Your data volume is growing, but manual script execution and cron jobs can no longer keep up? We've encountered situations where N+1 queries in an ETL pipeline caused hours of delays, with no monitoring in place. A pipeline loading 5 million orders on LocalExecutor took 4 hours, and during peak loads, downtime reached 30 minutes. We solve this problem with Apache Airflow—a mature platform for orchestrating and automating data processing workflows. Our engineers have built dozens of DAG pipelines for companies in e-commerce, fintech, and logistics. Each pipeline undergoes load testing: a typical DAG processes up to 10 million records overnight, meeting the SLA with 99.9% execution time. Stability is confirmed by 50+ projects. Get a consultation for your project—we'll evaluate it in 1 day.
Airflow vs. Temporal and Camunda
Airflow is optimized for batch data processing:
- ETL/ELT pipelines (PostgreSQL → transformation → Data Warehouse)
- Daily reports and exports
- ML pipelines (data preparation → training → model deployment)
- Periodic aggregations and synchronizations
If you have event-driven business processes with human tasks, consider Temporal or Camunda. Airflow cannot wait for users for hours. But for data engineers, it's the best choice: it is 3–5 times faster than Temporal in batch scenarios.
Why We Choose KubernetesExecutor
| Executor | Scaling | Isolation | Resource Management |
|---|---|---|---|
| LocalExecutor | Limited to one node | None | Manual |
| CeleryExecutor | Horizontal via workers | Medium | Requires Redis/RabbitMQ |
| KubernetesExecutor | Automatic | Each task runs in a Pod | Via requests/limits |
KubernetesExecutor provides isolation at the task level: each task runs in its own Pod with dedicated CPU and memory. Under peak load, Kubernetes automatically spins up Pods and scales down afterward. We use this approach in production and consider it the standard for modern data pipelines.
How KubernetesExecutor Speeds Up Data Processing
Under peak load, KubernetesExecutor scales resources horizontally: instead of one node, a cluster of 10+ Pods is used. ETL pipeline execution time is reduced by 3–5 times compared to LocalExecutor. For example, a pipeline loading 5 million orders on LocalExecutor takes 4 hours; on KubernetesExecutor, it takes 50 minutes. The difference is clear.
How We Build an ETL Pipeline: Detailed Case Study
Consider an example for an online store: daily loading of orders from PostgreSQL into a DWH (PostgreSQL) with transformation and aggregation. Below is a DAG we deployed for a customer in 2 days.
Installation via Helm
helm repo add apache-airflow https://airflow.apache.org helm upgrade --install airflow apache-airflow/airflow \ --namespace airflow \ --create-namespace \ --set executor=KubernetesExecutor \ --set postgresql.enabled=true \ --set redis.enabled=true \ --values airflow-values.yaml # airflow-values.yaml aiflow: image: repository: apache/airflow tag: 2.8.0 config: AIRFLOW__CORE__DAGS_FOLDER: /opt/airflow/dags AIRFLOW__CORE__MAX_ACTIVE_RUNS_PER_DAG: "3" AIRFLOW__SCHEDULER__MIN_FILE_PROCESS_INTERVAL: "30" dags: gitSync: enabled: true repo: https://github.com/your-org/airflow-dags.git branch: main subPath: dags/ DAG — Example ETL Pipeline
from airflow import DAG from airflow.operators.python import PythonOperator from airflow.providers.postgres.operators.postgres import PostgresOperator from airflow.providers.postgres.hooks.postgres import PostgresHook from datetime import datetime, timedelta import pandas as pd default_args = { 'owner': 'data-team', 'depends_on_past': False, 'start_date': datetime(2024, 1, 1), 'retries': 2, 'retry_delay': timedelta(minutes=5), 'email_on_failure': True, 'email': ['[email protected]'], } with DAG( 'daily_orders_etl', default_args=default_args, schedule_interval='0 2 * * *', catchup=False, tags=['etl', 'orders'], description='Load and transform orders into DWH', ) as dag: def extract_orders(**context): hook = PostgresHook(postgres_conn_id='production_db') ds = context['ds'] df = hook.get_pandas_df(f""" SELECT o.id, o.customer_id, o.total, o.status, o.created_at, c.email, c.country FROM orders o JOIN customers c ON c.id = o.customer_id WHERE o.created_at::date = '{ds}' AND o.status IN ('paid', 'shipped', 'delivered') """) context['ti'].xcom_push(key='orders_count', value=len(df)) df.to_parquet(f'/tmp/orders_{ds}.parquet') return len(df) def transform_orders(**context): ds = context['ds'] df = pd.read_parquet(f'/tmp/orders_{ds}.parquet') df['order_date'] = pd.to_datetime(df['created_at']).dt.date df['revenue_usd'] = df['total'] / 100 df['is_international'] = df['country'] != 'RU' df['customer_tier'] = df['revenue_usd'].apply( lambda x: 'vip' if x >= 500 else 'regular' ) df.to_parquet(f'/tmp/orders_transformed_{ds}.parquet') def load_to_dwh(**context): ds = context['ds'] df = pd.read_parquet(f'/tmp/orders_transformed_{ds}.parquet') hook = PostgresHook(postgres_conn_id='datawarehouse') engine = hook.get_sqlalchemy_engine() df.to_sql('fact_orders', engine, schema='dwh', if_exists='append', index=False, method='multi', chunksize=1000) aggregate_metrics = PostgresOperator( task_id='aggregate_metrics', postgres_conn_id='datawarehouse', sql=""" INSERT INTO dwh.daily_metrics (date, total_revenue, orders_count, avg_order) SELECT '{{ ds }}'::date, SUM(revenue_usd), COUNT(*), AVG(revenue_usd) FROM dwh.fact_orders WHERE order_date = '{{ ds }}' ON CONFLICT (date) DO UPDATE SET total_revenue = EXCLUDED.total_revenue, orders_count = EXCLUDED.orders_count, avg_order = EXCLUDED.avg_order; """, ) extract = PythonOperator(task_id='extract_orders', python_callable=extract_orders) transform = PythonOperator(task_id='transform_orders', python_callable=transform_orders) load = PythonOperator(task_id='load_to_dwh', python_callable=load_to_dwh) extract >> transform >> load >> aggregate_metrics Parallel Execution and Sensors
from airflow.utils.task_group import TaskGroup with TaskGroup('process_regions') as process_regions: for region in ['EU', 'US', 'APAC']: PythonOperator( task_id=f'process_{region.lower()}', python_callable=process_region_data, op_kwargs={'region': region} ) extract >> process_regions >> aggregate_all To wait for external events, we use Sensors: FileSensor for files, HttpSensor for APIs. This is a standard production pattern.
KubernetesExecutor in Action
executor_config = { 'KubernetesExecutor': { 'request_memory': '2Gi', 'request_cpu': '500m', 'limit_memory': '4Gi', 'image': 'custom-airflow:2.8.0-pandas', } } heavy_transform = PythonOperator( task_id='heavy_transform', python_callable=transform_large_dataset, executor_config=executor_config ) How DAG Failures Are Handled
When a task fails, Airflow automatically retries with exponential backoff. We configure alerts in Telegram/Slack for every failure and long-running task. Additionally, we integrate metrics into Prometheus/Grafana: dashboards show execution time, number of successful/failed tasks, and resource utilization. This enables rapid incident response.
Minimum infrastructure requirements:
- Kubernetes cluster version 1.24+
- PostgreSQL 13+ for metadata database
- Redis (optional, for CeleryExecutor)
- Storage capacity: from 100 GB for logs and artifacts
Process Overview
- Analysis: We study your data sources, volumes, SLA, current issues, Airflow configuration, and infrastructure.
- Design: We outline DAGs, choose an Executor, plan error handling and alerts.
- Implementation: We write DAG code, Helm configs, CI/CD, and tests.
- Deployment: We set up infrastructure (Kubernetes or Docker), configure GitSync and monitoring.
- Testing: We run backfill on historical data and verify correctness.
- Documentation and Training: We deliver a runbook and train your team.
- Support: 2 weeks post-launch for stabilization.
What's Included in the Work
- Developed DAGs (Python code ready for production)
- Airflow configuration (values.yaml, variables, connections)
- CI/CD pipeline for automatic DAG deployment via Git
- Docker image with dependencies (libraries, drivers)
- Documentation: DAG descriptions, run instructions, troubleshooting
- Team training (2–3 sessions)
- Monitoring: alerts in Telegram/Slack, Grafana dashboards
Timeline Estimates
| Phase | Duration |
|---|---|
| Airflow deployment (Helm/Docker) + first DAG | 3–5 days |
| ETL pipeline with 5–8 tasks, transformations, and DWH loading | 1–2 weeks |
| Complex pipeline with parallelism, sensors, and backfill | 2–4 weeks |
Our Experience and Guarantees
We have 10+ years of experience in developing data infrastructure. We have delivered 50+ Airflow projects for e-commerce, fintech, and logistics. We use best practices: GitSync, KubernetesExecutor, alerting, DAG versioning. We guarantee stable pipeline operation and enterprise-level documentation.
We are ready to build a similar pipeline for your data. Contact us for an evaluation—we'll respond within 1 day. Or request a consultation to discuss details without obligation.