Understand and Optimize RAI Native App Costs

Using RelationalAI (RAI) on Snowflake incurs costs from both Snowflake and RAI itself. Similar to Snowflake, RAI’s cloud architecture separates storage and compute usage, allowing for cost optimization based on specific needs.

Download the Cost Monitoring NotebookCustomizable SQL queries for monitoring costs associated with the RAI Native App.

Before you start, make sure that:

• You have the RAI Native App installed and activated on your Snowflake account.

What drives costs

Before you install, it helps to know what actually costs money:

• Reasoner compute is the main driver. Reasoners consume Snowflake compute while running, and RelationalAI adds a surcharge on that compute. Larger reasoner sizes cost more per hour, so choose the smallest size that meets your performance needs and let reasoners auto-suspend.
• The CDC engine consumes Snowflake compute continuously while CDC is enabled, sized by its instance family. If you never enable CDC, this cost never appears.
• The app warehouse and serverless tasks consume Snowflake compute for interoperability and background work. Some serverless tasks keep running even after you deactivate the app.
• Storage (the app database and deployed model schemas) has no RelationalAI surcharge and a small Snowflake storage cost.

You can pause most of these by deactivating the app when it is idle. For the full list of what the app creates and how to inspect it, see Resources the app creates.

How Costs Are Calculated

Snowflake costs are based on compute and storage resource usage, detailed in the Snowflake User Guide. RAI costs appear as a surcharge for compute usage, primarily from reasoners. There’s no surcharge for storage usage.

The following resources all contribute to costs generated by the RAI Native App:

Resource	Description	Snowflake Charge	RAI Surcharge	Notes
App Database	Stores data used by the app.	Yes	No	This database is named RELATIONALAI unless you customized the app name during install.
App Warehouse	Used for Snowflake interoperability.	Yes	No	The RELATIONAL_AI_ERP_WAREHOUSE is automatically created the first time you activate the app. The default deactivate() suspends it; deactivate('all') drops it and recreates it on the next activation.
Serverless tasks	Used to schedule billing, garbage collection, upgrades, and other tasks.	Yes	No	Some tasks, like those related to billing, may generate small amounts of compute consumption even when the app is deactivated.
Compute Pools	Used for internal app operations and to host reasoners.	No	No	Compute pools themselves generate no charge. Costs are incurred only by reasoners that run inside them, and by reserved-node capacity if you configure it. See Compute Pools for the names and instance families of each compute pool managed by the app.
CDC Reasoner	Used by the CDC Service to process data streams.	Yes	No	Runs continuously while CDC is enabled. Costs are based on uptime and the CDC engine size. Suspending CDC pauses this cost.
User Reasoners	Used to evaluate queries from PyRel semantic models.	Yes	Yes	Costs are based on uptime and are determined by the host compute pool’s instance family and the number of compute nodes used by the reasoner.

Optimize costs

Configure reasoners to auto-suspend

Reasoners consume compute resources even when they’re not actively processing queries. To reduce costs, reasoners automatically suspend after a period of inactivity. You can adjust the auto-suspend timeout to better fit your needs.

Note

Suspended reasoners do not consume compute resources, but they retain their configuration and data. This incurs a small, but negligible cost for storage. Delete a reasoner when it’s no longer needed to eliminate costs entirely.

Related resources:

• Change a logic reasoner’s auto-suspend settings
• Change a prescriptive reasoner’s auto-suspend settings
• Change a predictive reasoner’s auto-suspend settings

Suspend data streams that are rarely updated

Streams that don’t require minute-by-minute processing can be suspended to reduce costs. You may also disable the CDC Service to suspend the CDC engine and stop processing change tracking data.

Note

While the CDC service is disabled, some compute is consumed as changes to data streams are tracked and queued for processing. To eliminate all costs associated with data streams, you must delete the streams.

Shut down the app when not in use

You can deactivate the app to reduce costs while the app is not in use. When the app is deactivated, the CDC service is suspended and all reasoners are deleted. The default deactivate() suspends the app warehouse and compute pools; deactivate('all') drops them. Tasks related to diagnostics and billing stay active and generate small amounts of compute consumption even after deactivation. Deactivation is not an uninstall: the app database and your account roles remain.

Monitor costs

Use the queries below to monitor costs associated with your RAI Native App. For more detailed monitoring, administrators may tag the native app’s storage and compute resources to track RAI costs with Snowflake’s cost management framework.

Note

COMING SOON: Snowflake will add support for detailed cost monitoring for native apps through standard cost monitoring tools.

Download the Cost Monitoring NotebookAll of the SQL queries in this section, ready to customize and run.

Credits consumed by all reasoners per month

Requires the billing_admin application role.

SQL

SELECT
  DATE_TRUNC('MONTH', hour) AS first_day_of_month,
  SUM(credits) AS sum_of_rai_units
FROM
  relationalai.consumption.engine_credit_hours
GROUP BY
  first_day_of_month
ORDER BY
  first_day_of_month DESC;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    DATE_TRUNC('MONTH', hour) AS first_day_of_month,
    SUM(credits) AS sum_of_rai_units
FROM
    relationalai.consumption.engine_credit_hours
GROUP BY
    first_day_of_month
ORDER BY
    first_day_of_month DESC;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.engine_credit_hours.month")
    print(rows)

Credits consumed by all reasoners per day

Requires the billing_admin application role.

SQL

SELECT
  DATE_TRUNC('day',c.HOUR) as day, sum(c.credits) AS total_daily_rai_units
FROM
  relationalai.consumption.engine_credit_hours c
WHERE
  hour >= DATEADD(month, -1, CURRENT_TIMESTAMP())
GROUP BY all ORDER BY 1 DESC;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    DATE_TRUNC('day', c.HOUR) AS day,
    SUM(c.credits) AS total_daily_rai_units
FROM
    relationalai.consumption.engine_credit_hours c
WHERE
    hour >= DATEADD(month, -1, CURRENT_TIMESTAMP())
GROUP BY all ORDER BY 1 DESC;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.engine_credit_hours.day")
    print(rows)

Credits consumed per reasoner in the past month

Requires the billing_admin application role.

SQL

SELECT
  engine_name,
  SUM(credits) AS total_credits,
  COUNT(hour) AS hours
FROM
  relationalai.consumption.engine_credit_hours
WHERE
  hour >= DATEADD(month, -1, CURRENT_TIMESTAMP())
GROUP BY
  engine_name
ORDER BY
  total_credits DESC;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    engine_name,
    SUM(credits) AS total_credits,
    COUNT(hour) AS hours
FROM
    relationalai.consumption.engine_credit_hours
WHERE
    hour >= DATEADD(month, -1, CURRENT_TIMESTAMP())
GROUP BY
    engine_name
ORDER BY
    total_credits DESC;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.engine_credit_hours.by_reasoner")
    print(rows)

Credits consumed per compute pool per month

Requires access to the SNOWFLAKE.ACCOUNT_USAGE.SNOWPARK_CONTAINER_SERVICES_HISTORY view, such as through the USAGE_VIEWER database role.

SQL

SELECT
    DATE_TRUNC('MONTH', START_TIME) AS first_day_of_month,
    compute_pool_name,
    SUM(credits_used) as total_credits
FROM
   SNOWFLAKE.ACCOUNT_USAGE.SNOWPARK_CONTAINER_SERVICES_HISTORY
WHERE
    APPLICATION_NAME = 'RELATIONALAI'
GROUP BY
    first_day_of_month, compute_pool_name
ORDER BY
    first_day_of_month DESC,
    compute_pool_name;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    DATE_TRUNC('MONTH', START_TIME) AS first_day_of_month,
    compute_pool_name,
    SUM(credits_used) as total_credits
FROM
    SNOWFLAKE.ACCOUNT_USAGE.SNOWPARK_CONTAINER_SERVICES_HISTORY
WHERE
    APPLICATION_NAME = 'RELATIONALAI'
GROUP BY
    first_day_of_month, compute_pool_name
ORDER BY
    first_day_of_month DESC,
    compute_pool_name;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.compute_pools.month")
    print(rows)

Credits consumed by the app warehouse per month

Requires access to the SNOWFLAKE.ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY view, such as through the USAGE_VIEWER database role.

SQL

SELECT
    DATE_TRUNC('MONTH', START_TIME) AS first_day_of_month,
    SUM(credits_used) as total_credits
FROM
    SNOWFLAKE.ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY
WHERE
    WAREHOUSE_NAME = 'RELATIONAL_AI_ERP_WAREHOUSE'
GROUP BY
  first_day_of_month
ORDER BY
  first_day_of_month DESC;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    DATE_TRUNC('MONTH', START_TIME) AS first_day_of_month,
    SUM(credits_used) as total_credits
FROM
    SNOWFLAKE.ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY
WHERE
    WAREHOUSE_NAME = 'RELATIONAL_AI_ERP_WAREHOUSE'
GROUP BY
    first_day_of_month
ORDER BY
    first_day_of_month DESC;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.warehouse.month")
    print(rows)

Jobs performed by all reasoners in the past month

Requires the eng_user application role.

To help understand which queries or other jobs are consuming the most credits, you could look at which jobs took the longest to complete. Although duration alone does not determine a job’s cost—other factors such as the configured reasoner size also affect cost—you can use this query to identify jobs that may be responsible for high costs. You can then investigate these jobs further to determine if they can be optimized.

Use the following query to see all jobs performed by all reasoners in the past month sorted by duration in descending order:

SQL

SELECT
  *
FROM
    relationalai.api.jobs
WHERE
  created_on >= DATEADD(month, -1, CURRENT_TIMESTAMP())
ORDER BY
  duration DESC;

Python

from relationalai.client import connect_sync

SQL = """
SELECT
    *
FROM
    relationalai.api.jobs
WHERE
    created_on >= DATEADD(month, -1, CURRENT_TIMESTAMP())
ORDER BY
    duration DESC;
"""

with connect_sync() as client:
    sql = client.core.sql_executor
    assert sql is not None

    rows = sql.collect(SQL, operation="costs.jobs.by_duration")
    print(rows)