Autocapture inference logs for realtime inference¶

Use Auto Capture in Snowflake ML to automatically log every request and response processed by a model service. Auto Capture provides immediate visibility into the request successes, request failures, and the inputs behind unexpected predictions.

Instead of piping request or response data into a table or view, you can automatically persist inference request and response data. Instead of needing to correctly create pipelines for data ingestion and monitoring, you can use Auto Capture.

With Auto Capture, you can do the following:

• Rapidly debug: Analyze historical inference data to diagnose edge cases and understand model behavior.

• Continuously improve your models: Use real-world production data to create high-quality datasets to train new models.

• Test: Use the data collected from the logs for A/B testing and shadow testing.

For each inference request, Auto Capture logs the following:

• Request payload

• Response payload

• Model version identifier

• Service identifier

• Gateway routing metadata

• Request/response timestamps

• Response code (such as 200).

This data is read-only and cannot be modified by users.

Snowflake only captures response data for successful requests. If a request fails, Snowflake doesn’t capture any data.

Prerequisites and model version compatibility¶

The following sections describe the prerequisites and model version compatibility for Auto Capture.

CREATE [ OR REPLACE ] MODEL [ IF NOT EXISTS ] <name> [ WITH VERSION <version_name> ]
FROM MODEL <source_model_name> [ VERSION <source_version_or_alias_name> ]

Activate Auto Capture¶

After you’ve created a new model or cloned an existing model, enable Auto Capture for the model service using the Python SDK. For more information about the model service, see Model Serving in Snowpark Container Services.

Use the following Python code to enable Auto Capture:

mv.create_service(
    service_name="my_service",
    service_compute_pool="my_compute_pool",
    autocapture=True
)

The mv variable is the model version object. You defined it when you logged the model to the model registry.

The default value for autocapture is False. Make sure you’re enabling autocapture for a model that you’ve created after January 23, 2026 and logged to the model registry. Otherwise, the service creation fails because the model doesn’t have an inference table.

Query inference data¶

To access your logs, use the INFERENCE_TABLE function. This function returns inference logs for a model and supports filtering by version, service, and gateway. Only model owners are able to see the data when they have USAGE privileges on the gateway and service.

-- Fetch all inference logs for a specific model
SELECT * FROM TABLE(INFERENCE_TABLE('my_model'));

SELECT * FROM TABLE(
INFERENCE_TABLE(
'MY_MODEL',
VERSION => 'V1',
SERVICE => 'MY_PREDICTION_SERVICE',
GATEWAY => 'MY_GATEWAY'
)
);

WHERE RECORD_ATTRIBUTES:"snow.model_serving.function.name" = 'predict'

SELECT *
FROM TABLE(
  INFERENCE_TABLE('my_model')
);

SELECT *
FROM TABLE(
  INFERENCE_TABLE(
    'my_model',
    MODEL_VERSION => 'v1'
  )
);

SELECT *
FROM TABLE(
  INFERENCE_TABLE(
    'my_model',
    MODEL_VERSION => 'v1',
    SERVICE => 'my_service'
  )
);

SELECT *
FROM TABLE(
  INFERENCE_TABLE(
    'my_model',
    MODEL_VERSION => 'v1',
    GATEWAY => 'my_gateway'
  )
);

Data schema and metadata¶

Snowflake only captures response data for successful requests. If a request fails, Snowflake doesn’t capture any data.

The following are the record attributes that are captured:

Data truncation logic¶

To maintain system performance, there’s a 1 MB limit for each inference event. If the request and response reaches the limit, Snowflake applies a multi-stage truncation process to preserve as much utility as possible.

The following table shows the truncation process:

*Stage 3 occurs if metadata alone exceeds the threshold after content reduction.

Limits¶

Keep in mind the following limitations and considerations when using auto capture:

• LLM Support: Auto capture isn’t supported for Large Language Models (LLMs).

• Throughput: Auto capture is designed for a system throughput of approximately 300-400 requests per second (or 10MB/s) per service.

• Replication: You can’t replicate inference tables. Replicated models will have no inference tables in the target account.

• Retention: Inference data persists even if the Service or Gateway is deleted.

• Warning: Deleting the Model object will permanently delete all associated inference data.

• Ground Truth: To perform drift analysis, maintain a separate ground truth table and join it with the INFERENCE_TABLE output using common request IDs.

• Consumer Accounts: Consumer accounts can’t create a service with autocapture enabled for shared models with inference tables.

• Performance: Autocapture is designed to not add latency to inference requests. However, it may drop some captures during periods of extremely high request volume.

Schema¶

As part of this feature, the following values are added to the respective columns.