Compute system tables reference

Preview

This system table is in Public Preview. To access the table, the schema must be enabled in your system catalog. For more information, see Enable system table schemas.

This article provides you with a reference guide for the compute system tables. You can use these tables to monitor the activity and metrics of all-purpose and jobs compute in your account:

  • clusters: Records compute configurations in your account.

  • node_types: Includes a single record for each of the currently available node types, including hardware information.

  • node_timeline: Includes minute-by-minute records of your compute’s utilization metrics.

Cluster table schema

Note

This table is only available in us-east1 and asia-southeast1.

The cluster table is a slow-changing dimension table that contains the full history of compute configurations over time for all-purpose and jobs compute.

The clusters system table is located at system.compute.clusters and has the following schema:

Column name

Data type

Description

Example

account_id

string

ID of the account where this cluster was created.

23e22ba4-87b9- 4cc2-9770-d10b894b7118

workspace_id

string

ID of the workspace where this cluster was created.

1234567890123456

cluster_id

string

ID of the cluster for which this record is associated.

0000-123456-xxxxxxxx

cluster_name

string

User defined name for the cluster.

My cluster

owned_by

string

Username of the cluster owner. Defaults to the cluster creator, but can be changed through the Clusters API.

sample_user@email.com

create_time

timestamp

Timestamp of the change to this compute definition.

2023-01-09 11:00:00.000

delete_time

timestamp

Timestamp of when the cluster was deleted. The value is null if the cluster is not deleted.

2023-01-09 11:00:00.000

driver_node_type

string

Driver node type name. This matches the instance type name from the cloud provider.

n2-highmem-8

worker_node_type

string

Worker node type name. This matches the instance type name from the cloud provider.

n2-highmem-4

worker_count

bigint

Number of workers. Defined for fixed-size clusters only.

4

min_autoscale_workers

bigint

The set minimum number of workers. This field is valid only for autoscaling clusters.

1

max_autoscale_workers

bigint

The set maximum number of workers. This field is valid only for autoscaling clusters.

1

auto_termination_minutes

bigint

The configured autotermination duration.

120

enable_elastic_disk

boolean

Autoscaling disk enablement status.

true

tags

map

User-defined tags for the cluster (does not include default tags).

{"ResourceClass":"SingleNode"}

cluster_source

string

Indicates the creator for the cluster: UI, API, JOB, etc.

UI

init_scripts

array

Set of paths for init scripts.

"/Users/example@email.com /files/scripts/install-python-pacakges.sh"

aws_attributes

struct

AWS specific settings. This field will be empty.

null

azure_attributes

struct

Azure specific settings. This field will be empty.

null

gcp_attributes

struct

GCP specific settings.

{ "first_on_demand": "0", "availability": "ON_DEMAND_GCP", "spot_bid_max_price": "—1" }

driver_instance_pool_id

string

Instance pool ID if the driver is configured on top of an instance pool.

1107-555555-crhod16-pool-DIdnjazB

worker_instance_pool_id

string

Instance Pool ID if the worker is configured on top of an instance pool.

1107-555555-crhod16-pool-DIdnjazB

dbr_version

string

The Databricks Runtime of the cluster.

14.x-snapshot-scala2.12

change_time

timestamp

Timestamp of change to the compute definition.

2023-01-09 11:00:00.000

change_date

date

Change date. Used for retention.

2023-01-09

Node types table schema

Note

This table is only available in us-east1 and asia-southeast1.

The node type table captures the currently available node types with their basic hardware information. The node type system table is located at system.compute.node_types and has the following schema:

Column name

Data type

Description

Example

account_id

string

ID of the account where this cluster was created.

23e22ba4-87b9-4cc2-9770-d10b894b7118

node_type

string

Unique identifier for node type.

<driver type>

core_count

double

Number of vCPUs for the instance.

48.0

memory_mb

long

Total memory for the instance.

393216

gpu_count

long

Number of GPUs for the instance.

0

Node timeline table schema

The node timeline table captures node-level resource utilization data at minute granularity. Each record contains data for a given minute of time per instance.

The node timeline system table is located at system.compute.node_timeline and has the following schema:

Column name

Data type

Description

Example

account_id

string

ID of the account where this compute resource is running.

23e22ba4-87b9-4cc2-9770-d10b894b7118

workspace_id

string

ID of the workspace where this compute resource is running.

1234567890123456

cluster_id

string

ID of the compute resource.

0000-123456-crmpt124

instance_id

string

ID for the specific instance.

i-1234a6c12a2681234

start_time

timestamp

Start time for the record in UTC.

2024-07-16T12:00:00Z

end_time

timestamp

End time for the record in UTC.

2024-07-16T13:00:00Z

driver

boolean

Whether the instance is a driver or worker node.

true

cpu_user_percent

double

Percentage of time the CPU spent in userland.

34.76163817234407

cpu_system_percent

double

Percentage of time the CPU spent in the kernel.

1.0895310279488264

cpu_wait_percent

double

Percentage of time the CPU spent waiting for I/O.

0.03445157400629276

mem_used_percent

double

Percentage of the compute’s memory that was used during the time period (including memory used by background processes running on the compute).

45.34858216779041

mem_swap_percent

double

Percentage of memory usage attributed to memory swap.

0.014648443087939

network_sent_bytes

bigint

The number of bytes sent out in network traffic.

517376

network_received_bytes

bigint

The number of received bytes from network traffic.

179234

disk_free_bytes_per_mount_point

map

The disk utilization grouped by mount point. This is ephemeral storage provisioned only while the compute is running.

{"/var/lib/lxc":123455551234,"/":

123456789123,"/local_disk0":123412341234}

node_type

string

The name of the node type. This will match the instance type name from the cloud provider.

<driver type>

Known limitations

  • Compute resources that were marked deleted before October 23, 2023 do not appear in the clusters table. This might result in joins from the system.billing.usage table not matching records in the clusters table. All active compute resources have been backfilled.

  • These tables only includes records for all-purpose and jobs compute. They do not contain records for serverless compute, Delta Live Tables compute, or SQL warehouses.

  • Nodes that ran for less than 10 minutes might not appear in the node_timeline table.

Sample queries

You can use the following sample queries to answer common questions:

Note

Some of these examples join the cluster table with the system.billing.usage table. Since billing records are cross-regional and cluster records region-sepcific, billing records only match cluster records for the region in which you are querying. To see records from another region, please execute the query in that region.

Join cluster records with the most recent billing records

This query can help you understand spending over time. Once you update the usage_start_time to the most current billing period, it grabs the most recent updates to the billing records to join into clusters data.

Each record is associated with the cluster owner during that particular run. So, if the cluster owner changes, costs will roll up to the correct owner based on when the cluster was used.

SELECT
  u.record_id,
  c.cluster_id,
  c.owned_by,
  c.change_time,
  u.usage_start_time,
  u.usage_quantity
FROM
  system.billing.usage u
  JOIN system.compute.clusters c
  JOIN (SELECT u.record_id, c.cluster_id, max(c.change_time) change_time
    FROM system.billing.usage u
    JOIN system.compute.clusters c
    WHERE
      u.usage_metadata.cluster_id is not null
      and u.usage_start_time >= '2023-01-01'
      and u.usage_metadata.cluster_id = c.cluster_id
      and date_trunc('HOUR', c.change_time) <= date_trunc('HOUR', u.usage_start_time)
    GROUP BY all) config
WHERE
  u.usage_metadata.cluster_id is not null
  and u.usage_start_time >= '2023-01-01'
  and u.usage_metadata.cluster_id = c.cluster_id
  and u.record_id = config.record_id
  and c.cluster_id = config.cluster_id
  and c.change_time = config.change_time
ORDER BY cluster_id, usage_start_time desc;

Attribute costs to the cluster owner

If you are looking to reduce compute costs, you can use this query to find out which cluster owners in your account are using the most DBUs.

SELECT
  u.record_id record_id,
  c.cluster_id cluster_id,
  max_by(c.owned_by, c.change_time) owned_by,
  max(c.change_time) change_time,
  any_value(u.usage_start_time) usage_start_time,
  any_value(u.usage_quantity) usage_quantity
FROM
  system.billing.usage u
  JOIN system.compute.clusters c
WHERE
  u.usage_metadata.cluster_id is not null
  and u.usage_start_time >= '2023-01-01'
  and u.usage_metadata.cluster_id = c.cluster_id
  and c.change_time <= u.usage_start_time
GROUP BY 1, 2
ORDER BY cluster_id, usage_start_time desc;

Identify the compute resources with the highest average utilization and peak utilization

Identify the all-purpose and jobs compute that have the highest average CPU utilization and the highest peak CPU utilization.

SELECT
        distinct cluster_id,
driver,
avg(cpu_user_percent + cpu_system_percent) as `Avg CPU Utilization`,
max(cpu_user_percent + cpu_system_percent) as `Peak CPU Utilization`,
        avg(cpu_wait_percent) as `Avg CPU Wait`,
        max(cpu_wait_percent) as `Max CPU Wait`,
        avg(mem_used_percent) as `Avg Memory Utilization`,
        max(mem_used_percent) as `Max Memory Utilization`,
avg(network_received_bytes)/(1024^2) as `Avg Network MB Received per Minute`,
avg(network_sent_bytes)/(1024^2) as `Avg Network MB Sent per Minute`
FROM
        node_timeline
WHERE
        start_time >= date_add(now(), -1)
GROUP BY
        cluster_id,
        driver
ORDER BY
        3 desc;