SQLMesh macros

Macro systems: two approaches

SQLMesh macros behave differently than those of templating systems like Jinja.

Macro systems are based on string substitution. The macro system scans code files, identifies special characters that signify macro content, and replaces the macro elements with other text.

In a general sense, that is the entire functionality of templating systems. They have tools that provide control flow logic (if-then) and other functionality, but that functionality is solely to support substituting in the correct strings.

Templating systems are intentionally agnostic to the programming language being templated, and most of them work for everything from blog posts to HTML to SQL.

In contrast, SQLMesh macros are designed specifically for generating SQL code. They have semantic understanding of the SQL code being created by analyzing it with the Python sqlglot library, and they allow use of Python code so users can tidily implement sophisticated macro logic.

SQLMesh macro approach

This section describes how SQLMesh macros work under the hood. Feel free to skip over this section and return if and when it is useful. This information is not required to use SQLMesh macros, but it will be useful for debugging any macros exhibiting puzzling behavior.

The critical distinction between the SQLMesh macro approach and templating systems is the role string substitution plays. In templating systems, string substitution is the entire and only point.

In SQLMesh, string substitution is just one step toward modifying the semantic representation of the SQL query. SQLMesh macros work by building and modifying the semantic representation of the SQL query.

After processing all the non-SQL text, it uses the substituted values to modify the semantic representation of the query to its final state.

It uses the following five step approach to accomplish this:

1. Parse the text with the appropriate sqlglot SQL dialect (e.g., Postgres, BigQuery, etc.). During the parsing, it detects the special macro symbol @ to differentiate non-SQL from SQL text. The parser builds a semantic representation of the SQL code's structure, capturing non-SQL text as "placeholder" values to use in subsequent steps.

2. Examine the placeholder values to classify them as one of the following types:

   • Creation of user-defined macro variables with the @DEF operator (see more about user-defined macro variables)
   • Macro variables: SQLMesh pre-defined, user-defined local, and user-defined global
   • Macro functions, both SQLMesh's and user-defined

3. Substitute macro variable values where they are detected. In most cases, this is direct string substitution as with a templating system.

4. Execute any macro functions and substitute the returned values.

5. Modify the semantic representation of the SQL query with the substituted variable values from (3) and functions from (4).

User-defined variables

SQLMesh supports three kinds of user-defined macro variables: global, gateway, and local.

Global and gateway macro variables are defined in the project configuration file and can be accessed in any project model. Local macro variables are defined in a model definition and can only be accessed in that model.

Macro variables with the same name may be specified at any or all of the global, gateway, and local levels. When variables are specified at multiple levels, the value of the most specific level takes precedence. For example, the value of a local variable takes precedence over the value of a gateway variable with the same name, and the value of a gateway variable takes precedence over the value of a global variable.

Global variables

Global variables are defined in the project configuration file variables key.

Global variable values may be any of the following data types or lists or dictionaries containing these types: int, float, bool, str.

Access global variable values in a model definition using the @<VAR_NAME> macro or the @VAR() macro function. The latter function requires the name of the variable in single quotes as the first argument and an optional default value as the second argument. The default value is a safety mechanism used if the variable name is not found in the project configuration file.

For example, this SQLMesh configuration key defines six variables of different data types:

YAMLPython

variables:
  int_var: 1
  float_var: 2.0
  bool_var: true
  str_var: "cat"
  list_var: [1, 2, 3]
  dict_var:
    key1: 1
    key2: 2

variables = {
    "int_var": 1,
    "float_var": 2.0,
    "bool_var": True,
    "str_var": "cat",
    "list_var": [1, 2, 3],
    "dict_var": {"key1": 1, "key2": 2},
}

config = Config(
    variables=variables,
    ... # other Config arguments
)

A model definition could access the int_var value in a WHERE clause like this:

SELECT *
FROM table
WHERE int_variable = @INT_VAR

Alternatively, the same variable can be accessed by passing the variable name into the @VAR() macro function. Note that the variable name is in single quotes in the call @VAR('int_var'):

SELECT *
FROM table
WHERE int_variable = @VAR('int_var')

A default value can be passed as a second argument to the @VAR() macro function, which will be used as a fallback value if the variable is missing from the configuration file.

In this example, the WHERE clause would render to WHERE some_value = 0 because no variable named missing_var was defined in the project configuration file:

SELECT *
FROM table
WHERE some_value = @VAR('missing_var', 0)

A similar API is available for Python macro functions via the evaluator.var method and Python models via the context.var method.

Gateway variables

Like global variables, gateway variables are defined in the project configuration file. However, they are specified in a specific gateway's variables key:

YAMLPython

gateways:
  my_gateway:
    variables:
      int_var: 1
    ...

gateway_variables = {
  "int_var": 1
}

config = Config(
    gateways={
      "my_gateway": GatewayConfig(
        variables=gateway_variables
        ... # other GatewayConfig arguments
        ),
      }
)

Access them in models using the same methods as global variables.

Gateway-specific variable values take precedence over variables with the same name specified in the root variables key.

Local variables

Local macro variables are defined in a model. Local variable values take precedence over global or gateway-specific variables with the same name.

Define your own local macro variables with the @DEF macro operator. For example, you could set the macro variable macro_var to the value 1 with:

@DEF(macro_var, 1);

SQLMesh has three basic requirements for using the @DEF operator:

1. The MODEL statement must end with a semi-colon ;
2. All @DEF uses must come after the MODEL statement and before the SQL query
3. Each @DEF use must end with a semi-colon ;

For example, consider the following model sqlmesh_example.full_model from the SQLMesh quickstart guide:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  audits (assert_positive_order_ids),
);

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
GROUP BY item_id

This model could be extended with a user-defined macro variable to filter the query results based on item_size like this:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  audits (assert_positive_order_ids),
); -- NOTE: semi-colon at end of MODEL statement

@DEF(size, 1); -- NOTE: semi-colon at end of @DEF operator

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
WHERE
  item_size > @size -- Reference to macro variable `@size` defined above with `@DEF()`
GROUP BY item_id

This example defines the macro variable size with @DEF(size, 1). When the model is run, SQLMesh will substitute in the number 1 where @size appears in the WHERE clause.

Macro functions

In addition to inline user-defined variables, SQLMesh also supports inline macro functions. These functions can be used to express more readable and reusable logic than is possible with variables alone. Lets look at an example:

MODEL(...);

@DEF(
  rank_to_int,
  x -> case when left(x, 1) = 'A' then 1 when left(x, 1) = 'B' then 2 when left(x, 1) = 'C' then 3 end
);

SELECT
  id,
  cust_rank_1,
  cust_rank_2,
  cust_rank_3
  @rank_to_int(cust_rank_1) as cust_rank_1_int,
  @rank_to_int(cust_rank_2) as cust_rank_2_int,
  @rank_to_int(cust_rank_3) as cust_rank_3_int
FROM
  some.model

Multiple arguments can be expressed in a macro function as well:

@DEF(pythag, (x,y) -> sqrt(pow(x, 2) + pow(y, 2)));

SELECT
  sideA,
  sideB,
  @pythag(sideA, sideB) AS sideC
FROM
  some.triangle

@DEF(nrr, (starting_mrr, expansion_mrr, churned_mrr) -> (starting_mrr + expansion_mrr - churned_mrr) / starting_mrr);

SELECT
  @nrr(fy21_mrr, fy21_expansions, fy21_churns) AS fy21_net_retention_rate,
  @nrr(fy22_mrr, fy22_expansions, fy22_churns) AS fy22_net_retention_rate,
  @nrr(fy23_mrr, fy23_expansions, fy23_churns) AS fy23_net_retention_rate,
FROM
  some.revenue

You can nest macro functions like so:

MODEL (
  name dummy.model,
  kind FULL
);

@DEF(area, r -> pi() * r * r);
@DEF(container_volume, (r, h) -> @area(@r) * h);

SELECT container_id, @container_volume((cont_di / 2), cont_hi) AS volume

Macro operators

SQLMesh's macro system has multiple operators that allow different forms of dynamic behavior in models.

@EACH

@EACH is used to transform a list of items by applying a function to each of them, analogous to a for loop.

Learn more about for loops and @EACH

Before diving into the @EACH operator, let's dissect a for loop to understand its components.

for loops have two primary parts: a collection of items and an action that should be taken for each item. For example, here is a for loop in Python:

for number in [4, 5, 6]:
    print(number)

This for loop prints each number present in the brackets:

4
5
6

The first line of the example sets up the loop, doing two things:

1. Telling Python that code inside the loop will refer to each item as number
2. Telling Python to step through the list of items in brackets

The second line tells Python what action should be taken for each item. In this case, it prints the item.

The loop executes one time for each item in the list, substituting in the item for the word number in the code. For example, the first time through the loop the code would execute as print(4) and the second time as print(5).

The SQLMesh @EACH operator is used to implement the equivalent of a for loop in SQLMesh macros.

@EACH gets its name from the fact that a loop performs the action "for each" item in the collection. It is fundamentally equivalent to the Python loop above, but you specify the two loop components differently.

@EACH takes two arguments: a list of items and a function definition.

@EACH([list of items], [function definition])

The function definition is specified inline. This example specifies the identity function, returning the input unmodified:

SELECT
  @EACH([4, 5, 6], number -> number)
FROM table

The loop is set up by the first argument: @EACH([4, 5, 6] tells SQLMesh to step through the list of items in brackets.

The second argument number -> number tells SQLMesh what action should be taken for each item using an "anonymous" function (aka "lambda" function). The left side of the arrow states what name the code on the right side will refer to each item as (like name in for [name] in [items] in a Python for loop).

The right side of the arrow specifies what should be done to each item in the list. number -> number tells @EACH that for each item number it should return that item (e.g., 1).

SQLMesh macros use their semantic understanding of SQL code to take automatic actions based on where in a SQL query macro variables are used. If @EACH is used in the SELECT clause of a SQL statement:

1. It prints the item
2. It knows fields are separated by commas in SELECT, so it automatically separates the printed items with commas

Because of the automatic print and comma-separation, the anonymous function number -> number tells @EACH that for each item number it should print the item and separate the items with commas. Therefore, the complete output from the example is:

SELECT
  4,
  5,
  6
FROM table

This basic example is too simple to be useful. Many uses of @EACH will involve using the values as one or both of a literal value and an identifier.

For example, a column favorite_number in our data might contain values 4, 5, and 6, and we want to unpack that column into three indicator (i.e., binary, dummy, one-hot encoded) columns. We could write that by hand as:

SELECT
  CASE WHEN favorite_number = 4 THEN 1 ELSE 0 END as favorite_4,
  CASE WHEN favorite_number = 5 THEN 1 ELSE 0 END as favorite_5,
  CASE WHEN favorite_number = 6 THEN 1 ELSE 0 END as favorite_6
FROM table

In that SQL query each number is being used in two distinct ways. For example, 4 is being used:

1. As a literal numeric value in favorite_number = 4
2. As part of a column name in favorite_4

We describe each of these uses separately.

For the literal numeric value, @EACH substitutes in the exact value that is passed in the brackets, including quotes. For example, consider this query similar to the CASE WHEN example above:

SELECT
  @EACH([4,5,6], x -> CASE WHEN favorite_number = x THEN 1 ELSE 0 END as column)
FROM table

It renders to this SQL:

SELECT
  CASE WHEN favorite_number = 4 THEN 1 ELSE 0 END AS column,
  CASE WHEN favorite_number = 5 THEN 1 ELSE 0 END AS column,
  CASE WHEN favorite_number = 6 THEN 1 ELSE 0 END AS column
FROM table

Note that the number 4, 5, and 6 are unquoted in both the input @EACH array in brackets and the resulting SQL query.

We can instead quote them in the input @EACH array:

SELECT
  @EACH(['4','5','6'], x -> CASE WHEN favorite_number = x THEN 1 ELSE 0 END as column)
FROM table

And they will be quoted in the resulting SQL query:

SELECT
  CASE WHEN favorite_number = '4' THEN 1 ELSE 0 END AS column,
  CASE WHEN favorite_number = '5' THEN 1 ELSE 0 END AS column,
  CASE WHEN favorite_number = '6' THEN 1 ELSE 0 END AS column
FROM table

We can place the array values at the end of a column name by using the SQLMesh macro operator @ inside the @EACH function definition:

SELECT
  @EACH(['4','5','6'], x -> CASE WHEN favorite_number = x THEN 1 ELSE 0 END as column_@x)
FROM table

This query will render to:

SELECT
  CASE WHEN favorite_number = '4' THEN 1 ELSE 0 END AS column_4,
  CASE WHEN favorite_number = '5' THEN 1 ELSE 0 END AS column_5,
  CASE WHEN favorite_number = '6' THEN 1 ELSE 0 END AS column_6
FROM table

This syntax works regardless of whether the array values are quoted or not.

NOTE: SQLMesh macros support placing macro values at the end of a column name simply using column_@x. However if you wish to substitute the variable anywhere else in the identifier, you need to use the more explicit substitution syntax @{}. This avoids ambiguity. These are valid uses: @{x}_column or my_@{x}_column.

@IF

SQLMesh's @IF macro allows components of a SQL query to change based on the result of a logical condition.

It includes three elements:

1. A logical condition that evaluates to TRUE or FALSE
2. A value to return if the condition is TRUE
3. A value to return if the condition is FALSE [optional]

These elements are specified as:

@IF([logical condition], [value if TRUE], [value if FALSE])

The value to return if the condition is FALSE is optional - if it is not provided and the condition is FALSE, then the macro has no effect on the resulting query.

The logical condition should be written in SQL and is evaluated with SQLGlot's SQL executor. It supports the following operators:

• Equality: = for equals, != or <> for not equals
• Comparison: <, >, <=, >=,
• Between: [number] BETWEEN [low number] AND [high number]
• Membership: [item] IN ([comma-separated list of items])

For example, the following simple conditions are all valid SQL and evaluate to TRUE:

• 'a' = 'a'
• 'a' != 'b'
• 0 < 1
• 1 >= 1
• 2 BETWEEN 1 AND 3
• 'a' IN ('a', 'b')

@IF can be used to modify any part of a SQL query. For example, this query conditionally includes sensitive_col in the query results:

SELECT
  col1,
  @IF(1 > 0, sensitive_col)
FROM table

Because 1 > 0 evaluates to TRUE, the query is rendered as:

SELECT
  col1,
  sensitive_col
FROM table

Note that @IF(1 > 0, sensitive_col) does not include the third argument specifying a value if FALSE. Had the condition evaluated to FALSE, @IF would return nothing and only col1 would be selected.

Alternatively, we could specify that nonsensitive_col be returned if the condition evaluates to FALSE:

SELECT
  col1,
  @IF(1 > 2, sensitive_col, nonsensitive_col)
FROM table

Because 1 > 2 evaluates to FALSE, the query is rendered as:

SELECT
  col1,
  nonsensitive_col
FROM table

Pre/post-statements

@IF may be used to conditionally execute pre/post-statements:

@IF([logical condition], [statement to execute if TRUE]);

The @IF statement itself must end with a semi-colon, but the inner statement argument must not.

This example conditionally executes a pre/post-statement depending on the model's runtime stage, accessed via the pre-defined macro variable @runtime_stage. The @IF post-statement will only be executed at model evaluation time:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  grain item_id,
  audits (assert_positive_order_ids),
);

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
GROUP BY item_id
ORDER BY item_id;

@IF(
  @runtime_stage = 'evaluating',
  ALTER TABLE sqlmesh_example.full_model ALTER item_id TYPE VARCHAR
);

NOTE: alternatively, we could alter a column's type if the @runtime_stage = 'creating', but that would only be useful if the model is incremental and the alteration would persist. FULL models are rebuilt on each evaluation, so changes made at their creation stage will be overwritten each time the model is evaluated.

@EVAL

@EVAL evaluates its arguments with SQLGlot's SQL executor.

It allows you to execute mathematical or other calculations in SQL code. It behaves similarly to the first argument of the @IF operator, but it is not limited to logical conditions.

For example, consider a query adding 5 to a macro variable:

MODEL (
  ...
);

@DEF(x, 1);

SELECT
  @EVAL(5 + @x) as my_six
FROM table

After macro variable substitution, this would render as @EVAL(5 + 1) and be evaluated to 6, resulting in the final rendered query:

SELECT
  6 as my_six
FROM table

@FILTER

@FILTER is used to subset an input array of items to only those meeting the logical condition specified in the anonymous function. Its output can be consumed by other macro operators such as @EACH or @REDUCE.

The user-specified anonymous function must evaluate to TRUE or FALSE. @FILTER applies the function to each item in the array, only including the item in the output array if it meets the condition.

The anonymous function should be written in SQL and is evaluated with SQLGlot's SQL executor. It supports standard SQL equality and comparison operators - see @IF above for more information about supported operators.

For example, consider this @FILTER call:

@FILTER([1,2,3], x -> x > 1)

It applies the condition x > 1 to each item in the input array [1,2,3] and returns [2,3].

@REDUCE

@REDUCE is used to combine the items in an array.

The anonymous function specifies how the items in the input array should be combined. In contrast to @EACH and @FILTER, the anonymous function takes two arguments whose values are named in parentheses.

For example, an anonymous function for @EACH might be specified x -> x + 1. The x to the left of the arrow tells SQLMesh that the array items will be referred to as x in the code to the right of the arrow.

Because the @REDUCE anonymous function takes two arguments, the text to the left of the arrow must contain two comma-separated names in parentheses. For example, (x, y) -> x + y tells SQLMesh that items will be referred to as x and y in the code to the right of the arrow.

Even though the anonymous function takes only two arguments, the input array can contain as many items as necessary.

Consider the anonymous function (x, y) -> x + y. Conceptually, only the y argument corresponds to items in the array; the x argument is a temporary value created when the function is evaluated.

For the call @REDUCE([1,2,3,4], (x, y) -> x + y), the anonymous function is applied to the array in the following steps:

1. Take the first two items in the array as x and y. Apply the function to them: 1 + 2 = 3.
2. Take the output of step (1) as x and the next item in the array 3 as y. Apply the function to them: 3 + 3 = 6.
3. Take the output of step (2) as x and the next item in the array 4 as y. Apply the function to them: 6 + 4 = 10.
4. No items remain. Return value from step (3): 10.

@REDUCE will almost always be used with another macro operator. For example, we might want to build a WHERE clause from multiple column names:

SELECT
  my_column
FROM
  table
WHERE
  col1 = 1 and col2 = 1 and col3 = 1

We can use @EACH to build each column's predicate (e.g., col1 = 1) and @REDUCE to combine them into a single statement:

SELECT
  my_column
FROM
  table
WHERE
  @REDUCE(
    @EACH([col1, col2, col3], x -> x = 1), -- Builds each individual predicate `col1 = 1`
    (x, y) -> x AND y -- Combines individual predicates with `AND`
  )

@STAR

@STAR is used to return a set of column selections in a query.

@STAR is named after SQL's star operator *, but it allows you to programmatically generate a set of column selections and aliases instead of just selecting all available columns. A query may use more than one @STAR and may also include explicit column selections.

@STAR uses SQLMesh's knowledge of each table's columns and data types to generate the appropriate column list.

If the column data types are known, the resulting query CASTs columns to their data type in the source table. Otherwise, the columns will be listed without any casting.

@STAR supports the following arguments, in this order:

• relation: The relation/table whose columns are being selected
• alias (optional): The alias of the relation (if it has one)
• exclude (optional): A list of columns to exclude
• prefix (optional): A string to use as a prefix for all selected column names
• suffix (optional): A string to use as a suffix for all selected column names
• quote_identifiers (optional): Whether to quote the resulting identifiers, defaults to true

NOTE: the exclude argument used to be named except_. The latter is still supported but we discourage its use because it will be deprecated in the future.

Like all SQLMesh macro functions, omitting an argument when calling @STAR requires passing subsequent arguments with their name and the special := keyword operator. For example, we might omit the alias argument with @STAR(foo, exclude := [c]). Learn more about macro function arguments below.

As a @STAR example, consider the following query:

SELECT
  @STAR(foo, bar, [c], 'baz_', '_qux')
FROM foo AS bar

The arguments to @STAR are:

1. The name of the table foo (from the query's FROM foo)
2. The table alias bar (from the query's AS bar)
3. A list of columns to exclude from the selection, containing one column c
4. A string baz_ to use as a prefix for all column names
5. A string _qux to use as a suffix for all column names

foo is a table that contains four columns: a (TEXT), b (TEXT), c (TEXT) and d (INT). After macro expansion, if the column types are known the query would be rendered as:

SELECT
  CAST("bar"."a" AS TEXT) AS "baz_a_qux",
  CAST("bar"."b" AS TEXT) AS "baz_b_qux",
  CAST("bar"."d" AS INT) AS "baz_d_qux"
FROM foo AS bar

Note these aspects of the rendered query:

• Each column is CAST to its data type in the table foo (e.g., a to TEXT)
• Each column selection uses the alias bar (e.g., "bar"."a")
• Column c is not present because it was passed to @STAR's exclude argument
• Each column alias is prefixed with baz_ and suffixed with _qux (e.g., "baz_a_qux")

Now consider a more complex example that provides different prefixes to a and b than to d and includes an explicit column my_column:

SELECT
  @STAR(foo, bar, exclude := [c, d], 'ab_pre_'),
  @STAR(foo, bar, exclude := [a, b, c], 'd_pre_'),
  my_column
FROM foo AS bar

As before, foo is a table that contains four columns: a (TEXT), b (TEXT), c (TEXT) and d (INT). After macro expansion, the query would be rendered as:

SELECT
  CAST("bar"."a" AS TEXT) AS "ab_pre_a",
  CAST("bar"."b" AS TEXT) AS "ab_pre_b",
  CAST("bar"."d" AS INT) AS "d_pre_d",
  my_column
FROM foo AS bar

Note these aspects of the rendered query:

• Columns a and b have the prefix "ab_pre_" , while column d has the prefix "d_pre_"
• Column c is not present because it was passed to the exclude argument in both @STAR calls
• my_column is present in the query

@GENERATE_SURROGATE_KEY

@GENERATE_SURROGATE_KEY generates a surrogate key from a set of columns. The surrogate key is a sequence of alphanumeric digits returned by the MD5 hash function on the concatenated column values.

The surrogate key is created by: 1. CASTing each column's value to TEXT (or the SQL engine's equivalent type) 2. Replacing NULL values with the text '_sqlmesh_surrogate_key_null_' for each column 3. Concatenating the column values after steps (1) and (2) 4. Applying the MD5() hash function to the concatenated value returned by step (3)

For example, the following query:

SELECT
  @GENERATE_SURROGATE_KEY(a, b, c)
FROM foo

would be rendered as:

SELECT
  MD5(
    CONCAT(
      COALESCE(CAST(a AS TEXT), '_sqlmesh_surrogate_key_null_'),
      '|',
      COALESCE(CAST(b AS TEXT), '_sqlmesh_surrogate_key_null_'),
      '|',
      COALESCE(CAST(c AS TEXT), '_sqlmesh_surrogate_key_null_')
    )
  )
FROM foo

@SAFE_ADD

@SAFE_ADD adds two or more operands, substituting NULLs with 0s. It returns NULL if all operands are NULL.

For example, the following query:

SELECT
  @SAFE_ADD(a, b, c)
FROM foo

would be rendered as:

SELECT
  CASE WHEN a IS NULL AND b IS NULL AND c IS NULL THEN NULL ELSE COALESCE(a, 0) + COALESCE(b, 0) + COALESCE(c, 0) END
FROM foo

@SAFE_SUB

@SAFE_SUB subtracts two or more operands, substituting NULLs with 0s. It returns NULL if all operands are NULL.

For example, the following query:

SELECT
  @SAFE_SUB(a, b, c)
FROM foo

would be rendered as:

SELECT
  CASE WHEN a IS NULL AND b IS NULL AND c IS NULL THEN NULL ELSE COALESCE(a, 0) - COALESCE(b, 0) - COALESCE(c, 0) END
FROM foo

@SAFE_DIV

@SAFE_DIV divides two numbers, returning NULL if the denominator is 0.

For example, the following query:

SELECT
  @SAFE_DIV(a, b)
FROM foo

would be rendered as:

SELECT
  a / NULLIF(b, 0)
FROM foo

@UNION

@UNION returns a UNION query that selects all columns with matching names and data types from the tables.

Its first argument is the UNION "type", 'DISTINCT' (removing duplicated rows) or 'ALL' (returning all rows). Subsequent arguments are the tables to be combined.

Let's assume that:

• foo is a table that contains three columns: a (INT), b (TEXT), c (TEXT)
• bar is a table that contains three columns: a (INT), b (INT), c (TEXT)

Then, the following expression:

@UNION('distinct', foo, bar)

would be rendered as:

SELECT
  CAST(a AS INT) AS a,
  CAST(c AS TEXT) AS c
FROM foo
UNION
SELECT
  CAST(a AS INT) AS a,
  CAST(c AS TEXT) AS c
FROM bar

@HAVERSINE_DISTANCE

@HAVERSINE_DISTANCE returns the haversine distance between two geographic points.

It supports the following arguments, in this order:

• lat1: Latitude of the first point
• lon1: Longitude of the first point
• lat2: Latitude of the second point
• lon2: Longitude of the second point
• unit (optional): The measurement unit, currently only 'mi' (miles, default) and 'km' (kilometers) are supported

SQLMesh macro operators do not accept named arguments. For example, @HAVERSINE_DISTANCE(lat1=lat_column) will error.

For example, the following query:

SELECT
  @HAVERSINE_DISTANCE(driver_y, driver_x, passenger_y, passenger_x, 'mi') AS dist
FROM rides

would be rendered as:

SELECT
  7922 * ASIN(SQRT((POWER(SIN(RADIANS((passenger_y - driver_y) / 2)), 2)) + (COS(RADIANS(driver_y)) * COS(RADIANS(passenger_y)) * POWER(SIN(RADIANS((passenger_x - driver_x) / 2)), 2)))) * 1.0 AS dist
FROM rides

@PIVOT

@PIVOT returns a set of columns as a result of pivoting an input column on the specified values. This operation is sometimes described a pivoting from a "long" format (multiple values in a single column) to a "wide" format (one value in each of multiple columns).

It supports the following arguments, in this order:

• column: The column to pivot
• values: The values to use for pivoting (one column is created for each value in values)
• alias (optional): Whether to create aliases for the resulting columns, defaults to true
• agg (optional): The aggregation function to use, defaults to SUM
• cmp (optional): The comparison operator to use for comparing the column values, defaults to =
• prefix (optional): A prefix to use for all aliases
• suffix (optional): A suffix to use for all aliases
• then_value (optional): The value to be used if the comparison succeeds, defaults to 1
• else_value (optional): The value to be used if the comparison fails, defaults to 0
• quote (optional): Whether to quote the resulting aliases, defaults to true
• distinct (optional): Whether to apply a DISTINCT clause for the aggregation function, defaults to false

Like all SQLMesh macro functions, omitting an argument when calling @PIVOT requires passing subsequent arguments with their name and the special := keyword operator. For example, we might omit the agg argument with @PIVOT(status, ['cancelled', 'completed'], cmp := '<'). Learn more about macro function arguments below.

For example, the following query:

SELECT
  date_day,
  @PIVOT(status, ['cancelled', 'completed'])
FROM rides
GROUP BY 1

would be rendered as:

SELECT
  date_day,
  SUM(CASE WHEN status = 'cancelled' THEN 1 ELSE 0 END) AS "'cancelled'",
  SUM(CASE WHEN status = 'completed' THEN 1 ELSE 0 END) AS "'completed'"
FROM rides
GROUP BY 1

@DEDUPLICATE

@DEDUPLICATE is used to deduplicate rows in a table based on the specified partition and order columns with a window function.

It supports the following arguments, in this order:

• relation: The table or CTE name to deduplicate
• partition_by: column names, or expressions to use to identify a window of rows out of which to select one as the deduplicated row
• order_by: A list of strings representing the ORDER BY clause, optional - you can add nulls ordering like this: [' desc nulls last']

For example, the following query:

with raw_data as (
@deduplicate(my_table, [id, cast(event_date as date)], ['event_date DESC', 'status ASC'])
)

select * from raw_data

would be rendered as:

WITH "raw_data" AS (
  SELECT
    *
  FROM "my_table" AS "my_table"
  QUALIFY
    ROW_NUMBER() OVER (PARTITION BY "id", CAST("event_date" AS DATE) ORDER BY "event_date" DESC, "status" ASC) = 1
)
SELECT
  *
FROM "raw_data" AS "raw_data"

@DATE_SPINE

@DATE_SPINE returns the SQL required to build a date spine. The spine will include the start_date (if it is aligned to the datepart), AND it will include the end_date. This is different from the date_spine macro in dbt-utils which will NOT include the end_date. It's typically used to join in unique, hard-coded, date ranges to with other tables/views, so people don't have to constantly adjust date ranges in where clauses across many SQL models.

It supports the following arguments, in this order:

• datepart: The datepart to use for the date spine - day, week, month, quarter, year
• start_date: The start date for the date spine in format YYYY-MM-DD
• end_date: The end date for the date spine in format YYYY-MM-DD

For example, the following query:

WITH discount_promotion_dates AS (
  @date_spine('day', '2024-01-01', '2024-01-16')
)

SELECT * FROM discount_promotion_dates

would be rendered as:

WITH "discount_promotion_dates" AS (
  SELECT
    "_exploded"."date_day" AS "date_day"
  FROM UNNEST(CAST(GENERATE_SERIES(CAST('2024-01-01' AS DATE), CAST('2024-01-16' AS DATE), INTERVAL '1' DAY) AS
DATE[])) AS "_exploded"("date_day")
)
SELECT
  "discount_promotion_dates"."date_day" AS "date_day"
FROM "discount_promotion_dates" AS "discount_promotion_dates"

Note: This is DuckDB SQL and other dialects will be transpiled accordingly. - Recursive CTEs (common table expressions) will be used for Redshift / MySQL / MSSQL. - For MSSQL in particular, there's a recursion limit of approximately 100. If this becomes a problem, you can add an OPTION (MAXRECURSION 0) clause after the date spine macro logic to remove the limit. This applies for long date ranges.

@AND

@AND combines a sequence of operands using the AND operator, filtering out any NULL expressions.

For example, the following expression:

@AND(TRUE, NULL)

would be rendered as:

TRUE

@OR

@OR combines a sequence of operands using the OR operator, filtering out any NULL expressions.

For example, the following expression:

@OR(TRUE, NULL)

would be rendered as:

TRUE

SQL clause operators

SQLMesh's macro system has six operators that correspond to different clauses in SQL syntax. They are:

• @WITH: common table expression WITH clause
• @JOIN: table JOIN clause(s)
• @WHERE: filtering WHERE clause
• @GROUP_BY: grouping GROUP BY clause
• @HAVING: group by filtering HAVING clause
• @ORDER_BY: ordering ORDER BY clause
• @LIMIT: limiting LIMIT clause

Each of these operators is used to dynamically add the code for its corresponding clause to a model's SQL query.

How SQL clause operators work

The SQL clause operators take a single argument that determines whether the clause is generated.

If the argument is TRUE the clause code is generated, if FALSE the code is not. The argument should be written in SQL and its value is evaluated with SQLGlot's SQL engine.

Each SQL clause operator may only be used once in a query, but any common table expressions or subqueries may contain their own single use of the operator as well.

As an example of SQL clause operators, let's revisit the example model from the User-defined Variables section above.

As written, the model will always include the WHERE clause. We could make its presence dynamic by using the @WHERE macro operator:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  audits (assert_positive_order_ids),
);

@DEF(size, 1);

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
@WHERE(TRUE) item_id > @size
GROUP BY item_id

The @WHERE argument is set to TRUE, so the WHERE code is included in the rendered model:

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
WHERE item_id > 1
GROUP BY item_id

If the @WHERE argument were instead set to FALSE the WHERE clause would be omitted from the query.

These operators aren't too useful if the argument's value is hard-coded. Instead, the argument can consist of code executable by the SQLGlot SQL executor.

For example, the WHERE clause will be included in this query because 1 less than 2:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  audits (assert_positive_order_ids),
);

@DEF(size, 1);

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
@WHERE(1 < 2) item_id > @size
GROUP BY item_id

The operator's argument code can include macro variables.

In this example, the two numbers being compared are defined as macro variables instead of being hard-coded:

MODEL (
  name sqlmesh_example.full_model,
  kind FULL,
  cron '@daily',
  audits (assert_positive_order_ids),
);

@DEF(left_number, 1);
@DEF(right_number, 2);
@DEF(size, 1);

SELECT
  item_id,
  count(distinct id) AS num_orders,
FROM
  sqlmesh_example.incremental_model
@WHERE(@left_number < @right_number) item_id > @size
GROUP BY item_id

The argument to @WHERE will be "1 < 2" as in the previous hard-coded example after the macro variables left_number and right_number are substituted in.

SQL clause operator examples

This section provides brief examples of each SQL clause operator's usage.

The examples use variants of this simple select statement:

SELECT *
FROM all_cities

@WITH operator

The @WITH operator is used to create common table expressions, or "CTEs."

CTEs are typically used in place of derived tables (subqueries in the FROM clause) to make SQL code easier to read. Less commonly, recursive CTEs support analysis of hierarchical data with SQL.

@WITH(True) all_cities as (select * from city)
select *
FROM all_cities

renders to

WITH all_cities as (select * from city)
select *
FROM all_cities

@JOIN operator

The @JOIN operator specifies joins between tables or other SQL objects; it supports different join types (e.g., INNER, OUTER, CROSS, etc.).

select *
FROM all_cities
LEFT OUTER @JOIN(True) country
  ON city.country = country.name

renders to

select *
FROM all_cities
LEFT OUTER JOIN country
  ON city.country = country.name

The @JOIN operator recognizes that LEFT OUTER is a component of the JOIN specification and will omit it if the @JOIN argument evaluates to False.

@WHERE operator

The @WHERE operator adds a filtering WHERE clause(s) to the query when its argument evaluates to True.

SELECT *
FROM all_cities
@WHERE(True) city_name = 'Toronto'

renders to

SELECT *
FROM all_cities
WHERE city_name = 'Toronto'

@GROUP_BY operator

SELECT *
FROM all_cities
@GROUP_BY(True) city_id

renders to

SELECT *
FROM all_cities
GROUP BY city_id

@HAVING operator

SELECT
count(city_pop) as population
FROM all_cities
GROUP BY city_id
@HAVING(True) population > 1000

renders to

SELECT
count(city_pop) as population
FROM all_cities
GROUP BY city_id
HAVING population > 1000

@ORDER_BY operator

SELECT *
FROM all_cities
@ORDER_BY(True) city_pop

renders to

SELECT *
FROM all_cities
ORDER BY city_pop

@LIMIT operator

SELECT *
FROM all_cities
@LIMIT(True) 10

renders to

SELECT *
FROM all_cities
LIMIT 10

User-defined macro functions

User-defined macro functions allow the same macro code to be used in multiple models.

SQLMesh supports user-defined macro functions written in two languages - SQL and Python:

• SQL macro functions must use the Jinja templating system.
• Python macro functions use the SQLGlot library to allow more complex operations than macro variables and operators provide alone.

Python macro functions

Setup

Python macro functions should be placed in .py files in the SQLMesh project's macros directory. Multiple functions can be defined in one .py file, or they can be distributed across multiple files.

An empty __init__.py file must be present in the SQLMesh project's macros directory. It will be created automatically when the project scaffold is created with sqlmesh init.

Each .py file containing a macro definition must import SQLMesh's macro decorator with from sqlmesh import macro.

Python macros are defined as regular python functions adorned with the SQLMesh @macro() decorator. The first argument to the function must be evaluator, which provides the macro evaluation context in which the macro function will run.

Inputs and outputs

Python macros parse all arguments passed to the macro call with SQLGlot before they are used in the function body. Therefore, unless argument type annotations are provided in the function definition, the macro function code must process SQLGlot expressions and may need to extract the expression's attributes/contents for use.

Python macro functions may return values of either string or SQLGlot expression types. SQLMesh will automatically parse returned strings into a SQLGlot expression after the function is executed so they can be incorporated into the model query's semantic representation.

Macro functions may return a list of strings or expressions that all play the same role in the query (e.g., specifying column definitions). For example, a list containing multiple CASE WHEN statements would be incorporated into the query properly, but a list containing both CASE WHEN statements and a WHERE clause would not.

Macro function basics

This example demonstrates the core requirements for defining a python macro - it takes no user-supplied arguments and returns the string text.

from sqlmesh import macro

@macro() # Note parentheses at end of `@macro()` decorator
def print_text(evaluator):
  return 'text'

We could use this in a SQLMesh SQL model like this:

SELECT
  @print_text() as my_text
FROM table

After processing, it will render to this:

SELECT
  text as my_text
FROM table

Note that the python function returned a string 'text', but the rendered query uses text as a column name. That is due to the function's returned text being parsed as SQL code by SQLGlot and integrated into the query's semantic representation.

The rendered query will treat text as a string if we double-quote the single-quoted value in the function definition as "'text'":

from sqlmesh import macro

@macro()
def print_text(evaluator):
    return "'text'"

When run in the same model query as before, this will render to:

SELECT
  'text' as my_text
FROM table

Argument data types

Most macro functions provide arguments so users can supply custom values when the function is called. The data type of the argument plays a key role in how the macro code processes its value, and providing type annotations in the macro definition ensures that the macro code receives the data type it expects. This section provides a brief description of SQLMesh macro type annotation - find additional information below.

As mentioned above, argument values passed to the macro call are parsed by SQLGlot before they become available to the function code. If an argument does not have a type annotation in the macro function definition, its value will always be a SQLGlot expression in the function body. Therefore, the macro function code must operate directly on the expression (and may need to extract information from it before usage).

If an argument does have a type annotation in the macro function definition, the value passed to the macro call will be coerced to that type after parsing by SQLGlot and before the values are used in the function body. Essentially, SQLMesh will extract the relevant information of the annotated data type from the expression for you (if possible).

For example, this macro function determines whether an argument's value is any of the integers 1, 2, or 3:

from sqlmesh import macro

@macro()
def arg_in_123(evaluator, my_arg):
    return my_arg in [1,2,3]

When this macro is called, it will return FALSE even if an integer was passed in the call. Consider this macro call:

SELECT
  @arg_in_123(1)

It returns SELECT FALSE because:

1. The passed value 1 is parsed by SQLGlot into a SQLGlot expression before the function code executes and
2. There is no matching SQLGlot expression in [1,2,3]

However, the macro will treat the argument like a normal Python function does if we annotate my_arg with the integer int type in the function definition:

from sqlmesh import macro

@macro()
def arg_in_123(evaluator, my_arg: int): # Type annotation `my_arg: int`
    return my_arg in [1,2,3]

Now the macro call will return SELECT TRUE because the value is coerced to a Python integer before the function code executes and 1 is in [1,2,3].

If an argument has a default value, the value is not parsed by SQLGlot before the function code executes. Therefore, take care to ensure that the default's data type matches that of a user-supplied argument by adding a type annotation, making the default value a SQLGlot expression, or making the default value None.

Positional and keyword arguments

In a macro call, the arguments may be provided by position if none are skipped.

For example, consider the add_args() function - it has three arguments with default values provided in the function definition:

from sqlmesh import macro

@macro()
def add_args(
    evaluator,
    argument_1: int = 1,
    argument_2: int = 2,
    argument_3: int = 3
):
    return argument_1 + argument_2 + argument_3

An @add_args call providing values for all arguments accepts positional arguments like this: @add_args(5, 6, 7) (which returns 5 + 6 + 7 = 18). A call omitting and using the default value for the the final argument_3 can also use positional arguments: @add_args(5, 6) (which returns 5 + 6 + 3 = 14).

However, skipping an argument requires specifying the names of subsequent arguments (i.e., using "keyword arguments"). For example, skipping the second argument above by just omitting it - @add_args(5, , 7) - results in an error.

Unlike Python, SQLMesh keyword arguments must use the special operator :=. To skip and use the default value for the second argument above, the call must name the third argument: @add_args(5, argument_3 := 8) (which returns 5 + 2 + 8 = 15).

Variable-length arguments

The add_args() macro defined in the previous section accepts only three arguments and requires that all three have a value. This greatly limits the macro's flexibility because users may want to add any number of values together.

The macro can be improved by allowing users to provide any number of arguments at call time. We use Python's "variable-length arguments" to accomplish this:

from sqlmesh import macro

@macro()
def add_args(evaluator, *args: int): # Variable-length arguments of integer type `*args: int`
    return sum(args)

This macro can be called with one or more arguments. For example:

• @add_args(1) returns 1
• @add_args(1, 2) returns 3
• @add_args(1, 2, 3) returns 6

Returning more than one value

Macro functions are a convenient way to tidy model code by creating multiple outputs from one function call. Python macro functions do this by returning a list of strings or SQLGlot expressions.

For example, we might want to create indicator variables from the values in a string column. We can do that by passing in the name of column and a list of values for which it should create indicators, which we then interpolate into CASE WHEN statements.

Because SQLMesh parses the input objects, they become SQLGLot expressions in the function body. Therefore, the function code cannot treat the input list as a regular Python list.

Two things will happen to the input Python list before the function code is executed:

1. Each of its entries will be parsed by SQLGlot. Different inputs are parsed into different SQLGlot expressions:

   • Numbers are parsed into Literal expressions
   • Quoted strings are parsed into Literal expressions
   • Unquoted strings are parsed into Column expressions

2. The parsed entries will be contained in a SQLGlot Array expression, the SQL entity analogous to a Python list

Because the input Array expression named values is not a Python list, we cannot iterate over it directly - instead, we iterate over its expressions attribute with values.expressions:

from sqlmesh import macro

@macro()
def make_indicators(evaluator, string_column, values):
    cases = []

    for value in values.expressions: # Iterate over `values.expressions`
        cases.append(f"CASE WHEN {string_column} = '{value}' THEN '{value}' ELSE NULL END AS {string_column}_{value}")

    return cases

We call this function in a model query to create CASE WHEN statements for the vehicle column values truck and bus like this:

SELECT
  @make_indicators(vehicle, [truck, bus])
FROM table

Which renders to:

SELECT
  CASE WHEN vehicle = 'truck' THEN 'truck' ELSE NULL END AS vehicle_truck,
  CASE WHEN vehicle = 'bus' THEN 'bus' ELSE NULL END AS vehicle_bus,
FROM table

Note that in the call @make_indicators(vehicle, [truck, bus]) none of the three values is quoted.

Because they are unquoted, SQLGlot will parse them all as Column expressions. In the places we used single quotes when building the string ('{value}'), they will be single-quoted in the output. In the places we did not quote them ({string_column} = and {string_column}_{value}), they will not.

Accessing predefined and local variable values

Pre-defined variables and user-defined local variables can be accessed within the macro's body via the evaluator.locals attribute.

The first argument to every macro function, the macro evaluation context evaluator, contains macro variable values in its locals attribute. evaluator.locals is a dictionary whose key:value pairs are macro variables names and the associated values.

For example, a function could access the predefined execution_epoch variable containing the epoch timestamp of when the execution started.

from sqlmesh import macro

@macro()
def get_execution_epoch(evaluator):
    return evaluator.locals['execution_epoch']

The function would return the execution_epoch value when called in a model query:

SELECT
  @get_execution_epoch() as execution_epoch
FROM table

The same approach works for user-defined local macro variables, where the key "execution_epoch" would be replaced with the name of the user-defined variable to be accessed.

One downside of that approach to accessing user-defined local variables is that the name of the variable is hard-coded into the function. A more flexible approach is to pass the name of the local macro variable as a function argument:

from sqlmesh import macro

@macro()
def get_macro_var(evaluator, macro_var):
    return evaluator.locals[macro_var]

We could define a local macro variable my_macro_var with a value of 1 and pass it to the get_macro_var function like this:

MODEL (...);

@DEF(my_macro_var, 1); -- Define local macro variable 'my_macro_var'

SELECT
  @get_macro_var('my_macro_var') as macro_var_value -- Access my_macro_var value from Python macro function
FROM table

The model query would render to:

SELECT
  1 as macro_var_value
FROM table

Accessing global variable values

User-defined global variables can be accessed within the macro's body using the evaluator.var method.

If a global variable is not defined, the method will return a Python None value. You may provide a different default value as the method's second argument.

For example:

from sqlmesh.core.macros import macro

@macro()
def some_macro(evaluator):
    var_value = evaluator.var("<var_name>") # Default value is `None`
    another_var_value = evaluator.var("<another_var_name>", "default_value") # Default value is `"default_value"`
    ...

Accessing model schemas

Model schemas can be accessed within a Python macro function through its evaluation context's column_to_types() method, if the column types can be statically determined. For instance, a schema of an external model can be accessed only after the sqlmesh create_external_models command has been executed.

This macro function renames the columns of an upstream model by adding a prefix to them:

from sqlglot import exp
from sqlmesh.core.macros import macro

@macro()
def prefix_columns(evaluator, model_name, prefix: str):
    renamed_projections = []

    # The following converts `model_name`, which is a SQLGlot expression, into a lookup key,
    # assuming that it does not contain quotes. If it did, we would have to generate SQL for
    # each part of `model_name` separately and then concatenate these parts, because in that
    # case `model_name.sql()` would produce an invalid lookup key.
    model_name_sql = model_name.sql()

    for name in evaluator.columns_to_types(model_name_sql):
        new_name = prefix + name
        renamed_projections.append(exp.column(name).as_(new_name))

    return renamed_projections

This can then be used in a SQL model like this:

MODEL (
  name schema.child,
  kind FULL
);

SELECT
  @prefix_columns(schema.parent, 'stg_')
FROM
  schema.parent

Note that columns_to_types expects an unquoted model name, such as schema.parent. Since macro arguments without type annotations are SQLGlot expressions, the macro code must extract meaningful information from them. For instance, the lookup key in the above macro definition is extracted by generating the SQL code for model_name using the sql() method.

Accessing the schema of an upstream model can be useful for various reasons. For example:

• Renaming columns so that downstream consumers are not tightly coupled to external or source tables
• Selecting only a subset of columns that satisfy some criteria (e.g. columns whose names start with a specific prefix)
• Applying transformations to columns, such as masking PII or computing various statistics based on the column types

Thus, leveraging columns_to_types can also enable one to write code according to the DRY principle, as a single macro function can implement the transformations instead of creating a different macro for each model of interest.

Note: there may be models whose schema is not available when the project is being loaded, in which case a special placeholder column will be returned, aptly named: __schema_unavailable_at_load__. In some cases, the macro's implementation will need to account for this placeholder in order to avoid issues due to the schema being unavailable.

Accessing snapshots

After a SQLMesh project has been successfully loaded, its snapshots can be accessed in Python macro functions and Python models that generate SQL through the get_snapshot method of MacroEvaluator.

This enables the inspection of physical table names or the processed intervals for certain snapshots at runtime, as shown in the example below:

from sqlmesh.core.macros import macro

@macro()
def some_macro(evaluator):
    if evaluator.runtime_stage == "evaluating":
        # Check the intervals a snapshot has data for and alter the behavior of the macro accordingly
        intervals = evaluator.get_snapshot("some_model_name").intervals
        ...
    ...

Using SQLGlot expressions

SQLMesh automatically parses strings returned by Python macro functions into SQLGlot expressions so they can be incorporated into the model query's semantic representation. Functions can also return SQLGlot expressions directly.

For example, consider a macro function that uses the BETWEEN operator in the predicate of a WHERE clause. A function returning the predicate as a string might look like this, where the function arguments are substituted into a Python f-string:

from sqlmesh import macro, SQL

@macro()
def between_where(evaluator, column_name: SQL, low_val: SQL, high_val: SQL):
    return f"{column_name} BETWEEN {low_val} AND {high_val}"

The function could then be called in a query:

SELECT
  a
FROM table
WHERE @between_where(a, 1, 3)

And it would render to:

SELECT
  a
FROM table
WHERE a BETWEEN 1 and 3

Alternatively, the function could return a SQLGLot expression equivalent to that string by using SQLGlot's expression methods for building semantic representations:

from sqlmesh import macro

@macro()
def between_where(evaluator, column, low_val, high_val):
    return column.between(low_val, high_val)

The methods are available because the column argument is parsed as a SQLGlot Column expression when the macro function is executed.

Column expressions are sub-classes of the Condition class, so they have builder methods like between and like.

Macro pre/post-statements

Macro functions may be used to generate pre/post-statements in a model.

By default, when you first add the pre/post-statement macro functions to a model, SQLMesh will treat those models as directly modified and require a backfill in the next plan. SQLMesh will also treat edits to or removals of pre/post-statement macros as a breaking change.

If your macro does not affect the data returned by a model and you do not want its addition/editing/removal to trigger a backfill, you can specify in the macro definition that it only affects the model's metadata. SQLMesh will still detect changes and create new snapshots for a model when you add/edit/remove the macro, but it will not view the change as breaking and require a backfill.

Specify that a macro only affects a model's metadata by setting the @macro() decorator's metadata_only argument to True. For example:

from sqlmesh import macro

@macro(metadata_only=True)
def print_message(evaluator, message):
  print(message)

Typed Macros

Typed macros in SQLMesh bring the power of type hints from Python, enhancing readability, maintainability, and usability of your SQL macros. These macros enable developers to specify expected types for arguments, making the macros more intuitive and less error-prone.

Benefits of Typed Macros

1. Improved Readability: By specifying types, the intent of the macro is clearer to other developers or future you.
2. Reduced Boilerplate: No need for manual type conversion within the macro function, allowing you to focus on the core logic.
3. Enhanced Autocompletion: IDEs can provide better autocompletion and documentation based on the specified types.

Defining a Typed Macro

Typed macros in SQLMesh use Python's type hints. Here's a simple example of a typed macro that repeats a string a given number of times:

from sqlmesh import macro

@macro()
def repeat_string(evaluator, text: str, count: int):
    return text * count

This macro takes two arguments: text of type str and count of type int, and it returns a string.

Without type hints, the inputs are two SQLGlot exp.Literal objects you would need to manually convert to Python str and int types. With type hints, you can work with them as string and integer types directly.

Let's try to use the macro in a SQLMesh model:

SELECT
  @repeat_string('SQLMesh ', 3) as repeated_string
FROM some_table;

Unfortunately, this model generates an error when rendered:

Error: Invalid expression / Unexpected token. Line 1, Col: 23.
  SQLMesh SQLMesh SQLMesh

Why? The macro returned SQLMesh SQLMesh SQLMesh as expected, but that string is not valid SQL in the rendered query:

SELECT
  SQLMesh SQLMesh SQLMesh as repeated_string ### invalid SQL code
FROM some_table;

The problem is a mismatch between our macro's Python return type str and the type expected by the parsed SQL query.

Recall that SQLMesh macros work by modifying the query's semantic representation. In that representation, a SQLGlot string literal type is expected. SQLMesh will do its best to return the type expected by the query's semantic representation, but that is not possible in all scenarios.

Therefore, we must explicitly convert the output with SQLGlot's exp.Literal.string() method:

from sqlmesh import macro

@macro()
def repeat_string(evaluator, text: str, count: int):
    return exp.Literal.string(text * count)

Now the query will render with a valid single-quoted string literal:

SELECT
  'SQLMesh SQLMesh SQLMesh ' AS "repeated_string"
FROM "some_table" AS "some_table"

Typed macros coerce the inputs to a macro function, but the macro code is responsible for coercing the output to the type expected by the query's semantic representation.

Supported Types

SQLMesh supports common Python types for typed macros including:

• str -- This handles string literals and basic identifiers, but won't coerce anything more complicated.
• int
• float
• bool
• SQL -- When you want the SQL string representation of the argument that's passed in
• list[T] - where T is any supported type including sqlglot expressions
• tuple[T] - where T is any supported type including sqlglot expressions
• T1 | T2 | ... - where T1, T2, etc. are any supported types including sqlglot expressions

We also support SQLGlot expressions as type hints, allowing you to ensure inputs are coerced to the desired SQL AST node your intending on working with. Some useful examples include:

• exp.Table
• exp.Column
• exp.Literal
• exp.Identifier

While these might be obvious examples, you can effectively coerce an input into any SQLGlot expression type, which can be useful for more complex macros. When coercing to more complex types, you will almost certainly need to pass a string literal since expression to expression coercion is limited. When a string literal is passed to a macro that hints at a SQLGlot expression, the string will be parsed using SQLGlot and coerced to the correct type. Failure to coerce to the correct type will result in the original expression being passed to the macro and a warning being logged for the user to address as-needed.

@macro()
def stamped(evaluator, query: exp.Select) -> exp.Subquery:
    return query.select(exp.Literal.string(str(datetime.now())).as_("stamp")).subquery()

# Coercing to a complex node like `exp.Select` works as expected given a string literal input
# SELECT * FROM @stamped('SELECT a, b, c')

When coercion fails, there will always be a warning logged but we will not crash. We believe the macro system should be flexible by default, meaning the default behavior is preserved if we cannot coerce. Given that, the user can express whatever level of additional checks they want. For example, if you would like to raise an error when the coercion fails, you can use an assert statement. For example:

@macro()
def my_macro(evaluator, table: exp.Table) -> exp.Column:
    assert isinstance(table, exp.Table)
    table.set("catalog", "dev")
    return table

# Works
# SELECT * FROM @my_macro('some.table')
# SELECT * FROM @my_macro(some.table)

# Raises an error thanks to the users inclusion of the assert, otherwise would pass through the string literal and log a warning
# SELECT * FROM @my_macro('SELECT 1 + 1')

In using assert this way, you still get the benefits of reducing/removing the boilerplate needed to coerce types; but you also get guarantees about the type of the input. This is a useful pattern and is user-defined, so you can use it as you see fit. It ultimately allows you to keep the macro definition clean and focused on the core business logic.

Advanced Typed Macros

You can create more complex macros using advanced Python features like generics. For example, a macro that accepts a list of integers and returns their sum:

from typing import List
from sqlmesh import macro

@macro()
def sum_integers(evaluator, numbers: List[int]) -> int:
    return sum(numbers)

Usage in SQLMesh:

SELECT
  @sum_integers([1, 2, 3, 4, 5]) as total
FROM some_table;

Generics can be nested and are resolved recursively allowing for fairly robust type hinting.

See examples of the coercion function in action in the test suite here.

Conclusion

Typed macros in SQLMesh not only enhance the development experience by making macros more readable and easier to use but also contribute to more robust and maintainable code. By leveraging Python's type hinting system, developers can create powerful and intuitive macros for their SQL queries, further bridging the gap between SQL and Python.

Mixing macro systems

SQLMesh supports both SQLMesh and Jinja macro systems. We strongly recommend using only one system in a model - if both are present, they may fail or behave in unintuitive ways.