Trino Resource Group Governance

Trino Resource Group Governance

When to Use

• Multiple teams share one Trino cluster and need query isolation
• Some workloads (ETL pipelines) should not crowd out interactive (BI/adhoc) queries
• Enforcing memory limits or concurrency caps per team or query type
• Setting SLA tiers for critical vs non-critical workloads
• Preventing a single user from monopolizing cluster resources

Configuration Files

# etc/config.properties — point to resource groups config file
resource-groups.config-file=etc/resource-groups.json

# Optional: use database for dynamic reconfiguration (reloads every ~1s)
# resource-groups.config-file=etc/resource-groups.json    # file-based

Complete Multi-Tenant Resource Groups Config

{
  "rootGroups": [
    {
      "name": "global",
      "softMemoryLimit": "80%",
      "hardConcurrencyLimit": 200,
      "maxQueued": 500,
      "schedulingPolicy": "weighted",
      "subGroups": [

        {
          "name": "admin",
          "softMemoryLimit": "100%",
          "hardConcurrencyLimit": 20,
          "maxQueued": 10,
          "schedulingPolicy": "query_priority",
          "schedulingWeight": 10,
          "jmxExport": true
        },

        {
          "name": "pipeline",
          "softMemoryLimit": "60%",
          "hardConcurrencyLimit": 60,
          "maxQueued": 200,
          "schedulingPolicy": "weighted_fair",
          "schedulingWeight": 3,
          "softCpuLimit": "2h",
          "hardCpuLimit": "3h",
          "jmxExport": true,
          "subGroups": [
            {
              "name": "pipeline_${USER}",
              "softMemoryLimit": "20%",
              "hardConcurrencyLimit": 8,
              "maxQueued": 50,
              "schedulingPolicy": "fair"
            }
          ]
        },

        {
          "name": "adhoc",
          "softMemoryLimit": "30%",
          "hardConcurrencyLimit": 80,
          "maxQueued": 300,
          "schedulingPolicy": "weighted_fair",
          "schedulingWeight": 2,
          "jmxExport": true,
          "subGroups": [
            {
              "name": "adhoc_${USER}",
              "softMemoryLimit": "10%",
              "hardConcurrencyLimit": 3,
              "maxQueued": 20,
              "schedulingPolicy": "fair"
            }
          ]
        },

        {
          "name": "batch",
          "softMemoryLimit": "50%",
          "hardConcurrencyLimit": 20,
          "maxQueued": 100,
          "schedulingPolicy": "fair",
          "schedulingWeight": 1,
          "softCpuLimit": "4h",
          "hardCpuLimit": "8h",
          "jmxExport": true
        },

        {
          "name": "reporting",
          "softMemoryLimit": "40%",
          "hardConcurrencyLimit": 30,
          "maxQueued": 100,
          "schedulingPolicy": "weighted_fair",
          "schedulingWeight": 2,
          "jmxExport": true,
          "subGroups": [
            {
              "name": "reporting_${USER}",
              "softMemoryLimit": "15%",
              "hardConcurrencyLimit": 5,
              "maxQueued": 20
            }
          ]
        }

      ]
    }
  ],

  "selectors": [
    {
      "user": "admin",
      "group": "global.admin"
    },
    {
      "source": ".*airflow.*",
      "queryType": "DATA_DEFINITION",
      "group": "global.batch"
    },
    {
      "source": ".*airflow.*",
      "group": "global.pipeline.pipeline_${USER}"
    },
    {
      "source": ".*dbt.*",
      "group": "global.pipeline.pipeline_${USER}"
    },
    {
      "source": ".*superset.*",
      "group": "global.reporting.reporting_${USER}"
    },
    {
      "clientTags": ["batch"],
      "group": "global.batch"
    },
    {
      "clientTags": ["adhoc"],
      "group": "global.adhoc.adhoc_${USER}"
    },
    {
      "user": ".*",
      "group": "global.adhoc.adhoc_${USER}"
    }
  ],

  "cpuQuotaPeriod": "1h"
}

Resource Group Properties Reference

Scheduling Policies Explained

-- Set query priority (used by query_priority policy)
SET SESSION query_priority = 10;   -- higher = more important

Selector Rules

Selectors match queries to resource groups. All conditions in a selector use AND logic. First matching selector wins.

"selectors": [
  {
    "user": "etl_service",                     -- exact user match
    "group": "global.pipeline.pipeline_etl_service"
  },
  {
    "source": ".*dbt-cloud.*",                 -- regex on source string
    "queryType": "SELECT",
    "group": "global.adhoc.adhoc_${USER}"
  },
  {
    "clientTags": ["high-priority", "etl"],   -- ALL tags must be present
    "group": "global.pipeline.pipeline_${USER}"
  },
  {
    "queryText": ".*iceberg\\.gold\\..*",      -- regex on SQL text
    "group": "global.reporting.reporting_${USER}"
  }
]

Query type values: SELECT, INSERT, DELETE, UPDATE, ANALYZE, DATA_DEFINITION, EXPLAIN.

Setting client tags from JDBC:

Properties props = new Properties();
props.setProperty("user", "my_user");
props.setProperty("clientTags", "high-priority,etl");   // comma-separated
Connection conn = DriverManager.getConnection("jdbc:trino://coordinator:8080", props);

Setting source via CLI:

trino --server http://coordinator:8080 \
      --user dbt_user \
      --source "dbt-cloud-production"

CPU Quota Enforcement

CPU quotas throttle groups that exceed CPU time budgets per period:

{
  "cpuQuotaPeriod": "1h",
  "subGroups": [
    {
      "name": "batch",
      "softCpuLimit": "30m",     // start queuing above 30 min CPU/hour
      "hardCpuLimit": "60m"      // hard stop at 60 min CPU/hour
    }
  ]
}

• Queries above softCpuLimit are deprioritized (other groups get preference)
• Above hardCpuLimit, queries queue until next cpuQuotaPeriod resets

Database-Backed Configuration (Dynamic)

For dynamic reconfiguration without restarts (Starburst Enterprise / large deployments):

# etc/config.properties
resource-groups.config-file=etc/resource-groups.json

# For database-backed (reloads every ~1s from DB)
# Supported: MySQL, PostgreSQL, Oracle
resource-groups.config-db.url=jdbc:postgresql://db:5432/trinoconfig
resource-groups.config-db.user=trino_admin
resource-groups.config-db.password=${ENV:DB_PASSWORD}

Tables required: resource_groups, selectors, exact_match_source_selectors.

Monitoring Resource Groups

# Active group state via JMX (groups with jmxExport: true)
curl -s http://coordinator:8080/v1/jmx/mbean/trino.execution:name=QueryManager \
  | jq .

# Query resource group assignment
curl -s http://coordinator:8080/v1/query/<query_id> \
  | jq '.resourceGroupId'

-- Check current group utilization via JMX connector
SELECT node_id, name, value
FROM jmx.current."trino.execution:name=QueryManager"
WHERE name LIKE '%Running%' OR name LIKE '%Queued%';

Typical Isolation Patterns

Pattern 1: ETL vs Interactive

"subGroups": [
  {"name": "etl",         "hardConcurrencyLimit": 20, "schedulingWeight": 1},
  {"name": "interactive", "hardConcurrencyLimit": 50, "schedulingWeight": 5}
]

Interactive queries get 5× more scheduler priority than ETL.

Pattern 2: Per-Team Isolation

"subGroups": [
  {"name": "team_orders",   "softMemoryLimit": "20%", "hardConcurrencyLimit": 15},
  {"name": "team_marketing","softMemoryLimit": "15%", "hardConcurrencyLimit": 10},
  {"name": "team_finance",  "softMemoryLimit": "20%", "hardConcurrencyLimit": 15}
]

Each team has a guaranteed memory budget and max concurrency.

Pattern 3: Rate Limiting a User

{
  "name": "adhoc_${USER}",
  "hardConcurrencyLimit": 3,       -- user can run at most 3 queries simultaneously
  "maxQueued": 10,                  -- max 10 waiting in queue before rejection
  "softMemoryLimit": "5%",          -- user queues if consuming > 5% cluster memory
  "hardPhysicalDataScanLimit": "1TB"  -- reject queries scanning > 1TB
}

Anti-Patterns

1. Single flat resource group (no hierarchy) — all queries compete equally; ETL jobs crowd out interactive queries; always build at least two tiers.
2. hardConcurrencyLimit set too high — allows more queries than workers can handle, causing memory pressure; set based on worker count × 5 (rule of thumb starting point).
3. Missing catch-all selector — queries that don't match any selector fail with a routing error; always add a "user": ".*" catch-all at the end.
4. Not using schedulingWeight — with weighted policy and all weights = 1, groups compete equally regardless of SLA tier; high-priority groups must have higher weight.
5. Exporting all groups via JMX — high cardinality from ${USER} groups creates thousands of JMX beans; only export named tier-level groups.

References

• Resource groups docs: trino.io/docs/current/admin/resource-groups.html
• Starburst resource groups: docs.starburst.io/latest/admin/resource-groups.html
• Related skills: [[trino-admin-cluster-health]], [[trino-memory-and-spill-tuning]], [[trino-production-readiness-review]]