Airflow Observability

When to Use

• Setting up visibility into Airflow scheduler and task performance
• Building Grafana dashboards for a data platform overview
• Diagnosing why tasks are slow or queuing
• Alerting on SLA misses and error spikes
• Tracing a task failure through its log chain

---

Metrics Collection Setup

StatsD + Prometheus

# airflow.cfg
[metrics]
statsd_on = True
statsd_host = statsd-exporter.monitoring.svc.cluster.local
statsd_port = 8125
statsd_prefix = airflow
statsd_allow_list = scheduler,executor,pool,dag,task  # allowlist prefixes

# statsd-exporter mapping (converts StatsD to Prometheus labels)
# statsd_exporter_config.yaml
mappings:
  # dag.etl_orders.task.load_orders.duration
  - match: "airflow.dag.*.task.*.duration"
    name: "airflow_task_duration_seconds"
    labels:
      dag_id: "$1"
      task_id: "$2"

  # dag.etl_orders.task.load_orders.failures
  - match: "airflow.dag.*.task.*.failures"
    name: "airflow_task_failures_total"
    labels:
      dag_id: "$1"
      task_id: "$2"

  # pool.open_slots.default_pool
  - match: "airflow.pool.open_slots.*"
    name: "airflow_pool_open_slots"
    labels:
      pool_name: "$1"

  # scheduler.tasks.starving
  - match: "airflow.scheduler.tasks.starving"
    name: "airflow_scheduler_tasks_starving"

OpenTelemetry (Airflow 2.7+)

# airflow.cfg
[metrics]
otel_on = True
otel_host = otel-collector.monitoring.svc.cluster.local
otel_port = 4318
otel_prefix = airflow
otel_interval_milliseconds = 30000

# OpenTelemetry Collector config
receivers:
  otlp:
    protocols:
      grpc:
        endpoint: 0.0.0.0:4317
      http:
        endpoint: 0.0.0.0:4318

exporters:
  prometheus:
    endpoint: "0.0.0.0:9090"
  jaeger:
    endpoint: jaeger-collector:14250

service:
  pipelines:
    metrics:
      receivers: [otlp]
      exporters: [prometheus]
    traces:
      receivers: [otlp]
      exporters: [jaeger]

---

Key Metrics to Monitor

Prometheus Queries

# Task failure rate (last 1 hour, by DAG)
sum(increase(airflow_task_failures_total[1h])) by (dag_id)

# Task success rate percentage
100 * sum(increase(airflow_task_success_total[1h])) by (dag_id) /
(sum(increase(airflow_task_success_total[1h])) by (dag_id) +
 sum(increase(airflow_task_failures_total[1h])) by (dag_id))

# P95 task duration by task_id
histogram_quantile(0.95,
  sum(rate(airflow_task_duration_seconds_bucket[1h])) by (le, dag_id, task_id)
)

# Pool slot utilization
1 - (airflow_pool_open_slots / airflow_pool_total_slots)

# Scheduler heartbeat age (seconds since last heartbeat)
time() - airflow_scheduler_last_scheduler_run

# DAG parse time
airflow_dag_processing_total_parse_time

# Queued tasks (tasks waiting for execution)
airflow_executor_queued_tasks

# Running tasks
airflow_executor_running_tasks

---

Grafana Dashboard Panels

// Grafana dashboard JSON snippet for key panels

// Panel 1: Task Success/Failure Rate
{
  "title": "Task outcomes (last 24h)",
  "type": "timeseries",
  "targets": [
    {
      "expr": "sum(increase(airflow_task_success_total[5m])) by (dag_id)",
      "legendFormat": "success - {{dag_id}}"
    },
    {
      "expr": "sum(increase(airflow_task_failures_total[5m])) by (dag_id)",
      "legendFormat": "failure - {{dag_id}}"
    }
  ]
}

# Grafana dashboard provisioning (deploy via ConfigMap)
apiVersion: v1
kind: ConfigMap
metadata:
  name: airflow-dashboard
  namespace: monitoring
  labels:
    grafana_dashboard: "1"
data:
  airflow-overview.json: |
    {
      "title": "Airflow Overview",
      "panels": [...]
    }

---

Structured Task Logging

# airflow.cfg: structured JSON logs
[logging]
log_format = {"time": "%(asctime)s", "level": "%(levelname)s", "dag": "%(dag_id)s", "task": "%(task_id)s", "run": "%(run_id)s", "message": "%(message)s"}

# Or configure programmatically
import logging
import json

class JsonFormatter(logging.Formatter):
    def format(self, record):
        log_entry = {
            "timestamp": self.formatTime(record),
            "level": record.levelname,
            "dag_id": getattr(record, "dag_id", None),
            "task_id": getattr(record, "task_id", None),
            "run_id": getattr(record, "run_id", None),
            "message": record.getMessage(),
        }
        if record.exc_info:
            log_entry["exception"] = self.formatException(record.exc_info)
        return json.dumps(log_entry)

# In a task: use structured log fields
@task
def load_orders(ds: str):
    import logging
    log = logging.getLogger(__name__)
    log.info("Loading orders", extra={
        "partition_date": ds,
        "source_table": "bronze.orders_raw",
        "target_table": "silver.orders",
    })
    rows = execute_query(...)
    log.info("Load complete", extra={"rows_loaded": rows, "partition_date": ds})

---

Alert Rules

# Prometheus alerting rules
groups:
  - name: airflow
    rules:

    - alert: AirflowSchedulerHeartbeatStale
      expr: (time() - airflow_scheduler_last_scheduler_run) > 120
      for: 2m
      labels:
        severity: critical
        team: data-engineering
      annotations:
        summary: "Airflow scheduler not heartbeating"
        runbook: "https://wiki.my-org.com/runbooks/airflow-scheduler-down"

    - alert: AirflowTaskFailureSpike
      expr: |
        sum(increase(airflow_task_failures_total[15m])) by (dag_id) > 5
      for: 5m
      labels:
        severity: warning
      annotations:
        summary: "High task failure rate for DAG {{ $labels.dag_id }}"

    - alert: AirflowPoolStarved
      expr: airflow_pool_open_slots{pool_name="default_pool"} == 0
      for: 10m
      labels:
        severity: warning
      annotations:
        summary: "Airflow default pool fully occupied for 10 min"

    - alert: AirflowSLAMiss
      expr: increase(airflow_sla_missed_total[1h]) > 0
      labels:
        severity: warning
      annotations:
        summary: "SLA missed for DAG {{ $labels.dag_id }}"

    - alert: AirflowDagParseError
      expr: airflow_dag_processing_import_errors > 0
      labels:
        severity: warning
      annotations:
        summary: "DAG import errors detected"

---

DAG Run Audit Queries

-- Failed DAG runs in last 24 hours
SELECT dag_id, run_id, state, start_date, end_date,
       EXTRACT(EPOCH FROM (end_date - start_date)) / 60 AS duration_min
FROM dag_run
WHERE state = 'failed'
  AND start_date > NOW() - INTERVAL '24 hours'
ORDER BY start_date DESC;

-- Tasks taking > 2x their historical average
WITH task_stats AS (
    SELECT dag_id, task_id,
           AVG(EXTRACT(EPOCH FROM (end_date - start_date))) AS avg_duration_sec,
           STDDEV(EXTRACT(EPOCH FROM (end_date - start_date))) AS std_duration_sec
    FROM task_instance
    WHERE state = 'success'
      AND start_date > NOW() - INTERVAL '30 days'
    GROUP BY dag_id, task_id
),
recent_tasks AS (
    SELECT dag_id, task_id, run_id,
           EXTRACT(EPOCH FROM (end_date - start_date)) AS duration_sec
    FROM task_instance
    WHERE state = 'success'
      AND start_date > NOW() - INTERVAL '24 hours'
)
SELECT r.dag_id, r.task_id, r.run_id,
       ROUND(r.duration_sec / 60, 1) AS duration_min,
       ROUND(s.avg_duration_sec / 60, 1) AS avg_duration_min,
       ROUND(r.duration_sec / NULLIF(s.avg_duration_sec, 0), 2) AS slowdown_ratio
FROM recent_tasks r
JOIN task_stats s USING (dag_id, task_id)
WHERE r.duration_sec > 2 * s.avg_duration_sec
ORDER BY slowdown_ratio DESC
LIMIT 20;

-- DagBag import errors
SELECT dag_id, last_parsed_time, last_pickled, fileloc
FROM dag
WHERE is_paused = FALSE AND is_active = FALSE
ORDER BY last_parsed_time DESC;

---

Observability Checklist

[ ] StatsD or OTEL metrics enabled in airflow.cfg
[ ] StatsD exporter → Prometheus scrape configured
[ ] Grafana dashboard with: task success/failure, duration, pool slots, scheduler heartbeat
[ ] Alert: scheduler heartbeat stale (> 2 min)
[ ] Alert: task failure spike (> 5 failures in 15 min)
[ ] Alert: pool starved (0 open slots for > 10 min)
[ ] Alert: SLA miss callback + Prometheus counter
[ ] Structured JSON log format enabled
[ ] Remote log storage (S3/GCS) — not local pods
[ ] DAG parse error alert (airflow_dag_processing_import_errors > 0)
[ ] Weekly SLA report generated from metadata DB

---

Anti-Patterns

1. No metrics backend configured — default Airflow has no metrics; scheduler hangs go undetected for hours; always configure StatsD or OTEL.
2. Alerting only on task failures — pool starvation and scheduler heartbeat issues prevent tasks from running without any failures visible; monitor queue depth and heartbeat too.
3. Logs only on local pods — pod replacement during upgrade or OOM wipes all logs; configure remote logging from day one.
4. No SLA configured — pipeline latency issues detected only by downstream consumers; set SLA on every critical DAG.
5. Using print() instead of logging — print output isn't captured by Airflow log viewer with structured context; always use logging.getLogger(__name__).

---

References

• Airflow metrics: airflow.apache.org/docs/apache-airflow/stable/administration-and-deployment/logging-monitoring/metrics.html
• StatsD exporter: github.com/prometheus/statsd_exporter
• OpenTelemetry: opentelemetry.io
• Related skills: [[dataops-airflow-production-readiness]], [[dataops-airflow-ha-review]], [[infra-prometheus-optimization]]