Skip to main content

Angle Guide

The following guide will explain Angle from first principles, leading you through from simple queries to more complex ones.

If you want to try the examples for yourself, or experiment with changes to the example schema, you should first follow the instructions in Walkthrough to get set up.

Just the facts​

Data in Glean is described by a schema, which we normally put in a file with the extension angle. For the purposes of this guide we’ll use the example schema in example.angle. Full details about defining schemas can be found in Schemas. The example.angle file contains a schema definition like this:

schema example.1 {

# definitions go here

}

This says we’re defining a schema called example, with version 1.

The schema contains definitions for predicates. A predicate is the type of facts, which are the individual pieces of information that Glean stores. Our example schema models a simplified class hierarchy for an object-oriented language, starting with a predicate for a Class:

predicate Class :
  {
     name : string,
     line : nat,
  }

This says that the facts of Class are records with two fields, a name field which contains a string, and a line field which contains a nat (“nat” is short for “natural number”, which is limited to 64 bits in Glean).

The simplest type of Angle query is one that just selects facts from the database that match a pattern. For our first Angle query, let’s find a class by its name:

facts> example.Class { name = "Pet" }
{ "id": 1024, "key": { "name": "Pet", "line": 10 } }

1 results, 1 facts, 4.61ms, 117632 bytes, 677 compiled bytes

(The last line contains statistics about query performance from Glean; I’ll leave this out in the rest of the examples.)

What’s going on here?

  • The query consists of the predicate name example.Class followed by a pattern { name = "Pet" }
  • Note that when we refer to a predicate in a query, the name is qualified by prefixing the schema name, so it’s example.Class rather than just Class.
  • The query returns all the facts of example.Class that match the pattern

The shell shows results in JSON format. When you’re making Glean queries from code, the results will normally be decoded into native data types that you can manipulate directly in whatever language you’re using; for more details see Thrift and JSON.

Note that each fact has a unique id. This is how Glean identifies facts in its database. As a user you normally won’t have to worry about fact ids; you can think of them like memory addresses.

The pattern specifies which facts to return. In the example above, our pattern is matching a record type and specifying a subset of the fields: just the name field. We could match the line field instead:

facts> example.Class { line = 20 }
{ "id": 1025, "key": { "name": "Lizard", "line": 20 } }
note

Your patterns should normally match fields at the beginning of the record, because facts in the database are indexed by a prefix of the fields. Matching a field in the middle of the record works by scanning all the facts, which could be expensive. We’ll get into this in more detail in Query Efficiency.

What other kinds of patterns can we use? Well, the simplest patterns are the wildcard, “_”, which matches anything, and "never", which always fails to match.

facts> example.Class _
{ "id": 1026, "key": { "name": "Fish", "line": 30 } }
{ "id": 1027, "key": { "name": "Goldfish", "line": 40 } }
{ "id": 1025, "key": { "name": "Lizard", "line": 20 } }
{ "id": 1024, "key": { "name": "Pet", "line": 10 } }
facts> example.Class never
(no results)

We’ll introduce more kinds of pattern in the following sections. The full list of patterns can be found in Angle Reference.

Matching nested facts​

The real power of Glean comes from relationships between facts. Facts can refer directly to other facts, and we can write queries that directly match on these connections.

Our example schema has a predicate that expresses the inheritance relationship between classes:

predicate Parent :
  {
     child : Class,
     parent : Class,
  }

Let’s find what Fish inherits from:

facts> example.Parent { child = { name = "Fish" }}
{
  "id": 1029,
  "key": { "child": { "id": 1026, "key": { "name": "Fish", "line": 30 } }, "parent": { "id": 1024, "key": { "name": "Pet", "line": 10 } } }
}

Let’s break this down.

  • { child = { name = "Fish" }} is a pattern that matches the key type of Parent
  • So, looking at the schema, { name = "Fish" } is a pattern that should match the Class in the field child.

By default Angle queries recursively expand facts in the results. We can see in the above result that the child and parent fields contain the full facts they point to. If we want the result to be “shallow”, meaning it contains just the facts that match and not the nested facts, we can ask Glean to not expand the content of those references. In the shell this is done by running the command :expand off:

facts> :expand off
facts> example.Parent { child = { name = "Fish" }}
{ "id": 1029, "key": { "child": { "id": 1026 }, "parent": { "id": 1024 } } }

We can of course go the other way and find all the children of a class:

facts> example.Parent { parent = { name = "Pet" }}
{
  "id": 1028,
  "key": {
    "child": { "id": 1025, "key": { "name": "Lizard", "line": 20 } },
    "parent": { "id": 1024, "key": { "name": "Pet", "line": 10 } }
  }
}
{
  "id": 1029,
  "key": {
    "child": { "id": 1026, "key": { "name": "Fish", "line": 30 } },
    "parent": { "id": 1024, "key": { "name": "Pet", "line": 10 } }
  }
}

But as before, note that this would be an inefficient query if we had a lot of data because the pattern is matching on the second field of Parent (namely parent). Later we’ll see how to make these queries more efficient using a derived predicate.

Union types​

Our examples so far have dealt with record types. Glean also supports union types, also called sum types, which are used to express multiple alternatives. For example, let’s expand our schema to include class members which can be either a method or a variable:

predicate Has :
  {
    class_ : Class,
    has : Member,
    access : enum { Public | Private },
  }

predicate Member :
  {
    method : { name : string, doc : maybe string } |
    variable : { name : string }
  }
}

The predicate Has maps a Class to a Member (with a Public or Private annotation), and a Member is either method or variable, with some associated data. Note that a Class might have more than one Member, which is fine: there can be multiple Has facts for a given Class.

note

The schema uses class_ rather than class as a field name, because class is a reserved word in Angle. There are many such reserved words, which are reserved not because Angle uses them, but because they cause problems for code that is automatically generated from the schema. To avoid having too many ad-hoc language-specific naming rules, Glean prevents certain problematic names from being used in the schema. The Angle compiler will tell you if you try to use a reserved word.

Let’s find classes that have a variable called fins:

facts> example.Has { has = { variable = { name = "fins" }}}
{
  "id": 1036,
  "key": {
    "class_": { "id": 1026, "key": { "name": "Fish", "line": 30 } },
    "has": { "id": 1035, "key": { "variable": { "name": "fins" } } },
    "access": 1
  }
}

The key thing here is that we matched on Member which is a union type, using the pattern { variable = { name = "fins" }}. A pattern to match a union type looks very much like a record pattern, but it can have only a single field, in this case either variable or method.

Maybe​

Glean has one built-in union type called maybe, which is useful when we want to have optional values in the data. It's used in our example schema to attach optional documentation to a class member:

predicate Member :
  {
    method : { name : string, doc : maybe string } |
    variable : { name : string }
  }

The type maybe string behaves exactly as if it were defined as the union type { nothing | just : string }. That means we can write a pattern that matches it, exactly as we would write a pattern for { nothing | just : string }:

Methods without documentation:

facts> example.Member { method = { doc = nothing } }

Methods with documentation:

facts> example.Member { method = { doc = {  just = _ }}}

Or-patterns​

In a pattern we can express multiple alternatives by separating patterns with a vertical bar |.

For example, we can find classes on lines 20 or 30:

facts> example.Class { line = 20 | 30 }
{ "id": 1025, "key": { "name": "Lizard", "line": 20 } }
{ "id": 1026, "key": { "name": "Fish", "line": 30 } }

Or we can find all the classes that have either a method called feed or a variable with any name:

facts> example.Has { has = { method = { name = "feed" }} | { variable = _ }}

(results omitted)

If-patterns​

We can conditionally match patterns using if then else.

Variables matched in the condition will be available in the then branch.

Whilst an or-pattern will always evaluate both of its branches, the else branch of an if-pattern will never be evaluated if the condition succeeds at least once.

For example, we could get all child classes if inheritance is being used in the codebase, or retrieve all classes if it isn't.

facts > if (example.Parent { child = X }) then X else example.Class _
  { "id": 1025, "key": { "name": "Lizard", "line": 20 } }
  { "id": 1026, "key": { "name": "Fish", "line": 30 } }
  { "id": 1027, "key": { "name": "Goldfish", "line": 40 } }

Please note that if-patterns cannot be used in stored derived predicates. This is the case because they require the use of negation, which is disallowed in stored predicates.

More complex queries​

So far we’ve seen how to query for facts by matching patterns, including matching nested facts. In this section we’ll see how to construct more complex queries that combine matching facts from multiple predicates.

Suppose we want to find all the parents of classes that have a variable called fins. We need to build a query that will

  • find the classes with a variable called fins using example.Has as we did above
  • find their parents using example.Parent

We can combine these two as follows:

example.Has
  {
    class_ = C,
    has = { variable = { name = "fins" }}
  };
example.Parent { child = C }
note

I’ve written this on several lines with indentation to illustrate it better, to do this in the shell you will need to use the :edit command to put the query in a temporary file.

The key thing here is that we used a variable C to stand for the class_ field when matching facts of example.Has, and then we searched for example.Parent facts with the same value of C for the child field.

Note that variables must always begin with an upper-case letter, while schema names (example) and field names (child) begin with a lower-case letter.

The semicolon separates multiple statements in a query. When there are multiple statements the results of the query are the facts that match the last statement, in this case the example.Parent. Let’s try it:

facts> example.Has { class_ = C, has = { variable = { name = "fins" }}}; example.Parent { child = C }
{
  "id": 1029,
  "key": {
    "child": { "id": 1026, "key": { "name": "Fish", "line": 30 } },
    "parent": { "id": 1024, "key": { "name": "Pet", "line": 10 } }
  }
}

Suppose we don’t care too much about the child here, we only care about getting a list of the parents. We can avoid returning the redundant information by specifying explicitly what it is we want to return from the query:

P where
    example.Has
      {
        class_ = C,
        has = { variable = { name = "fins" }}
      };
    example.Parent { child = C, parent = P }

The general form of the query is expression where statements, where expression is an arbitrary expression and each statement is a pattern that matches some facts. The results of the query are the distinct values of expression for which all the statements match facts in the database.

facts> P where example.Has { class_ = C, has = { variable = { name = "fins" }}}; example.Parent { child = C, parent = P }
{ "id": 1024, "key": { "name": "Pet", "line": 10 } }

Statements​

In general, a statement can be of the form A = B. For example, if we write

C = example.Class { name = "Fish" };
example.Parent { child = C }

that’s the same as

example.Parent { child = { name = "Fish" }}

A statement can have a pattern on either side, for example

C where
  C = example.Class { name = N };
  N = "Fish" | "Goldfish"

A statement can itself be a set of alternatives separated by a vertical bar |. For example, we can find classes that are either a parent of the Goldfish or have a feed method:

C where
  example.Parent { child = { name = "Goldfish" }, parent = C } |
  example.Has { class_ = C, has = { method = { name = "feed" }}}

Arrays​

When the schema uses an array, we need to be able to write queries that traverse the elements of the array. For example, a common use of an array is to represent the list of declarations in a source file. Our example schema defines the FileClasses predicate:

predicate FileClasses :
  {
    file : string,
    classes : [Class]
  }

The goal here is to map efficiently from a filename to the list of classes defined in that file. Suppose we want to write a query that finds all the classes called Goldfish in the file petshop.example, we could do it like this:

example.FileClasses { file = "petshop.example", classes = Cs };
{ name = "Goldfish" } = Cs[..]

The second line is the interesting one: { name = "Goldfish" } = Cs[..] means

  • on the right-hand side, Cs[..] means “each element of the array Cs”
  • the left-hand side is a pattern, filtering only those Class facts that match { name = "Goldfish" }

We can also match the whole array with a pattern of the form [ p1, p2, .. ]

facts> X where [_,X,_] = [1,2,3]
{ "id": 1040, "key": 2 }

Or if we don't care about the length of the array:

facts> X where [_,X, ..] = [1,2,3]
{ "id": 1040, "key": 2 }

String prefix​

We’ve seen many examples of patterns that match strings. Glean also supports matching strings by prefix; for example:

facts> example.Class { name = "F".. }
{ "id": 1026, "key": { "name": "Fish", "line": 30 } }

The syntax "F".. means strings beginning with the prefix ”F".

note

Why only prefix and not substring matching in general? Prefix matching can be supported efficiently by Glean’s prefix-tree representation of the fact database. Other kinds of string matching could be supported, but they wouldn’t be able to exploit the database representation so there’s little advantage to implementing them in Angle compared with filtering on the client-side.

Tuples​

A tuple is just a a way of writing a record without the field names. So for example, instead of

example.Parent { child = C }

we could write

example.Parent { C, _ }

When using a tuple you have to list all the fields, in the same order as they are declared in the schema. That's why { child = C } becomes { C, _ } when written as a tuple.

There are upsides and downsides to using the tuple notation:

  • Pro: more concise
  • Con: brittle and sensitive to changes in the schema. If we add a field, then tuple patterns will break whereas record patterns won't.

As a rule of thumb we tend to use tuple syntax in cases where the predicate is "obviously" a relation, such as example.Parent, but we wouldn't use tuple syntax for more complex records.

Enums and bool​

An enum type is a set of named constants. In the Has predicate we used an enum type to indicate whether a class member is Public or Private:

predicate Has :
  {
    class_ : Class,
    has : Member,
    access : enum { Public | Private },
  }

To match an enum we just use the appropriate identifier, in this case Public or Private:

facts> example.Has { access = Private }
{ "id": 1036, "key": { "class_": { "id": 1026 }, "has": { "id": 1035 }, "access": 1 } }

Note that in the JSON format results, an enum is represented by an integer. When you make queries in code, the enum will be represented by an appropriate type, such as a data type in Haskell.

The boolean type bool is a special case of an enum, defined like this:

type bool = enum { false | true }

Negation​

If we want results that do not match a certain criterion, we can use ! to specify a subquery that should fail. A subquery fails if it doesn't return any result.

For example, we can find classes that don't have methods

facts> C where C = example.Class _; !(example.Has { class_ = C, has = { method = _ } })
{ "id": 1026, "key": { "name": "Fish", "line": 30 } }
{ "id": 1027, "key": { "name": "Goldfish", "line": 40 } }
{ "id": 1025, "key": { "name": "Lizard", "line": 20 } }

Or we could find the maximum element in an array

facts> X where Values = [5,1,2,3]; X = Values[..]; !(Y = Values[..]; Y > X)
{ "id": 1091, "key": 5 }

The query asks for the X for which given all values of Y none is greater than it. If Y = Values[..] were outside of the negation, the meaning would be give me all X for which there is at least one Y that is not greater than it. The answer to that would be all elements.