Quotas and Limits | NVIDIA Structured Data and Graph Models

Data Scale Limits

The following table lists the maximum data scale limits for Kumo deployments.

Your specific environment may be configured with lower limits, depending on your Kumo account. To request a limit increase, contact your Kumo point of contact.

		Default
1	Number of rows in a table	15 billion
2	Number of columns in a table	2,048
3	Table size	2 TB
4	Row size	10 KB
5	Number of tables in a graph	12
6	Number of rows in a graph	30 billion
7	Graph size	2 TB
8	Number of edges in a graph	30 billion
9	Parallel job execution	10

“Number of edges in a graph” denotes twice the total number of matching (primary key, foreign key) pairs.

Graph limits are summed across all tables in the graph.

All limits are admin-configurable through environment variables (MAX_TABLE_ROWS, MAX_TABLE_COLS, MAX_TABLE_SIZE_BYTES, MAX_GRAPH_TOTAL_ROWS, MAX_GRAPH_TOTAL_SIZE_BYTES, MAX_GRAPH_TOTAL_EDGES, MAX_TABLES_IN_A_GRAPH, ML_MAX_CONCURRENT_JOBS_TOTAL). For Helm deployments, pass these through kumoConfig.extra in values.yaml. For Docker deployments, set them as environment variables in your Docker Compose configuration.

Queueing of Jobs and Parallel Execution

Kumo allows queueing of jobs, which start one by one as ongoing jobs complete. By default, up to 10 jobs (training or batch prediction) may be queued. The queue length is admin-configurable through the ML_JOB_QUEUE_SIZE environment variable.
Jobs can also be executed in parallel in environments where this feature is enabled. The number of parallel jobs is admin-configurable through ML_MAX_CONCURRENT_JOBS_TOTAL.