AutoGineer
Generated banner for Managed Airflow Recommendations
Generated logo for Managed Airflow Recommendations

Managed Airflow Recommendations

GoogleOfficialPublished by Google

  • Skill
  • Research and analysis
  • Cloud and DevOps
  • Apache-2.0
  • Low risk
  • Has scripts
Download ZIP

From the author

Provides recommendations and best practices for creating, configuring, tuning and optimizing Managed Service for Apache Airflow (MSAA; formerly Cloud Composer) environments. Use when the user asks for guidance, recommendations, or best practices on configuring Managed Airflow environments, scaling Airflow environments, preventing workload restarts, or analyzing system health. Don't use for troubleshooting individual DAG or task run failures, authoring DAGs, or migrating DAGs to newer Airflow versions.
— the description by Google

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill acts as an expert advisor for optimizing Managed Service for Apache Airflow (formerly Cloud Composer) environments. It uses a suite of diagnostic scripts to collect telemetry on CPU, memory, disk usage, and component restarts, allowing for evidence-based recommendations rather than speculative scaling. The skill guides users through environment health checks, configuration audits, and best practices for DAG efficiency, ensuring that infrastructure adjustments are justified by performance data.

What it does

  • Analyzes environment health and performance metrics using custom diagnostic scripts.
  • Identifies resource bottlenecks like CPU saturation, memory pressure, and disk usage spikes.
  • Evaluates DAG parsing efficiency to detect scheduler performance issues.
  • Provides actionable recommendations for infrastructure scaling and configuration tuning.
  • Offers guidance on Airflow best practices such as using deferrable operators and avoiding top-level code.
  • Retrieves environment configurations and logs for detailed troubleshooting.

When to use it

  • When you need to optimize the performance or cost of an existing Managed Airflow environment.
  • When investigating unexplained workload restarts or system instability.
  • When you need to verify if your current environment sizing is appropriate for your workload.
  • When seeking expert advice on implementing Airflow best practices to improve scheduler and worker efficiency.

Advantages

  • Evidence-based: Recommendations are backed by actual telemetry data.
  • Context-aware: Tailored advice based on your specific environment configuration.
  • Comprehensive: Covers infrastructure, configuration, and DAG-level best practices.
  • Safe: Operates in read-only mode to prevent accidental environment changes.

Usage examples

“My Airflow environment is experiencing frequent worker restarts. Can you help me investigate?”
The agent will guide you through gathering restart and memory metrics to determine if the issue is caused by OOM events or resource saturation.
“What are the best practices for optimizing my Managed Airflow environment's performance?”
The agent will analyze your current configuration and provide a list of recommendations, such as adjusting worker scaling or optimizing DAG parsing.

Install

  1. Download the skill as a ZIP:

    https://mp-staging.autogineer.com/agent-skills/google/managed-airflow-recommendations/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip managed-airflow-recommendations.zip -d .agents/skills

Documentation by Google

Managed Service for Apache Airflow (formerly Cloud Composer) recommendations

This skill provides specialized instructions for providing recommendations, best practices, and performance-tuning for Managed Service for Apache Airflow (formerly Cloud Composer) environments. It leverages custom scripts to gather key telemetry data, enabling you to deliver data-backed, context-aware advice.

Usage Attribution

MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read and follow the instructions in Usage Attribution.

Role & Persona

You are a Managed Service for Apache Airflow and Airflow Performance Expert. You provide concrete, evidence-based recommendations for system architecture (scaling parameters, sizing) and offer advice to address reliability issues (parsing efficiency, workload restarts). You do not blindly recommend "upsizing" immediately; instead, you analyze metrics and code to find optimal tuning solutions.

Available Resources

The following scripts and references are available to assist in gathering data and diagnosing issues:

Scripts (scripts/):

  • dag_parsing_stats.py: Analyzes DAG parsing times and efficiency metrics to identify processing bottlenecks.

  • environment_health.py: Retrieves general environment health indicators and status.

  • workload_cpu_usage.py: Collects CPU utilization metrics for Managed Airflow workloads (workers, schedulers, webserver).

  • workload_disk_usage.py: Monitors disk space usage for environment workloads.

  • workload_memory_usage.py: Gathers memory consumption metrics to help identify potential Out-of-Memory issues.

  • workload_restarts.py: Retrieves restart counts for Airflow components to help identify unstable workloads.

  • lib/: Shared helpers used by the scripts above (Cloud Monitoring PromQL API client and flag parsing). Don't run them directly.

References (references/):

  • gcloud reference: A reference guide containing essential gcloud commands for retrieving and inspecting Managed Airflow environment configurations.

Prerequisites

Verify the following before gathering data. The scripts and gcloud commands fail without them.

  1. Google Cloud SDK: Verify gcloud is installed (Installation Guide).

  2. Python: The scripts require Python 3.9 or later and use only the standard library (no extra packages to install).

  3. Authentication: Authenticate the CLI and Application Default Credentials. The scripts obtain an access token with gcloud auth application-default print-access-token.

    gcloud auth login
gcloud auth application-default login

  4. Project & Billing: Use the project that hosts the environment ({env_project}). The project must have an active billing account attached (it already does if the environment is running).

    gcloud config set project {env_project}

  5. API Enablement: Make sure the Managed Airflow, Cloud Monitoring and Cloud Logging APIs are enabled:

    gcloud services enable composer.googleapis.com monitoring.googleapis.com logging.googleapis.com

  6. IAM Roles: Following the principle of least privilege, the user only needs read-only roles on the environment's project:

    • roles/composer.environmentAndStorageObjectViewer: Describe the environment and list files in the environment's bucket.

    • roles/monitoring.viewer: Read environment metrics (used by the scripts).

    • roles/logging.viewer: Read environment logs.

    See Access control for details.

  7. Region: Use the location of the existing environment as {env_region} (e.g., us-central1). See Google Cloud locations for region availability.

Task Execution Process

When the user requests recommendations or best practices for an Airflow environment, follow this structured workflow:

  1. Context Gathering:

    • Determine the Target Environment (environment name, project ID, region). If missing, kindly ask the user to provide them.

    • Establish the target Timeframe (e.g., past 24 hours, past 7 days) if the user is investigating a recent performance incident.

  2. Environment Setup & Configuration Verification:

    • Use the gcloud commands defined in gcloud reference to retrieve the current environment configuration.

    • Inspect environment scales (e.g. environment size, number of schedulers, max workers, cpu/ram limits).

  3. Metrics Gathering (Diagnostic Tools):

    • Execute the provided Python scripts in the scripts/ directory to gather system telemetry. Do NOT make generalizations without data.

    • Tip: run python3 scripts/{script_name}.py --help to discover the purpose and parameters of each script.

  4. Analysis & Diagnosis:

    • CPU/Memory/Disk: Look for saturation. Are workers consistently maxing out CPU? Are schedulers OOMing (Out of Memory) and causing restarts?

    • Restarts: High restart counts (especially for the scheduler or workers) often indicate memory issues or unoptimized DAGs blocking the event loop.

    • DAG Parsing: Are DAG parsing times high? This impacts the scheduler's ability to orchestrate efficiently (tip: inspect dag-processor-manager log).

  5. Recommendation Generation:

    • Based on the data collected, present an actionable, categorized list of recommendations.

    • Categories should usually include:

      • Infrastructure & Scaling: Recommendations around workload count, workload resources (cpu/memory), core infrastructure size (small/medium/large).

      • Airflow Configurations: Optimizing airflow.cfg overrides (e.g. parallelism, max_active_tasks_per_dag, dag_file_processor_timeout).

      • Bucket Hygiene: Optimizations related to the environment bucket (e.g. remove non-DAG files in dags/ directory).

      • Production Best Practises: Recommendations around features like high-resilience mode and database retention (only if not already enabled).

Airflow Best Practices (General Knowledge)

  • Top-Level Code: DAG files should NEVER contain heavy processing, database connections, or API calls outside of task definitions (top-level code). This blocks the DagProcessor and increases scheduler CPU. Code should be pushed into operators or hook methods.

  • Deferrable Operators: Encourage the use of Deferrable Operators (or Async operators) and the Triggerer component to run long-waiting tasks (like checking a sensor or waiting for a BigQuery job) without tying up worker slots and resources.

  • Dynamic DAGs: Creating DAGs dynamically should be done carefully (prefer Dynamic Task Mapping over dynamically generating DAG files in a loop) to keep parsing times low.

  • Variables/Connections: Remind users that reading Airflow Variables or Connections at the top level of a DAG forces an unnecessary database hit on every heartbeat. Use them inside task execution elements.

  • Storage Limits: Temporary data should not be written blindly to local task storage unless properly cleaned up, as it can cause workload_disk_usage spikes and task failures.

Important Constraints & Instructions

  • Evidence-Based Decisions: Do not blindly recommend increasing machine sizes without first checking if the memory or CPU is actually saturated.

  • Distinguish Gen 2 / Gen 3: Ensure recommendations match the architecture.

  • Format: Use Markdown to structure your report clearly. Use tables where appropriate for metrics summaries.

  • Sensitive Operations: This skill is read-only. NEVER apply recommended changes (e.g. gcloud composer environments update) without explicit user approval, because updating an environment can restart Airflow components, interrupt running tasks and increase cost.

  • Cost: Scaling recommendations (more workers, larger workloads, larger environment size, high-resilience mode) increase environment cost. Point users to the Managed Airflow pricing page instead of quoting prices, because pricing changes over time.

  • Cleanup: The scripts and gcloud commands in this skill only read configuration, metrics and logs. They don't create any resources, so no teardown is required.

Files

scripts/workload_disk_usage.py

# Copyright 2026 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Script to calculate disk usage of Managed Airflow components."""

import argparse
import collections
import datetime
import json

from lib import flagutils
from lib import monitoring

NUM_POINTS = 10


def main():
  parser = argparse.ArgumentParser(
      description="Calculate disk usage of Managed Airflow components."
  )
  parser.add_argument(
      "--project_id", type=str, required=True, help="GCP project ID"
  )
  parser.add_argument(
      "--location", type=str, required=True, help="GCP location"
  )
  parser.add_argument(
      "--environment_name",
      type=str,
      required=True,
      help="Managed Airflow environment name",
  )
  parser.add_argument(
      "--components",
      type=str,
      required=False,
      help=(
          "Comma-separated list of Managed Airflow components to query"
          " (default: all)"
      ),
  )
  parser.add_argument(
      "--query_duration_hours",
      type=int,
      required=True,
      help="Query duration in hours (minimum 1 hour)",
  )
  parser.add_argument(
      "--query_end_time",
      type=str,
      required=False,
      default=None,
      help="Query end time in ISO 8601 format (default: now)",
  )
  args = parser.parse_args()

  if args.query_duration_hours < 1:
    parser.error("--query_duration_hours must be at least 1.")

  try:
    components = flagutils.parse_components_flag(args.components)
  except ValueError as e:
    parser.error(str(e))

  # Calculate step (duration / 10)
  step_seconds = (args.query_duration_hours * 3600) // NUM_POINTS
  common_filter = (
      'monitored_resource="cloud_composer_workload",'
      f'project_id="{args.project_id}",'
      f'environment_name="{args.environment_name}",'
      f'location="{args.location}",'
      f'type=~"{"|".join(components)}"'
  )

  query = f"""
label_replace(
  sum by (type)(avg_over_time({{__name__="composer.googleapis.com/workload/disk/quota",{common_filter}}}[{step_seconds}s])),
  "stat", "totalLimit", "", ""
)
or
label_replace(
  sum by (type)(avg_over_time({{__name__="composer.googleapis.com/workload/disk/bytes_used",{common_filter}}}[{step_seconds}s])),
  "stat", "totalUsage", "", ""
)
or
label_replace(
  max by (type)(avg_over_time({{__name__="composer.googleapis.com/workload/disk/bytes_used",{common_filter}}}[{step_seconds}s])),
  "stat", "maxUsage", "", ""
)
or
label_replace(
  avg by (type)(avg_over_time({{__name__="composer.googleapis.com/workload/disk/bytes_used",{common_filter}}}[{step_seconds}s])),
  "stat", "avgUsage", "", ""
)
"""

  end_time = datetime.datetime.now(datetime.timezone.utc).timestamp()
  if args.query_end_time:
    end_time = datetime.datetime.fromisoformat(args.query_end_time).timestamp()
  start_time = end_time - (args.query_duration_hours * 3600)

  params = {
      "query": query,
      # The first evaluation point covers [start_time, start_time + step], so
      # the query returns exactly NUM_POINTS points.
      "start": start_time + step_seconds,
      "end": end_time,
      "step": f"{step_seconds}s",
  }

  try:
    data = monitoring.fetch_time_series_promql(args.project_id, params)
  except Exception as e:
    print(f"Failed to fetch data: {e}")
    return
  monitoring.exit_if_no_data(
      data, args.project_id, args.location, args.environment_name
  )

  output = collections.defaultdict(dict)

  for result in data.get("data", {}).get("result", []):
    component_upper = result["metric"].get("type")
    if not component_upper:
      continue
    component = component_upper.lower()

    stat = result["metric"].get("stat")

    for val in result["values"]:
      timestamp_sec = int(val[0])
      value_float = float(val[1]) / (1024 * 1024 * 1024)
      end_dt_str = datetime.datetime.fromtimestamp(
          timestamp_sec, datetime.timezone.utc
      ).isoformat()
      start_dt_str = datetime.datetime.fromtimestamp(
          timestamp_sec - step_seconds, datetime.timezone.utc
      ).isoformat()
      if timestamp_sec not in output[component]:
        output[component][timestamp_sec] = {
            "startTime": start_dt_str,
            "endTime": end_dt_str,
        }
      output[component][timestamp_sec][stat] = round(value_float, 3)

  final_output = {}
  for comp, points in output.items():
    sorted_points = [points[ts] for ts in sorted(points.keys())]
    final_output[comp] = sorted_points

  print("Workload disk usage (in GiB):")
  print(json.dumps(final_output, indent=2))


if __name__ == "__main__":
  main()

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs a command-line tool, Needs Python

Security

What our scanner noticed· Low risk

  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by
Google
Publisher
Google (Official)
Source
google/skills · skills/cloud/managed-airflow-recommendations · commit 2194d3b
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Google.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download managed-airflow-recommendations.zip