Joins

Joins are a powerful part of MetricFlow and simplify the process of making all valid dimensions available for your metrics at query time, regardless of where they are defined in different semantic models. With Joins, you can also create metrics using measures from different semantic models.

Joins use entities defined in your semantic model configs as the join keys between tables. Assuming entities are defined in the semantic model, MetricFlow creates a graph using the semantic models as nodes and the join paths as edges to perform joins automatically. MetricFlow chooses the appropriate join type and avoids fan-out or chasm joins with other tables based on the entity types.

What are fan-out or chasm joins?

— Fan-out joins are when one row in a table is joined to multiple rows in another table, resulting in more output rows than input rows.

— Chasm joins are when two tables have a many-to-many relationship through an intermediate table, and the join results in duplicate or missing data.

Types of joins

Joins are auto-generated

MetricFlow automatically generates the necessary joins to the defined semantic objects, eliminating the need for you to create new semantic models or configuration files.

This document explains the different types of joins that can be used with entities and how to query them using the CLI.

MetricFlow primarily uses left joins for joins, and restricts the use of fan-out and chasm joins. Refer to the table below to identify which joins are or aren't allowed based on specific entity types to prevent the creation of risky joins.

entity type - Table A	entity type - Table B	Join type
Primary	Primary	✅ Left
Primary	Unique	✅ Left
Primary	Foreign	❌ Fan-out (Not allowed)
Unique	Primary	✅ Left
Unique	Unique	✅ Left
Unique	Foreign	❌ Fan-out (Not allowed)
Foreign	Primary	✅ Left
Foreign	Unique	✅ Left
Foreign	Foreign	❌ Fan-out (Not allowed)

Example

The following example uses two semantic models with a common entity and shows a MetricFlow query that requires a join between the two semantic models.

Let's say you have two semantic models, transactions and user_signup as seen in the following example:

semantic_models:
  - name: transactions
    entities:
      - name: id
        type: primary
      - name: user
        type: foreign
        expr: user_id
    measures:
      - name: average_purchase_price
        agg: avg
        expr: purchase_price
  - name: user_signup
    entities:
      - name: user
        type: primary
        expr: user_id
    dimensions:
      - name: type
        type: categorical

MetricFlow will use user_id as the join key to join two semantic models, transactions and user_signup. This enables you to query the average_purchase_price metric in transactions, sliced by the type dimension in the user_signup semantic model.

Note that the average_purchase_price measure is defined in the transactions semantic model, where user_id is a foreign entity. However, the user_signup semantic model has user_id as a primary entity.

Since this is a foreign-to-primary relationship, a left join is implemented where the transactions semantic model joins the user_signup semantic model since the average_purchase_price measure is defined in the transactions semantic model.

When querying dimensions from different semantic models using the CLI, a double underscore (or dunder) is added to the dimension name after the joining entity. In the CLI query shown below, user_id__type is included as a dimension.

mf query --metrics average_purchase_price --dimensions metric_time,user_id__type 

Multi-hop joins

MetricFlow allows users to join measures and dimensions across a graph of entities by moving from one table to another within a graph. This is referred to as "multi-hop join".

MetricFlow can join up to three tables, supporting multi-hop joins with a limit of two hops. This does the following:

• Enables complex data analysis without ambiguous paths.
• Supports navigating through data models, like moving from orders to customers to country tables.

While direct three-hop paths are limited to prevent confusion from multiple routes to the same data, MetricFlow does allow joining more than three tables if the joins don’t exceed two hops to reach a dimension.

For example, if you have two models, country and region, where customers are linked to countries, which in turn are linked to regions, you can join all of them in a single SQL query and can dissect orders by customer__country_country_name but not by customer__country__region_name.

Notice how the schema can be translated into the following three MetricFlow semantic models to create the metric 'Average purchase price by country' using the purchase_price measure from the sales table and the country_name dimension from the country_dim table.

semantic_models:
  - name: sales
    defaults:
      agg_time_dimension: first_ordered_at
    entities:
      - name: id
        type: primary
      - name: user_id
        type: foreign
    measures:
      - name: average_purchase_price
        agg: avg
        expr: purchase_price
    dimensions:
      - name: metric_time
        type: time
        type_params:
  - name: user_signup
    entities:
      - name: user_id
        type: primary
      - name: country_id
        type: Unique
    dimensions:
      - name: signup_date
        type: time
      - name: country_dim

  - name: country
    entities:
      - name: country_id
        type: primary
    dimensions:
      - name: country_name
        type: categorical

Query multi-hop joins

To query dimensions without a multi-hop join involved, you can use the fully qualified dimension name with the syntax entity double underscore (dunder) dimension, like entity__dimension.

For dimensions retrieved by a multi-hop join, you need to additionally provide the entity path as a list, like user_id.