Snowflake configurations
Iceberg table format beta
The dbt-snowflake adapter supports the Iceberg table format. It is available for three of the Snowflake materializations:
For now, to create Iceberg tables, you must implement a behavior flag due to performance impact related to using Iceberg tables. Snowflake does not support is_iceberg
on the Show Objects
query, which dbt depends on for metadata.
To use Iceberg, set the enable_iceberg_materializations
flag to True
in your dbt_project.yml:
flags:
enable_iceberg_materializations: True
The following configurations are supported:
Field | Type | Required | Description | Sample input |
---|---|---|---|---|
Table Format | String | Yes | Configures the object table format. | iceberg |
External volume | String | Yes | Specifies the identifier (name) of the external volume where the Iceberg table stores its metadata files and data in Parquet format. If you don’t specify this parameter, the Iceberg table defaults to the external volume for the schema, database, or account. The schema takes precedence over the database, and the database takes precedence over the account. | my_s3_bucket |
Base location Subpath | String | No | Defines the directory path for Snowflake to write the table data and metadata files. If you change this after creating the table, it will write a new table but leave the existing one. | jaffle_marketing_folder |
Example configuration
To configure an Iceberg table materialization in dbt, refer to the example configuration:
{{
config(
materialized = "table",
table_format="iceberg",
external_volume="s3_iceberg_snow",
)
}}
select * from {{ ref('raw_orders') }}
Base location
Snowflake's CREATE ICEBERG TABLE
DDL requires that a base_location
be provided. dbt defines this parameter on the user's behalf to streamline usage and enforce basic isolation of table data within the EXTERNAL VOLUME
. The default behavior in dbt is to provide a base_location
string of the form: _dbt/{SCHEMA_NAME}/{MODEL_NAME}
Base Location Subpath
We recommend using dbt's auto-generated base_location
. However, if you need to customize the resulting base_location
, dbt allows users to configure a base_location_subpath
. When specified, the subpath concatenates to the end of the previously described pattern for base_location
string generation.
For example, config(base_location_subpath="prod")
will generate a base_location
of the form _dbt/{SCHEMA_NAME}/{MODEL_NAME}
.
A theoretical (but not recommended) use case is re-using an EXTERNAL VOLUME
while maintaining isolation across development and production environments. We recommend against this as storage permissions should configured on the external volume and underlying storage, not paths that any analytics engineer can modify.
Rationale
dbt manages base_location
on behalf of users to enforce best practices. With Snowflake-managed Iceberg format tables, the user owns and maintains the data storage of the tables in an external storage solution (the declared external volume
). The base_ location
parameter declares where to write the data within the external volume. The Snowflake Iceberg catalog keeps track of your Iceberg table regardless of where the data lives within the external volume
declared and the base_location
provided. However, Snowflake permits passing anything into the base_location
field, including an empty string, even reusing the same path across multiple tables. This behavior could result in future technical debt because it will limit the ability to:
- Navigate the underlying object store (S3/Azure blob)
- Read Iceberg tables via an object-store integration
- Grant schema-specific access to tables via object store
- Use a crawler pointed at the tables within the external volume to build a new catalog with another tool
To maintain best practices, we enforce an input. Currently, we do not support overriding the default base location
input but will consider it based on user feedback.
In summary, dbt-snowflake does not support arbitrary definition of base_location
for Iceberg tables. Instead, dbt, by default, writes your tables within a _dbt/{SCHEMA_NAME}/{TABLE_NAME}
prefix to ensure easier object-store observability and auditability.
Limitations
There are some limitations to the implementation you need to be aware of:
- Using Iceberg tables with dbt, the result is that your query is materialized in Iceberg. However, often, dbt creates intermediary objects as temporary and transient tables for certain materializations, such as incremental ones. It is not possible to configure these temporary objects also to be Iceberg-formatted. You may see non-Iceberg tables created in the logs to support specific materializations, but they will be dropped after usage.
- You cannot incrementally update a preexisting incremental model to be an Iceberg table. To do so, you must fully rebuild the table with the
--full-refresh
flag.
Dynamic tables
The Snowflake adapter supports dynamic tables.
This materialization is specific to Snowflake, which means that any model configuration that
would normally come along for the ride from dbt-core
(e.g. as with a view
) may not be available
for dynamic tables. This gap will decrease in future patches and versions.
While this materialization is specific to Snowflake, it very much follows the implementation
of materialized views.
In particular, dynamic tables have access to the on_configuration_change
setting.
Dynamic tables are supported with the following configuration parameters:
Learn more about these parameters in Snowflake's docs:
Target lag
Snowflake allows two configuration scenarios for scheduling automatic refreshes:
- Time-based — Provide a value of the form
<int> { seconds | minutes | hours | days }
. For example, if the dynamic table needs to be updated every 30 minutes, usetarget_lag='30 minutes'
. - Downstream — Applicable when the dynamic table is referenced by other dynamic tables. In this scenario,
target_lag='downstream'
allows for refreshes to be controlled at the target, instead of at each layer.
Learn more about target_lag
in Snowflake's docs.
Limitations
As with materialized views on most data platforms, there are limitations associated with dynamic tables. Some worth noting include:
- Dynamic table SQL has a limited feature set.
- Dynamic table SQL cannot be updated; the dynamic table must go through a
--full-refresh
(DROP/CREATE). - Dynamic tables cannot be downstream from: materialized views, external tables, streams.
- Dynamic tables cannot reference a view that is downstream from another dynamic table.
Find more information about dynamic table limitations in Snowflake's docs.
For dbt limitations, these dbt features are not supported:
Temporary tables
Incremental table merges for Snowflake prefer to utilize a view
rather than a temporary table
. The reasoning is to avoid the database write step that a temporary table would initiate and save compile time.
However, some situations remain where a temporary table would achieve results faster or more safely. The tmp_relation_type
configuration enables you to opt in to temporary tables for incremental builds. This is defined as part of the model configuration.
To guarantee accuracy, an incremental model using the delete+insert
strategy with a unique_key
defined requires a temporary table; trying to change this to a view will result in an error.
Defined in the project YAML:
name: my_project
...
models:
<resource-path>:
+tmp_relation_type: table | view ## If not defined, view is the default.
In the configuration format for the model SQL file:
{{ config(
tmp_relation_type="table | view", ## If not defined, view is the default.
) }}
Transient tables
Snowflake supports the creation of transient tables. Snowflake does not preserve a history for these tables, which can result in a measurable reduction of your Snowflake storage costs. Transient tables participate in time travel to a limited degree with a retention period of 1 day by default with no fail-safe period. Weigh these tradeoffs when deciding whether or not to configure your dbt models as transient
. By default, all Snowflake tables created by dbt are transient
.
Configuring transient tables in dbt_project.yml
A whole folder (or package) can be configured to be transient (or not) by adding a line to the dbt_project.yml
file. This config works just like all of the model configs defined in dbt_project.yml
.
name: my_project
...
models:
+transient: false
my_project:
...
Configuring transience for a specific model
A specific model can be configured to be transient by setting the transient
model config to true
.
{{ config(materialized='table', transient=true) }}
select * from ...
Query tags
Query tags are a Snowflake parameter that can be quite useful later on when searching in the QUERY_HISTORY view.
dbt supports setting a default query tag for the duration of its Snowflake connections in
your profile. You can set more precise values (and override the default) for subsets of models by setting
a query_tag
model config or by overriding the default set_query_tag
macro:
models:
<resource-path>:
+query_tag: dbt_special
{{ config(
query_tag = 'dbt_special'
) }}
select ...
In this example, you can set up a query tag to be applied to every query with the model's name.
{% macro set_query_tag() -%}
{% set new_query_tag = model.name %}
{% if new_query_tag %}
{% set original_query_tag = get_current_query_tag() %}
{{ log("Setting query_tag to '" ~ new_query_tag ~ "'. Will reset to '" ~ original_query_tag ~ "' after materialization.") }}
{% do run_query("alter session set query_tag = '{}'".format(new_query_tag)) %}
{{ return(original_query_tag)}}
{% endif %}
{{ return(none)}}
{% endmacro %}
Note: query tags are set at the session level. At the start of each model materialization, if the model has a custom query_tag
configured, dbt will run alter session set query_tag
to set the new value. At the end of the materialization, dbt will run another alter
statement to reset the tag to its default value. As such, build failures midway through a materialization may result in subsequent queries running with an incorrect tag.
Merge behavior (incremental models)
The incremental_strategy
config controls how dbt builds incremental models. By default, dbt will use a merge statement on Snowflake to refresh incremental tables.
Snowflake's merge
statement fails with a "nondeterministic merge" error if the unique_key
specified in your model config is not actually unique. If you encounter this error, you can instruct dbt to use a two-step incremental approach by setting the incremental_strategy
config for your model to delete+insert
.
Configuring table clustering
dbt supports table clustering on Snowflake. To control clustering for a table or incremental model, use the cluster_by
config. When this configuration is applied, dbt will do two things:
- It will implicitly order the table results by the specified
cluster_by
fields - It will add the specified clustering keys to the target table
By using the specified cluster_by
fields to order the table, dbt minimizes the amount of work required by Snowflake's automatic clustering functionality. If an incremental model is configured to use table clustering, then dbt will also order the staged dataset before merging it into the destination table. As such, the dbt-managed table should always be in a mostly clustered state.
Using cluster_by
The cluster_by
config accepts either a string, or a list of strings to use as clustering keys. The following example will create a sessions table that is clustered by the session_start
column.
{{
config(
materialized='table',
cluster_by=['session_start']
)
}}
select
session_id,
min(event_time) as session_start,
max(event_time) as session_end,
count(*) as count_pageviews
from {{ source('snowplow', 'event') }}
group by 1
The code above will be compiled to SQL that looks (approximately) like this:
create or replace table my_database.my_schema.my_table as (
select * from (
select
session_id,
min(event_time) as session_start,
max(event_time) as session_end,
count(*) as count_pageviews
from {{ source('snowplow', 'event') }}
group by 1
)
-- this order by is added by dbt in order to create the
-- table in an already-clustered manner.
order by session_start
);
alter table my_database.my_schema.my_table cluster by (session_start);
Automatic clustering
Automatic clustering is enabled by default in Snowflake today, no action is needed to make use of it. Though there is an automatic_clustering
config, it has no effect except for accounts with (deprecated) manual clustering enabled.
If manual clustering is still enabled for your account, you can use the automatic_clustering
config to control whether or not automatic clustering is enabled for dbt models. When automatic_clustering
is set to true
, dbt will run an alter table <table name> resume recluster
query after building the target table.
The automatic_clustering
config can be specified in the dbt_project.yml
file, or in a model config()
block.
models:
+automatic_clustering: true
Configuring virtual warehouses
The default warehouse that dbt uses can be configured in your Profile for Snowflake connections. To override the warehouse that is used for specific models (or groups of models), use the snowflake_warehouse
model configuration. This configuration can be used to specify a larger warehouse for certain models in order to control Snowflake costs and project build times.
- YAML code
- SQL code
The example config below changes the warehouse for a group of models with a config argument in the yml.
name: my_project
version: 1.0.0
...
models:
+snowflake_warehouse: "EXTRA_SMALL" # use the `EXTRA_SMALL` warehouse for all models in the project...
my_project:
clickstream:
+snowflake_warehouse: "EXTRA_LARGE" # ...except for the models in the `clickstream` folder, which will use the `EXTRA_LARGE` warehouse.
snapshots:
+snowflake_warehouse: "EXTRA_LARGE" # all Snapshot models are configured to use the `EXTRA_LARGE` warehouse.
The example config below changes the warehouse for a single model with a config() block in the sql model.
{{
config(
materialized='table',
snowflake_warehouse='EXTRA_LARGE'
)
}}
with
aggregated_page_events as (
select
session_id,
min(event_time) as session_start,
max(event_time) as session_end,
count(*) as count_page_views
from {{ source('snowplow', 'event') }}
group by 1
),
index_sessions as (
select
*,
row_number() over (
partition by session_id
order by session_start
) as page_view_in_session_index
from aggregated_page_events
)
select * from index_sessions
Copying grants
When the copy_grants
config is set to true
, dbt will add the copy grants
DDL qualifier when rebuilding tables and views. The default value is false
.
models:
+copy_grants: true
Secure views
To create a Snowflake secure view, use the secure
config for view models. Secure views can be used to limit access to sensitive data. Note: secure views may incur a performance penalty, so you should only use them if you need them.
The following example configures the models in the sensitive/
folder to be configured as secure views.
name: my_project
version: 1.0.0
models:
my_project:
sensitive:
+materialized: view
+secure: true
The Snowflake adapter supports dynamic tables.
This materialization is specific to Snowflake, which means that any model configuration that
would normally come along for the ride from dbt-core
(e.g. as with a view
) may not be available
for dynamic tables. This gap will decrease in future patches and versions.
While this materialization is specific to Snowflake, it very much follows the implementation
of materialized views.
In particular, dynamic tables have access to the on_configuration_change
setting.
Dynamic tables are supported with the following configuration parameters:
Learn more about these parameters in Snowflake's docs:
Target lag
Snowflake allows two configuration scenarios for scheduling automatic refreshes:
- Time-based — Provide a value of the form
<int> { seconds | minutes | hours | days }
. For example, if the dynamic table needs to be updated every 30 minutes, usetarget_lag='30 minutes'
. - Downstream — Applicable when the dynamic table is referenced by other dynamic tables. In this scenario,
target_lag='downstream'
allows for refreshes to be controlled at the target, instead of at each layer.
Learn more about target_lag
in Snowflake's docs.
Limitations
As with materialized views on most data platforms, there are limitations associated with dynamic tables. Some worth noting include:
- Dynamic table SQL has a limited feature set.
- Dynamic table SQL cannot be updated; the dynamic table must go through a
--full-refresh
(DROP/CREATE). - Dynamic tables cannot be downstream from: materialized views, external tables, streams.
- Dynamic tables cannot reference a view that is downstream from another dynamic table.
Find more information about dynamic table limitations in Snowflake's docs.
For dbt limitations, these dbt features are not supported:
Source freshness known limitation
Snowflake calculates source freshness using information from the LAST_ALTERED
column, meaning it relies on a field updated whenever any object undergoes modification, not only data updates. No action must be taken, but analytics teams should note this caveat.
Per the Snowflake documentation:
The
LAST_ALTERED
column is updated when the following operations are performed on an object:
- DDL operations.
- DML operations (for tables only).
- Background maintenance operations on metadata performed by Snowflake.