Apache Spark configurations

Configuring tables

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

OptionDescriptionRequired?Example
file_formatThe file format to use when creating tables (parquet, delta, csv, json, text, jdbc, orc, hive or libsvm).Optionalparquet
location_rootThe created table uses the specified directory to store its data. The table alias is appended to it.Optional/mnt/root
partition_byPartition the created table by the specified columns. A directory is created for each partition.Optionaldate_day
clustered_byEach partition in the created table will be split into a fixed number of buckets by the specified columns.Optionalcountry_code
bucketsThe number of buckets to create while clusteringRequired if clustered_by is specified8

Incremental models

Changelog

dbt seeks to offer useful, intuitive modeling abstractions by means of its built-in configurations and materializations. Because there is so much variance between Apache Spark clusters out in the world—not to mention the powerful features offered to Databricks users by the Delta file format and custom runtime—making sense of all the available options is an undertaking in its own right.

For that reason, the dbt-spark plugin leans heavily on the incremental_strategy config. This config tells the incremental materialization how to build models in runs beyond their first. It can be set to one of three values:

  • append (default): Insert new records without updating or overwriting any existing data.
  • insert_overwrite: If partition_by is specified, overwrite partitions in the table with new data. If no partition_by is specified, overwrite the entire table with new data.
  • merge (Delta Lake only): Match records based on a unique_key; update old records, insert new ones. (If no unique_key is specified, all new data is inserted, similar to append.)

Each of these strategies has its pros and cons, which we'll discuss below. As with any model config, incremental_strategy may be specified in dbt_project.yml or within a model file's config() block.

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.

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

spark_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

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.

Usage notes:

  • This strategy is not supported for tables with file_format: delta.
  • This strategy is not available when connecting via Databricks SQL endpoints (method: odbc + endpoint).
  • If connecting via a Databricks cluster + ODBC driver (method: odbc + cluster), you must include set spark.sql.sources.partitionOverwriteMode DYNAMIC in the cluster Spark Config in order for dynamic partition replacement to work (incremental_strategy: insert_overwrite + partition_by).
Databricks cluster: Spark Config

Databricks cluster: Spark Config

spark_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 events
group by 1

The merge strategy

Changelog

Usage notes: The merge incremental strategy requires:

  • file_format: delta
  • Databricks Runtime 5.1 and above

dbt will run an atomic merge statement which looks nearly identical 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).

delta_incremental.sql
{{ config(
materialized='incremental',
file_format='delta',
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

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 '*'.

Always schema, never database

Changelog

Apache Spark uses the terms "schema" and "database" interchangeably. dbt understands database to exist at a higher level than schema. As such, you should never use or set database as a node config or in the target profile when running dbt-spark.

If you want to control the schema/database in which dbt will materialize models, use the schema config and generate_schema_name macro only.

Databricks configurations

To access features exclusive to Databricks runtimes, such as snapshots and the merge incremental strategy, you will want to use the Delta 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
seeds:
+file_format: delta
snapshots:
+file_format: delta