databricks-cost-tuning（数据块成本调整）

Databricks Cost Tuning

Overview

Reduce Databricks spending through cluster policies, spot instances, SQL warehouse right-sizing, and cost governance. Databricks charges per DBU (Databricks Unit) with rates varying by compute type: Jobs Compute ($0.15/DBU), All-Purpose Compute ($0.40/DBU), SQL Compute ($0.22/DBU), Serverless ($0.07/DBU). System tables (system.billing.usage and system.billing.list_prices) provide cost visibility.

Prerequisites

• Databricks Premium or Enterprise workspace
• Access to system.billing.usage and system.billing.list_prices tables
• Workspace admin for cluster policy creation

Instructions

Step 1: Identify Top Cost Drivers

-- Top 10 most expensive resources this month
SELECT cluster_id,
       COALESCE(usage_metadata.cluster_name, 'unnamed') AS cluster_name,
       sku_name,
       SUM(usage_quantity) AS total_dbus,
       ROUND(SUM(usage_quantity * p.pricing.default), 2) AS estimated_cost_usd
FROM system.billing.usage u
LEFT JOIN system.billing.list_prices p ON u.sku_name = p.sku_name
WHERE u.usage_date >= date_trunc('month', current_date())
GROUP BY cluster_id, cluster_name, u.sku_name
ORDER BY estimated_cost_usd DESC
LIMIT 10;

-- Cost by team (requires cluster tags)
SELECT usage_metadata.cluster_tags.Team AS team,
       sku_name,
       ROUND(SUM(usage_quantity), 1) AS total_dbus,
       ROUND(SUM(usage_quantity * p.pricing.default), 2) AS cost_usd
FROM system.billing.usage u
LEFT JOIN system.billing.list_prices p ON u.sku_name = p.sku_name
WHERE u.usage_date >= date_trunc('month', current_date())
GROUP BY team, u.sku_name
ORDER BY cost_usd DESC;

Step 2: Enforce Cluster Policies

Cluster policies restrict what users can configure, preventing runaway costs.

from databricks.sdk import WorkspaceClient

w = WorkspaceClient()

# Create a cost-optimized policy for interactive clusters
policy = w.cluster_policies.create(
    name="cost-optimized-interactive",
    definition="""{
        "autotermination_minutes": {
            "type": "range", "minValue": 10, "maxValue": 60, "defaultValue": 20
        },
        "num_workers": {
            "type": "range", "minValue": 0, "maxValue": 8
        },
        "node_type_id": {
            "type": "allowlist",
            "values": ["m5.xlarge", "m5.2xlarge", "c5.xlarge", "c5.2xlarge"],
            "defaultValue": "m5.xlarge"
        },
        "aws_attributes.availability": {
            "type": "fixed", "value": "SPOT_WITH_FALLBACK"
        },
        "custom_tags.CostCenter": {
            "type": "fixed", "value": "engineering"
        }
    }""",
)

# Assign policy to a group
w.cluster_policies.set_permissions(
    cluster_policy_id=policy.policy_id,
    access_control_list=[{
        "group_name": "data-analysts",
        "all_permissions": [{"permission_level": "CAN_USE"}],
    }],
)

Step 3: Spot Instances for Batch Jobs

Spot instances save 60-90% on worker nodes. Always keep the driver on-demand.

# Job cluster config with spot instances
job_cluster_config = {
    "spark_version": "14.3.x-scala2.12",
    "node_type_id": "i3.xlarge",
    "num_workers": 4,
    "aws_attributes": {
        "availability": "SPOT_WITH_FALLBACK",
        "first_on_demand": 1,          # Driver is on-demand
        "spot_bid_price_percent": 100,  # Pay up to on-demand price
        "zone_id": "auto",             # Let Databricks pick cheapest AZ
    },
}
# Workers use spot, driver uses on-demand
# If spot instances unavailable, falls back to on-demand automatically

Step 4: Right-Size SQL Warehouses

-- Check warehouse utilization
SELECT warehouse_id, warehouse_name,
       COUNT(*) AS query_count,
       ROUND(AVG(total_duration_ms) / 1000, 1) AS avg_duration_sec,
       ROUND(MAX(queue_duration_ms) / 1000, 1) AS max_queue_sec,
       ROUND(AVG(queue_duration_ms) / 1000, 1) AS avg_queue_sec
FROM system.query.history
WHERE start_time > current_timestamp() - INTERVAL 7 DAYS
GROUP BY warehouse_id, warehouse_name;

-- If avg_queue_sec is near 0 → warehouse is oversized, reduce cluster_size
-- If avg_queue_sec > 30 → warehouse needs more capacity or auto-scaling

# Migrate to serverless for bursty workloads (cheaper per-second billing)
for wh in w.warehouses.list():
    print(f"{wh.name}: type={wh.warehouse_type}, size={wh.cluster_size}, "
          f"auto_stop={wh.auto_stop_mins}min, state={wh.state}")
    # Serverless (~$0.07/DBU) vs Classic (~$0.22/DBU) for bursty workloads

Step 5: Auto-Terminate Idle Development Clusters

# Find and set aggressive auto-termination on dev clusters
databricks clusters list --output JSON | \
  jq -r '.[] | select(.cluster_name | test("dev|sandbox|test")) | .cluster_id' | \
  while read CID; do
    echo "Setting 15min auto-stop on cluster $CID"
    databricks clusters edit --cluster-id "$CID" --json '{"autotermination_minutes": 15}'
  done

# Find idle running clusters wasting money
from datetime import datetime, timedelta

for c in w.clusters.list():
    if c.state.value == "RUNNING" and c.last_activity_time:
        idle_minutes = (datetime.now().timestamp() * 1000 - c.last_activity_time) / 60000
        if idle_minutes > 60:
            print(f"IDLE {idle_minutes:.0f}min: {c.cluster_name} "
                  f"({c.num_workers} x {c.node_type_id})")

Step 6: Instance Pools for Faster + Cheaper Startup

# Create a pool of pre-allocated instances
pool = w.instance_pools.create(
    instance_pool_name="etl-pool",
    node_type_id="i3.xlarge",
    min_idle_instances=2,     # Keep 2 warm for instant startup
    max_capacity=10,
    idle_instance_autotermination_minutes=15,
    aws_attributes={"availability": "SPOT_WITH_FALLBACK"},
)

# Reference in job cluster config
job_cluster = {
    "instance_pool_id": pool.instance_pool_id,
    "num_workers": 4,
    # No node_type_id needed — inherited from pool
}

Output

• Cost breakdown by cluster, team, and SKU
• Cluster policies enforcing auto-termination and instance limits
• Spot instance config for 60-90% savings on batch workers
• SQL warehouse utilization report for right-sizing
• Instance pools for faster cluster startup
• Idle cluster detection script

Error Handling

Examples

Monthly Cost Report

SELECT date_trunc('week', usage_date) AS week,
       sku_name,
       ROUND(SUM(usage_quantity), 0) AS total_dbus
FROM system.billing.usage
WHERE usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY week, sku_name
ORDER BY week, total_dbus DESC;

Cost Savings Checklist

• Auto-termination enabled on ALL interactive clusters (15-30 min)
• Spot instances enabled for all batch job workers
• Instance pools for frequently-used cluster configs
• Serverless SQL warehouses for bursty query workloads
• Cluster policies assigned to all non-admin groups
• Team tags on all clusters for cost attribution
• Weekly cost review using system.billing.usage

Resources

• Cost Optimization Best Practices
• Cluster Policies
• Billing Usage Tables
• Serverless SQL Warehouses