Trino admin cluster health

Trino Admin: Cluster Health

When to Use

• Cluster appears slow or queries are queuing unexpectedly
• Workers are reporting failures or dropping out of the cluster
• Memory pressure causing OOM query kills
• Setting up monitoring for a new Trino cluster
• Performing rolling upgrades or worker restarts

---

REST API Health Endpoints

Trino exposes a comprehensive REST API for monitoring without requiring a client connection.

COORDINATOR="http://trino-coordinator:8080"

# Cluster info: version, environment, uptime
curl -s $COORDINATOR/v1/info | jq .

# All registered nodes (coordinator + workers)
curl -s $COORDINATOR/v1/node | jq '.[].uri'

# Failed nodes (should be empty in healthy cluster)
curl -s $COORDINATOR/v1/node/failed | jq .

# Cluster-level stats: running queries, blocked queries, active workers
curl -s $COORDINATOR/v1/cluster | jq .

# All active queries with state
curl -s -H "X-Trino-User: admin" $COORDINATOR/v1/query | jq '.[] | {queryId, state, query}'

# Specific query details
curl -s -H "X-Trino-User: admin" $COORDINATOR/v1/query/<query_id> | jq .

# Kill a query
curl -X DELETE -H "X-Trino-User: admin" $COORDINATOR/v1/query/<query_id>

---

Key Health Metrics via REST

# Worker count and health
curl -s $COORDINATOR/v1/cluster | jq '{
  activeWorkers: .activeWorkers,
  runningQueries: .runningQueries,
  queuedQueries:  .queuedQueries,
  blockedQueries: .blockedQueries
}'

# Check if coordinator is ready to serve requests
curl -s $COORDINATOR/v1/info/state
# → "ACTIVE" = ready, "SHUTTING_DOWN" = graceful shutdown in progress

---

JMX MBeans for Prometheus

Configure Prometheus JMX exporter in jmx_exporter.yaml:

# prometheus-jmx-exporter.yaml
lowercaseOutputName: true
lowercaseOutputLabelNames: true
rules:
  # Running and queued queries
  - pattern: 'trino.execution<name=QueryManager><>RunningQueries'
    name: trino_running_queries
    type: GAUGE

  - pattern: 'trino.execution<name=QueryManager><>(StartedQueries|FailedQueries)\.FiveMinute\.Count'
    name: trino_queries_$1_5m
    type: GAUGE

  - pattern: 'trino.execution<name=QueryManager><>ExecutionTime\.FiveMinutes\.(P50|P90|P99)'
    name: trino_query_execution_time_$1_ms
    type: GAUGE

  # Memory
  - pattern: 'trino.memory<type=ClusterMemoryPool, name=general><>FreeDistributedBytes'
    name: trino_free_memory_bytes
    type: GAUGE

  - pattern: 'trino.memory<name=ClusterMemoryManager><>QueriesKilledDueToOutOfMemory'
    name: trino_queries_killed_oom_total
    type: COUNTER

  # Workers
  - pattern: 'trino.failuredetector<name=HeartbeatFailureDetector><>ActiveCount'
    name: trino_active_workers
    type: GAUGE

  # Task throughput
  - pattern: 'trino.execution<name=SqlTaskManager><>InputDataSize\.FiveMinute\.Count'
    name: trino_input_data_rate_5m_bytes
    type: GAUGE

JVM args to enable JMX exporter:

# etc/jvm.config (add to existing config)
-javaagent:/opt/jmx_exporter/jmx_prometheus_javaagent.jar=9090:/opt/jmx_exporter/prometheus-jmx-exporter.yaml

---

Prometheus Alert Rules

# trino-alerts.yaml
groups:
  - name: trino
    rules:
      - alert: TrinoWorkersLow
        expr: trino_active_workers < 3
        for: 2m
        labels:
          severity: critical
        annotations:
          summary: "Trino has {{ $value }} active workers (threshold: 3)"

      - alert: TrinoQueryQueueHigh
        expr: trino_running_queries > 80
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "Trino query queue depth {{ $value }} > 80"

      - alert: TrinoOOMKills
        expr: increase(trino_queries_killed_oom_total[5m]) > 0
        labels:
          severity: warning
        annotations:
          summary: "Trino killed {{ $value }} queries due to OOM in last 5m"

      - alert: TrinoHighQueryFailureRate
        expr: |
          rate(trino_queries_FailedQueries_5m[5m]) /
          rate(trino_queries_StartedQueries_5m[5m]) > 0.1
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "Trino failure rate {{ $value | humanizePercentage }}"

      - alert: TrinoFreeMemoryLow
        expr: trino_free_memory_bytes < 5 * 1024 * 1024 * 1024  # < 5GB free
        for: 5m
        labels:
          severity: warning
        annotations:
          summary: "Trino cluster free memory {{ $value | humanize1024 }}B"

---

Query State Diagnosis

# Find all BLOCKED queries (may indicate memory/resource pressure)
curl -s -H "X-Trino-User: admin" $COORDINATOR/v1/query \
  | jq '.[] | select(.state == "BLOCKED") | {queryId, state, queryTextPreview}'

# Find long-running queries (running > 10 minutes)
curl -s -H "X-Trino-User: admin" $COORDINATOR/v1/query \
  | jq --argjson threshold 600000 \
    '.[] | select(.state == "RUNNING") | select(.elapsedTime.toMillis // 0 > $threshold)'

# Count queries by state
curl -s -H "X-Trino-User: admin" $COORDINATOR/v1/query \
  | jq 'group_by(.state) | map({state: .[0].state, count: length})'

Query states reference:

State	Normal?	Action if Stuck
QUEUED	Briefly	Check resource group limits, increase maxConcurrentQueries
PLANNING	Briefly	May indicate stale stats or very complex query
RUNNING	Expected	Monitor progress via Web UI
BLOCKED	Briefly	Persistent = memory pressure, investigate spill or kill
FINISHING	Briefly	Normal finalisation
FAILED	No	Check query error in /v1/query/<id>

---

Log Monitoring

# Trino server log location
tail -f /var/trino/data/var/log/server.log

# Key patterns to alert on:
grep -E "WORKER_LOST|NodeManager|OutOfMemoryError|exceeded memory limit" server.log

# Worker registration events
grep "NodeManager" server.log | tail -50

# OOM query kills
grep "exceeded memory limit" server.log | tail -20

Key log patterns:

Pattern	Meaning	Action
Query exceeded memory limit	OOM kill	Increase query.max-memory or add spill
WORKER_LOST	Worker crashed/disconnected	Check worker logs, heap dumps
Too many tasks active	Worker overloaded	Scale out workers
Connection refused	Discovery server unreachable	Check coordinator health

---

Graceful Worker Shutdown (Rolling Upgrade)

# 1. Put worker in graceful shutdown mode (new queries won't be scheduled)
curl -X PUT -H "Content-Type: application/json" \
  http://worker-host:8080/v1/info/state \
  -d '"SHUTTING_DOWN"'

# 2. Wait for all tasks to finish (poll until running tasks = 0)
watch -n 5 'curl -s http://worker-host:8080/v1/info/state'

# 3. Once state = INACTIVE, safely stop the process
bin/launcher stop

Configure shutdown grace period:

# etc/config.properties
shutdown.grace-period=2m

---

Worker Node Checklist

For each worker node, verify:

# JVM heap usage (should be < 80% of -Xmx)
curl -s http://worker:8080/v1/jmx/mbean/java.lang:type=Memory \
  | jq '.attributes[] | select(.name == "HeapMemoryUsage")'

# Number of active tasks
curl -s http://worker:8080/v1/task | jq 'length'

# Worker is registered with coordinator
curl -s http://coordinator:8080/v1/node | jq '.[].uri' | grep worker-host

---

Web UI Key Views

Navigate to http://trino-coordinator:8080/ui/:

View	Path	What to Check
Query list	/ui/	QUEUED count, BLOCKED count, recent FAILED
Query details	/ui/query/<id>	Stage breakdown, task timings, CPU vs wall time
Worker status	/ui/worker/<id>	Heap used, tasks running, thread count
Resource groups	/ui/ (filter)	Concurrency limits, queue depth per group

Performance signals in Web UI:

• High "Blocked" percentage = memory or I/O bottleneck
• Many QUEUED queries = resource group concurrency limit hit
• wall_time >> cpu_time = waiting on network/I/O, not compute

---

Anti-Patterns

1. Running coordinator with node-scheduler.include-coordinator=true in production — coordinator becomes a bottleneck for both planning and execution; always set false for clusters with > 3 workers.
2. Not monitoring OOM kills — QueriesKilledDueToOutOfMemory counter silently increments; queries fail with cryptic errors; alert on this metric.
3. Ignoring BLOCKED query state — persistent BLOCKED queries hold memory slots and prevent other queries from running; set alert if any query is BLOCKED > 10 minutes.
4. Hard shutting down workers during query execution — kills in-flight tasks; always use graceful shutdown via /v1/info/state PUT.
5. Exposing JMX without authentication in production — JMX can expose sensitive cluster information; bind to localhost only or use TLS.

---

References

• Admin REST API: trino.io/docs/current/develop/client-protocol.html
• Web UI: trino.io/docs/current/admin/web-interface.html
• JMX connector: trino.io/docs/current/connector/jmx.html
• Related skills: [[trino-memory-and-spill-tuning]], [[trino-resource-group-governance]], [[trino-observability-platform]], [[trino-production-readiness-review]]