ash/documentation/topics/expressions.md

Expressions

Ash expressions are used in various places like calculations, filters, and policies, and are meant to be portable representations of elixir expressions. You can create an expression using the Ash.Expr.expr/1 macro, like so:

Ash.Expr.expr(1 + 2)
Ash.Expr.expr(x + y)
Ash.Expr.expr(post.title <> " | " <> post.subtitle)

Ash expressions have some interesting properties in their evaluation, primarily because they are made to be portable, i.e executable in some data layer (like SQL) or executable in Elixir. In general, these expressions will behave the same way they do in Elixir. The primary difference is how nil values work. They behave the way that NULL values behave in SQL. This is primarily because this pattern is easier to replicate to various popular data layers, and is generally safer when using expressions for things like authentication. The practical implications of this are that nil values will "poison" many expressions, and cause them to return nil. For example, x + nil would always evaluate to nil. Additionally, true and nil will always result in nil, this is also true with or and not, i.e true or nil will return nil, and not nil will return nil.

Operators

The following operators are available and they behave the same as they do in Elixir, except for the nil addendum above.

• ==
• !=
• >
• >=
• <
• <=
• in
• *
• -
• /
• <>
• ||
• &&
• is_nil | Only works as an operator in maps/keyword list syntax. i.e [x: [is_nil: true]]

Functions

The following functions are built in. Data Layers can add their own functions to expressions. For example, AshPostgres adds trigram_similarity function.

The following functions are built in:

• if | Works like elixir's if.
• is_nil/1 | Works like elixir's is_nil
• get_path/2 | i.e get_path(value, ["foo", "bar"]). This is what expressions like value[:foo]["bar"] are turned into under the hood.
• contains/2 | if one string contains another string, i.e contains("fred", "red")
• length/1 | the length of a list, i.e. length([:foo, :bar])
• type/2 | Cast a given value to a specific type, i.e type(^arg(:id), :uuid) or type(integer_field, :string)
• string_downcase/1 | Downcases a string
• string_join/1 | Concatenates a list of strings, and ignores any nil values
• string_join/2 | As above, but with a joiner
• string_split/1 | Splits a string on spaces
• string_split/2 | As above, but with a specific delimiter
• string_split/3 | As above, but with options. See the function for the available options.
• string_length/1 | Returns the length of a given string, as reported by String.length/1
• string_trim/1 | Trims unicode whitespace from the beginning and the end of a string
• at/2 | Get an element from a list, i.e at(list, 1)
• round/1 | Round a float, decimal or int to 0 precision, i.e round(num)
• round/2 | Round a float, decimal or int to the provided precision or less, i.e round(1.1234, 3) == 1.1234 and round(1.12, 3) == 1.12
• String interpolation | "#{first_name} #{last_name}", is remapped to the equivalent usage of <>
• fragment/* | Creates a fragment of a data layer expression. See the section on fragments below.

Fragments

Fragments come in two forms.

String Fragments

For SQL/query-backed data layers, they will be a string with question marks for interpolation. For example: fragment("(? + ?)", foo, bar).

Function Fragments

For elixir-backed data layers, they will be a function or an MFA that will be called with the provided arguments. For example: fragment(&Module.add/2, foo, bar) or fragment({Module, :add, []}, foo, bar). When using anonymous functions, you can only use the format &Module.function/arity. &Module.add/2 is okay, but fn a, b -> Module.add(a, b) end is not.

Sub-expressions

• exists/2 | exists(foo.bar, name == "fred") takes an expression scoped to the destination resource, and checks if any related entry matches. See the section on exists below.
• path.exists/2 | Same as exists but the source of the relationship is itself a nested relationship. See the section on exists below.
• parent/1 | Allows an expression scoped to a resource to refer to the "outer" context. Used in relationship filters and exists

DateTime Functions

• now/0 | Evaluates to the current time when the expression is evaluated
• today/0 | Evaluates to the current date when the expression is evaluated
• ago/2 | i.e deleted_at > ago(7, :day). The available time intervals are documented in Ash.Type.DurationName
• from_now/2 | Same as ago but adds instead of subtracting
• datetime_add/3 | add an interval to a datetime, i.e datetime_add(^datetime, 10, :hour)
• date/3 | add an interval to a date, i.e datetime_add(^date, 3, :day)

Primitives

• cond - cond is transformed to a series of if expressions under the hood
• item[:key] or item["key"] - accesses keys in a map. In both cases, it prefers a matching atom key, falling back to a matching string key. This is to aid with data stores that store embeds as JSON with string keys (like AshPostgres), so that this expression behaves the same in the data layer as it does in Elixir.

Escape Hatches

• lazy/1 - Takes an MFA and evaluates it just before running the query. This is important for calculations, because the expression/2 callback should be stable (returns the same value given the same input). For example lazy({ULID, :generate, [timestamp_input]})

Inline Aggregates

Aggregates can be referenced in-line, with their relationship path specified and any options provided that match the options given to Ash.Query.Aggregate.new/4. For example:

calculate :grade, :decimal, expr(
  count(answers, query: [filter: expr(correct == true)]) /
  count(answers, query: [filter: expr(correct == false)])
)

The available aggregate kinds can also be seen in the Ash.Query.Aggregate module documentation.

Templates

Most of the time, when you are using an expression, you will actually be creating a template. In this template, you have a few references that can be used, which will be replaced when before the expression is evaluated. The following references are available.

^actor(:key) # equivalent to `get_in(actor || %{}, [:key])`
^actor([:key1, :key2]) # equivalent to `get_in(actor || %{}, [:key, :key2])`
^arg(:arg_name) # equivalent to `Map.get(arguments, :arg_name)`
^context(:key) # equivalent to `get_in(context, :key)`
^context([:key1, :key2]) # equivalent to `get_in(context, [:key1, :key2])`
^ref(:key) # equivalent to referring to `key`. Allows for dynamic references
^ref(:key, [:path]) # equivalent to referring to `path.key`. Allows for dynamic references with dynamic (or static) paths.

Custom Expressions

Custom expressions allow you to extend Ash's expression language with your own expressions. To see more, see Ash.CustomExpression. To add a custom expression, configure it and recompile ash. For example:

config :ash, :custom_expressions, [
  MyApp.CustomExpression
]

mix deps.compile ash --force

These expressions will be available across all usages of Ash expressions within your application.

Use cases for expressions

Filters

The most obvious place we use expressions is when filtering data. For example:

Ash.Query.filter(Ticket, status == :open and opened_at >= ago(10, :day))

These filters will be run in the data layer, i.e in the SQL query.

Filter semantics & joins

The semantics of Ash filters are probably slightly different than what you are used to, and they are important to understand. Every filter expression is always talking about a single row, potentially "joined" to single related rows. By referencing relationships, you are implicitly doing a join. For those familiar with SQL terminology, it is equivalent to a left join, although AshPostgres can detect when it is safe to do an inner join (for performance reason). Lets use an example of posts and comments.

Given a filter like the following:

Ash.Query.filter(Post, comments.points > 10 and comments.tag.name == "elixir")

The filter refers to a single post/comment/tag combination. So in english, this is "posts where they have a comment with more than 10 points and that same comment has a tag with the name elixir". What this also means is that filters like the above do not compose nicely when new filters are added. For example:

def has_comment_with_more_points_than(query, score) do
  Ash.Query.filter(Post, comments.points > 10)
end

def has_comment_tagged(query, tag) do
  Ash.Query.filter(Post, comments.tag.name == ^tag)
end

Post
|> has_comment_with_more_points_than(query, 10)
|> has_comment_tagged("elixir")

That code seems like it ought to produce a filter over Post that would give us any post with a comment having more than 10 points, and with a comment tagged elixir. That is not the same thing as having a single comment that meets both those criteria. So how do we make this better?

Exists

Lets rewrite the above using exists:

def has_comment_with_more_points_than(query, score) do
  Ash.Query.filter(Post, exists(comments, points > ^score))
end

def has_comment_tagged(query, tag) do
  Ash.Query.filter(Post, exists(comments.tag.name == ^tag)
end

Post
|> has_comment_with_more_points_than(query, ^score)
|> has_comment_tagged("elixir")

Now, they will compose properly! Generally speaking, you should use exists when you are filtering across any relationships that are to_many relationships *even if you don't expect your filter to be composed. Currently, the filter syntax does not minimize(combine) these exists/2 statements, but doing so is not complex and can be added. While unlikely, please lodge an issue if you see any performance issues with exists.

Exists at path

Sometimes, you want the ability to say that some given row must have an existing related entry matching a filter. For example:

Ash.Query.filter(Post, author.exists(roles, name == :admin) and author.active)

While the above is not common, it can be useful in some specific circumstances, and is used under the hood by the policy authorizer when combining the filters of various resources to create a single filter.

Relationship Filters

When filtering relationships, you can use the parent/1 function to scope a part of the expression to "source" of the join. This allows for very expressive relationships! Keep in mind, however, that if you want to update and/or manage these relationships, you'll have to make sure that any attributes that make these things actually related are properly set.

has_many :descendents, __MODULE__ do
  description "All descendents in the same tree"
  no_attributes? true # this says that there is no matching source_attribute and destination_attribute on this relationship
  # This is an example using postgres' ltree extension.
  filter expr(tree_id == parent(tree_id) and fragment("? @> ?", parent(path), path))
end

Portability

Ash expressions being portable is more important than it sounds. For example, if you were using AshPostgres and had the following calculation, which is an expression capable of being run in elixir or translated to SQL:

calculate :full_name, :string, expr(first_name <> " " <> last_name)

And you did something like the following:

User
|> Ash.Query.load(:full_name)
|> Ash.Query.sort(:full_name)
|> Accounts.read!()

You would see that it ran a SQL query with the full_name calculation as SQL. This allows for sorting on that value. However, if you had something like this:

# data can be loaded in the query like above, or on demand later
Accounts.load!(user, :full_name)

you would see that no SQL queries are run. The calculation is run directly in Elixir and the value is set.

Parent

Parent is a way to "jump out" of a scoped expression. Here are some examples:

Ash.Query.filter(exists(open_tickets, severity >= parent(severity_threshold)))

has_many :relevant_tickets, Ticket do
  filter expr(status == :open and severity >= parent(severity_threshold))
end

count :count_of_relevant_tickets, :open_tickets do
  filter expr(status == :open and severity >= parent(severity_threshold))
end

Referencing related values

Related values can be references using dot delimiters, i.e Ash.Query.filter(user.first_name == "fred"). When referencing related values in filters, if the reference is a has_one or belongs_to, the filter does exactly what it looks like (matches if the related value matches). If it is a has_many or a many_to_many, it matches if any of the related records match.

Referencing aggregates and calculations

Aggregates are simple, as all aggregates can be referenced in filter expressions (if you are using a data layer that supports it).

For calculations, only those that define an expression can be referenced in other expressions.

Here are some examples:

# given a `full_name` calculation

Ash.Query.filter(User, full_name == "Hob Goblin")

# given a `full_name` calculation that accepts an argument called `delimiter`

Ash.Query.filter(User, full_name(delimiter: "~") == "Hob~Goblin")