Skip to content

Overview

Models are comprised of metadata and queries that create tables and views, which can be used by other models or even outside of SQLMesh. They are defined in the models/ directory of your SQLMesh project and live in .sql files.

SQLMesh will automatically determine the relationships among and lineage of your models by parsing SQL, so you don't have to worry about manually configuring dependencies.

Example

The following is an example of a model defined in SQL. Note the following aspects:

  • Models can include descriptive information as comments, such as the first line.
  • The first non-comment statement of a model.sql file is the MODEL DDL.
  • The last non-comment statement should be a SELECT statement that defines the logic needed to create the table
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
-- Customer revenue computed and stored daily.
MODEL (
  name sushi.customer_total_revenue,
  owner toby,
  cron '@daily',
  grain customer_id
);

SELECT
  o.customer_id::TEXT,
  SUM(o.amount)::DOUBLE AS revenue
FROM sushi.orders AS o
GROUP BY o.customer_id;

Conventions

SQLMesh attempts to infer as much as possible about your pipelines through SQL alone to reduce the cognitive overhead of switching to another format such as YAML.

The SELECT expression of a model must follow certain conventions for SQLMesh to detect the necessary metadata to operate.

Unique column names

The final SELECT of a model's query must contain unique column names.

Explicit types

SQLMesh encourages explicit type casting in the final SELECT of a model's query. It is considered a best practice to prevent unexpected types in the schema of a model's table.

SQLMesh uses the postgres x::int syntax for casting; the casts are automatically transpiled to the appropriate format for the execution engine.

1
2
3
4
WITH cte AS (
  SELECT 1 AS foo -- don't need to cast here
)
SELECT foo::int -- need to cast here because it's in the final select statement

Inferrable names

The final SELECT of a model's query must have inferrable names or aliases.

Explicit aliases are recommended, but not required. The SQLMesh formatter will automatically add aliases to columns without them when the model SQL is rendered.

This example demonstrates non-inferrable, inferrable, and explicit aliases:

1
2
3
4
5
6
7
8
SELECT
  1, -- not inferrable
  x + 1, -- not inferrable
  SUM(x), -- not inferrable
  x, -- inferrable as x
  x::int, -- inferrable as x
  x + 1 AS x, -- explicitly x
  SUM(x) as x, -- explicitly x

Model properties

The MODEL DDL statement takes various properties, which are used for both metadata and controlling behavior.

name

  • name specifies the name of the model. This name represents the production view name that the model outputs, so it generally takes the form of "schema"."view_name". The name of a model must be unique in a SQLMesh project.

    When models are used in non-production environments, SQLMesh automatically prefixes the names. For example, consider a model named "sushi"."customers". In production its view is named "sushi"."customers", and in dev its view is named "sushi__dev"."customers".

    Name is required and must be unique.

kind

  • Kind specifies what kind a model is. A model's kind determines how it is computed and stored. The default kind is VIEW, which means a view is created and your query is run each time that view is accessed. See below for properties that apply to incremental model kinds.

dialect

owner

  • Owner specifies who the main point of contact is for the model. It is an important field for organizations that have many data collaborators.

stamp

  • An optional arbitrary string sequence used to create new model versions without making changes to any of the functional components of the definition.

start

  • Start is used to determine the earliest time needed to process the model. It can be an absolute date/time (2022-01-01), or a relative one (1 year ago).

cron

  • Cron is used to schedule your model to process or refresh at a certain interval. It uses croniter under the hood, so expressions such as @daily can be used.

interval_unit

  • Interval unit determines the granularity of data intervals for this model. By default the interval unit is automatically derived from the cron expression. Supported values are: year, month, day, hour, half_hour, quarter_hour, and five_minute.

tags

  • Tags are one or more labels used to organize your models.

grain

  • A model's grain is the column or combination of columns that uniquely identify a row in the results returned by the model's query. If the grain is set, SQLMesh tools like table_diff are simpler to run because they automatically use the model grain for parameters that would otherwise need to be specified manually.

grains

  • A model can define multiple grains if it has more than one unique key or combination of keys.

references

  • References are non-unique columns or combinations of columns that identify a join relationship to an entity. For example, a model could define a reference account_id, which would indicate that it can now automatically join to any model with an account_id grain. It cannot safely join to a table with an account_id reference because references are not unique and doing so would constitute a many-to-many join. Sometimes columns are named differently, in that case you can alias column names to a common entity name. For example guest_id AS account_id would allow a model with the column guest_id to join to a model with the grain account_id.

storage_format

  • Storage format is a property for engines such as Spark or Hive that support storage formats such as parquet and orc.

partitioned_by

  • Partitioned by is an optional property for engines such as Spark or BigQuery that support partitioning. Use this to specify a multi-column partition key or to modify a date column for partitioning. For example, in BigQuery you could partition by day by extracting the day component of a timestamp column event_ts with partitioned_by TIMESTAMP_TRUNC(event_ts, DAY).

clustered_by

  • Clustered by is an optional property for engines such as Bigquery that support clustering.

table_properties

  • A key-value of arbitrary table properties specific to the target engine. For example:
1
2
3
4
5
6
7
MODEL (
  ...,
  table_properties (
    partition_expiration_days = 7,
    require_partition_filter = true
  )
);

allow_partials

  • Indicates that this model can be executed for partial (incomplete) data intervals. By default, each model processes only complete intervals to prevent common mistakes caused by partial data. The size of the interval is determined by the model's interval_unit. Setting allow_partials to true overrides this behavior, indicating that the model may process a segment of input data that is missing some of the data points. Please note that setting this attribute to true results in the disregard of the cron attribute.

Incremental Model Properties

For models that are incremental, the following parameters can be specified in the kind's definition.

time_column

  • Time column is a required property for incremental models. It is used to determine which records to overwrite when doing an incremental insert. Time column can have an optional format string specified in the SQL dialect of the model.
  • Engines that support partitioning, such as Spark and BigQuery, use the time column as the model's partition key. Multi-column partitions or modifications to columns can be specified with the partitioned_by property.

lookback

  • Lookback is used for incremental models to capture late arriving data. This must be a positive integer and refers to the number of units that late arriving data is expected.

batch_size

  • Batch size is used to optimize backfilling incremental data. It determines the maximum number of intervals to run in a single job. For example, if a model specifies a cron of @hourly and a batch_size of 12, when backfilling 3 days of data, the scheduler will spawn 6 jobs. (3 days * 24 hours/day = 72 hour intervals to fill. 72 intervals / 12 intervals per job = 6 jobs.)

forward_only

  • Set this to true to indicate that all changes to this model should be forward-only.

disable_restatement

  • Set this to true to indicate that data restatement is disabled for this model.

Macros

Macros can be used for passing in parameterized arguments such as dates, as well as for making SQL less repetitive. By default, SQLMesh provides several predefined macro variables that can be used. Macros are used by prefixing with the @ symbol. For more information, refer to macros.

Statements

Models can have additional statements that run before and/or after the main query. They can be useful for loading things such as UDFs or cleaning up after a model query has run.

In general, pre-statements statements should only be used for preparing the main query. They should not be used for creating or altering tables, as this could lead to unpredictable behavior if multiple models are running simultaneously.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
MODEL (
...
);

-- Additional statements preparing for main query
ADD JAR s3://special_udf.jar;
CREATE TEMPORARY FUNCTION UDF AS 'my.jar.udf';

SELECT UDF(x)::int AS x
FROM y;

Additional statements can also be provided after the main query, in which case they will run after each evaluation of the SELECT query. Note that the model query must end with a semi-colon prior to the post-statements.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
MODEL (
...
);

...

SELECT UDF(x)::int AS x
FROM y;

-- Cleanup statements
DROP TABLE temp_table;

Time column

Models that are loaded incrementally require a time column to partition data.

A time column is a column in a model with an optional format string in the dialect of the model; for example, '%Y-%m-%d' for DuckDB or 'yyyy-mm-dd' for Snowflake. For more information, refer to time column.

Advanced usage

The column used as your model's time column is not limited to a text or date type. In the following example, the time column, di, is an integer:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
-- Orders are partitioned by the di int column
MODEL (
  name sushi.orders,
  dialect duckdb,
  kind INCREMENTAL_BY_TIME_RANGE (
    time_column (order_date_int, '%Y%m%d')
  ),
);

SELECT
  id::INT AS id, -- Primary key
  customer_id::INT AS customer_id, -- Id of customer who made the order
  waiter_id::INT AS waiter_id, -- Id of waiter who took the order
  start_ts::TEXT AS start_ts, -- Start timestamp
  end_ts::TEXT AS end_ts, -- End timestamp
  di::INT AS order_date_int -- Date of order
FROM raw.orders
WHERE
  order_date_int BETWEEN @start_ds AND @end_ds
SQLMesh will handle casting the start and end dates to the type of your time column. The format is reflected in the time column format string.