Configuration

Training compute mode

Training runs on CPU by default. GPU is an explicit opt-in that requires GPU availability in the consumer’s region.

Switch to GPU:

CALL veil.core.set_training_compute('gpu');
CALL veil.core.start_services();

Switch back to CPU:

CALL veil.core.set_training_compute('cpu');
CALL veil.core.start_services();

No active training job can be running during a mode switch.

GPU mode selection uses SHOW COMPUTE POOL INSTANCE FAMILIES at runtime and matches against tested families. If no tested family is available in the consumer’s region, GPU mode is rejected with a clear error.

Compute sizing

The encoding (inference) pool and the CPU training pool each run on an instance family that an administrator can change. The default for both is CPU_X64_XS. Supported families: CPU_X64_XS, CPU_X64_S, CPU_X64_M, CPU_X64_SL, CPU_X64_L, HIGHMEM_X64_S, HIGHMEM_X64_M.

Set the encoding pool size:

CALL veil.core.set_encoding_instance('CPU_X64_M');
CALL veil.core.start_services();

Set the CPU training pool size:

CALL veil.core.set_training_instance('CPU_X64_M');
CALL veil.core.start_services();

The choice persists across restarts. start_services checks the family against those available in the consumer’s region before it changes anything, then recreates the pool on the selected family. Applying a size change restarts services, so run it during a maintenance window and not while a training job is active.

The training size applies to CPU training mode. In GPU mode, training runs on a GPU family selected automatically.

Training parameters

Parameter	Type	Default	Description
encoder_name	VARCHAR	required	Name for the encoder
source_table	VARCHAR	required	Fully-qualified table name
feature_columns	ARRAY	required	Columns to encode
latent_dims	INTEGER	16	Output vector dimension
epochs	INTEGER	100	Training epochs
batch_size	INTEGER	256	Batch size
passthrough_columns	ARRAY	[]	Columns returned alongside encoded vectors (not encoded)

Encoder names ending in _v<number> are rejected to avoid collisions with the archive naming convention.

Scheduled training

Schedule periodic retraining:

CALL veil.schedule_training(
    'sales',
    'my_db.my_schema.sales',
    ARRAY_CONSTRUCT('price', 'sqft', 'bedrooms', 'bathrooms'),
    16, 100, 256,
    'USING CRON 0 2 * * 0 America/Los_Angeles'
);

Inference service

The inference service runs on a CPU pool. The default family is CPU_X64_XS, available on all cloud providers and regions; an administrator can select a larger family with set_encoding_instance (see Compute sizing). The service supports up to 3 instances for horizontal scaling.

Billing

Usage-based billing via Snowflake Custom Event Billing. A single billing dimension: data processed through encoders (training and inference), metered per GB.

View billing events:

SELECT * FROM veil.billing_events_v;