, the language is just one of the available "frontends" that passes a stream of graphs to the engine "backend". While it is the recommended way of using mgmt, you're welcome to either use an alternate frontend, or write your own. To write your own frontend, you must implement the GAPI interface.
I am certainly no expert in FRP, and I've certainly got lots more to learn. One thing FRP experts might notice is that some of the concepts from FRP are either named differently, or are notably absent.
In mgmt, we don't talk about behaviours, events, or signals in the strict FRP definitons of the words. Firstly, because we only support discretized, streams of values with no plan to add continuous semantics. Secondly, because we prefer to use terms which are more natural and relatable to what our target audience is expecting. Our users are more likely to have a background in Physiology, or systems administration than a background in FRP.
Having said that, we hope that the FRP community will engage with us and help improve the parts that we got wrong. Even if that means adding continuous behaviours!
Thank you, and yes, probably. "Props" may also be accepted, although patches are preferred. If you can't do either, donations to support the project are welcome too!
Additional blog posts, videos and other material is available!.
If you have any ideas for changes or other improvements to the language, please let us know! We're still pre 1.0 and pre 0.1 and happy to change it in order to get it right!