Skip to main content

Databricks configurations

Configuring tables

When materializing a model as table, you may include several optional configs that are specific to the dbt-databricks plugin, in addition to the standard model configs.

The append strategy

Following the append strategy, dbt will perform an insert into statement with all new data. The appeal of this strategy is that it is straightforward and functional across all platforms, file types, connection methods, and Apache Spark versions. However, this strategy cannot update, overwrite, or delete existing data, so it is likely to insert duplicate records for many data sources.

databricks_incremental.sql
{{ config(
materialized='incremental',
incremental_strategy='append',
) }}

-- All rows returned by this query will be appended to the existing table

select * from {{ ref('events') }}
{% if is_incremental() %}
where event_ts > (select max(event_ts) from {{ this }})
{% endif %}

The insert_overwrite strategy

caution

This strategy is currently only compatible with All Purpose Clusters, not SQL Warehouses.

This strategy is most effective when specified alongside a partition_by clause in your model config. dbt will run an atomic insert overwrite statement that dynamically replaces all partitions included in your query. Be sure to re-select all of the relevant data for a partition when using this incremental strategy.

If no partition_by is specified, then the insert_overwrite strategy will atomically replace all contents of the table, overriding all existing data with only the new records. The column schema of the table remains the same, however. This can be desirable in some limited circumstances, since it minimizes downtime while the table contents are overwritten. The operation is comparable to running truncate + insert on other databases. For atomic replacement of Delta-formatted tables, use the table materialization (which runs create or replace) instead.

databricks_incremental.sql
{{ config(
materialized='incremental',
partition_by=['date_day'],
file_format='parquet'
) }}

/*
Every partition returned by this query will be overwritten
when this model runs
*/

with new_events as (

select * from {{ ref('events') }}

{% if is_incremental() %}
where date_day >= date_add(current_date, -1)
{% endif %}

)

select
date_day,
count(*) as users

from new_events
group by 1

The merge strategy

The merge incremental strategy requires:

  • file_format: delta or hudi
  • Databricks Runtime 5.1 and above for delta file format
  • Apache Spark for hudi file format

The Databricks adapter will run an atomic merge statement similar to the default merge behavior on Snowflake and BigQuery. If a unique_key is specified (recommended), dbt will update old records with values from new records that match on the key column. If a unique_key is not specified, dbt will forgo match criteria and simply insert all new records (similar to append strategy).

Specifying merge as the incremental strategy is optional since it's the default strategy used when none is specified.

merge_incremental.sql
{{ config(
materialized='incremental',
file_format='delta', # or 'hudi'
unique_key='user_id',
incremental_strategy='merge'
) }}

with new_events as (

select * from {{ ref('events') }}

{% if is_incremental() %}
where date_day >= date_add(current_date, -1)
{% endif %}

)

select
user_id,
max(date_day) as last_seen

from events
group by 1

The replace_where strategy

The replace_where incremental strategy requires:

  • file_format: delta
  • Databricks Runtime 12.0 and above

dbt will run an atomic replace where statement which selectively overwrites data matching one or more incremental_predicates specified as a string or array. Only rows matching the predicates will be inserted. If no incremental_predicates are specified, dbt will perform an atomic insert, as with append.

caution

replace_where inserts data into columns in the order provided, rather than by column name. If you reorder columns and the data is compatible with the existing schema, you may silently insert values into an unexpected column. If the incoming data is incompatible with the existing schema, you will instead receive an error.

replace_where_incremental.sql
{{ config(
materialized='incremental',
file_format='delta',
incremental_strategy = 'replace_where'
incremental_predicates = 'user_id >= 10000' # Never replace users with ids < 10000
) }}

with new_events as (

select * from {{ ref('events') }}

{% if is_incremental() %}
where date_day >= date_add(current_date, -1)
{% endif %}

)

select
user_id,
max(date_day) as last_seen

from events
group by 1

Selecting compute per model

Beginning in version 1.7.2, you can assign which compute resource to use on a per-model basis. For SQL models, you can select a SQL Warehouse (serverless or provisioned) or an all purpose cluster. For details on how this feature interacts with python models, see Specifying compute for Python models.

note

This is an optional setting. If you do not configure this as shown below, we will default to the compute specified by http_path in the top level of the output section in your profile. This is also the compute that will be used for tasks not associated with a particular model, such as gathering metadata for all tables in a schema.

To take advantage of this capability, you will need to add compute blocks to your profile:

profile.yml

profile-name:
target: target-name # this is the default target
outputs:
target-name:
type: databricks
catalog: optional catalog name if you are using Unity Catalog
schema: schema name # Required
host: yourorg.databrickshost.com # Required

### This path is used as the default compute
http_path: /sql/your/http/path # Required

### New compute section
compute:

### Name that you will use to refer to an alternate compute
Compute1:
http_path: '/sql/your/http/path' # Required of each alternate compute

### A third named compute, use whatever name you like
Compute2:
http_path: '/some/other/path' # Required of each alternate compute
...

target-name: # additional targets
...
### For each target, you need to define the same compute,
### but you can specify different paths
compute:

### Name that you will use to refer to an alternate compute
Compute1:
http_path: '/sql/your/http/path' # Required of each alternate compute

### A third named compute, use whatever name you like
Compute2:
http_path: '/some/other/path' # Required of each alternate compute
...

The new compute section is a map of user chosen names to objects with an http_path property. Each compute is keyed by a name which is used in the model definition/configuration to indicate which compute you wish to use for that model/selection of models. We recommend choosing a name that is easily recognized as the compute resources you're using, such as the name of the compute resource inside the Databricks UI.

note

You need to use the same set of names for compute across your outputs, though you may supply different http_paths, allowing you to use different computes in different deployment scenarios.

To configure this inside of dbt Cloud, use the extended attributes feature on the desired environments:


compute:
Compute1:
http_path: /SOME/OTHER/PATH
Compute2:
http_path: /SOME/OTHER/PATH

Specifying the compute for models

As with many other configuaration options, you can specify the compute for a model in multiple ways, using databricks_compute. In your dbt_project.yml, the selected compute can be specified for all the models in a given directory:

dbt_project.yml

...

models:
+databricks_compute: "Compute1" # use the `Compute1` warehouse/cluster for all models in the project...
my_project:
clickstream:
+databricks_compute: "Compute2" # ...except for the models in the `clickstream` folder, which will use `Compute2`.

snapshots:
+databricks_compute: "Compute1" # all Snapshot models are configured to use `Compute1`.

For an individual model the compute can be specified in the model config in your schema file.

schema.yml

models:
- name: table_model
config:
databricks_compute: Compute1
columns:
- name: id
data_type: int

Alternatively the warehouse can be specified in the config block of a model's SQL file.

model.sql

{{
config(
materialized='table',
databricks_compute='Compute1'
)
}}
select * from {{ ref('seed') }}

To validate that the specified compute is being used, look for lines in your dbt.log like:

Databricks adapter ... using default compute resource.

or

Databricks adapter ... using compute resource <name of compute>.

Specifying compute for Python models

Materializing a python model requires execution of SQL as well as python. Specifically, if your python model is incremental, the current execution pattern involves executing python to create a staging table that is then merged into your target table using SQL.

When you specify your databricks_compute for a python model, you are currently only specifying which compute to use when running the model-specific SQL. If you wish to use a different compute for executing the python itself, you must specify an alternate compute in the config for the model. For example:

model.py

def model(dbt, session):
dbt.config(
http_path="sql/protocolv1/..."
)

If your default compute is a SQL Warehouse, you will need to specify an all purpose cluster http_path in this way.

Persisting model descriptions

Relation-level docs persistence is supported in dbt v0.17.0. For more information on configuring docs persistence, see the docs.

When the persist_docs option is configured appropriately, you'll be able to see model descriptions in the Comment field of describe [table] extended or show table extended in [database] like '*'.

Default file format configurations

To access advanced incremental strategies features, such as snapshots and the merge incremental strategy, you will want to use the Delta or Hudi file format as the default file format when materializing models as tables.

It's quite convenient to do this by setting a top-level configuration in your project file:

dbt_project.yml
models:
+file_format: delta # or hudi

seeds:
+file_format: delta # or hudi

snapshots:
+file_format: delta # or hudi

Setting table properties

Table properties can be set with your configuration for tables or views using tblproperties:

with_table_properties.sql
{{ config(
tblproperties={
'delta.autoOptimize.optimizeWrite' : 'true',
'delta.autoOptimize.autoCompact' : 'true'
}
) }}
caution

These properties are sent directly to Databricks without validation in dbt, so be thoughtful with how you use this feature. You will need to do a full refresh of incremental materializations if you change their tblproperties.

One application of this feature is making delta tables compatible with iceberg readers using the Universal Format.

{{ config(
tblproperties={
'delta.enableIcebergCompatV2' = 'true'
'delta.universalFormat.enabledFormats' = 'iceberg'
}
) }}

tblproperties can be specified for python models, but they will be applied via an ALTER statement after table creation. This is due to a limitation in PySpark.

0