AIP-89: Privacy-First Telemetry for Apache Airflow

Status: Draft
Authors: Bolke de Bruin
Created: 2025-10-08
Target Airflow Version: 3.2.0
Related Issues:

Abstract

This AIP proposes implementing a privacy-first, transparent, and community-governed telemetry system for Apache Airflow using the Apache Software Foundation's Matomo instance as the default collection endpoint. This addresses the complete absence of telemetry data since the removal of Scarf, while learning from past mistakes to ensure community trust.

Motivation

Currently, Airflow has no telemetry collection capability following the removal of Scarf. While this removal was necessary due to community concerns about privacy and transparency, it has created significant challenges:

Current Problems:

• No visibility into feature usage - Maintainers cannot determine which features are actually being used
• Difficult deprecation decisions - No data on which operators, providers, or Python versions are in active use
• Wasted development effort - Resources spent on features that may have minimal adoption
• Inability to prioritize - Cannot make data-driven decisions about which providers or features to maintain
• Performance optimization challenges - No understanding of real-world deployment patterns

Why telemetry is crucial for Airflow's development:

• Feature prioritization - Understanding how features are used helps prioritize development efforts
• Deprecation decisions - Knowing which Python versions, databases, or providers are in use prevents breaking production systems
• API design decisions - Usage patterns inform API design and breaking change decisions
• Performance optimization - Deployment patterns guide optimization efforts
• Provider maintenance - Understanding which providers are actively used informs maintenance priorities
• Community resource allocation - Data helps direct volunteer and sponsored development time effectively

Historical Context

The Airflow 2.10.0 release included opt-out telemetry via Scarf, which created significant community backlash due to:

• Lack of advance communication about the feature
• Opt-out rather than opt-in approach
• Limited transparency about data collection
• Privacy concerns, even though no PII was collected
• Corporate compliance challenges

This led to Scarf's removal. However, the need for usage insights remains critical for the project's long-term health and sustainability.

Goals

• Privacy by design - Collect only minimal, non-personal data with differential privacy techniques
• Transparency - Users can see exactly what data is collected and sent
• User control - Easy opt-out with clear visibility into telemetry status
• Community governance - Changes to collected data require community approval
• Trust rebuilding - Demonstrate Airflow's commitment to user privacy and autonomy
• Statistical usefulness - Collect enough data to make meaningful decisions (targeting >20% adoption)
• No-impact on performance - Ensure that any data collection does not affect the performance of Airflow

Non-Goals

• Collecting personally identifiable information (PII)
• Collecting DAG names, task names, or other deployment-specific identifiers
• Collecting connection details, credentials, or sensitive configuration
• Real-time monitoring or performance profiling of individual deployments
• Commercial use of telemetry data
• Tracking individual users or installations over time

Proposal

Data Collection Endpoint

• Default: Apache Software Foundation's Matomo instance (https://analytics.apache.org/)
  • Hosted and managed by the ASF
  • Complies with ASF privacy policies
  • Data is owned by the Apache Airflow project
• Alternative endpoints: Users may configure alternative endpoints if required by organizational policies

Data Collected

The following minimal data will be collected when telemetry is enabled:

Installation Metrics (with noise)

• Airflow version: Major and minor version including patch version (e.g., 3.2.0) 
• Python version: Major and minor version only (e.g., 3.11)
• Deployment type: One of docker, kubernetes, systemd, standalone, unknown
• Operating system: Generic OS type (e.g., linux, darwin, windows)
• Architecture: System architecture (e.g., x86_64, arm64)
• Installation method: One of pip, docker, helm, managed_service, unknown

Usage Metrics (aggregated and noised)

• Active providers: List of installed provider packages including version (e.g., apache-airflow-providers-amazon==8.x.y)
• Executor type: Configured executor(s) (e.g., LocalExecutor, CeleryExecutor, KubernetesExecutor - only executors that are known to 'Official' Airflow otherwise "unspecified")
• Database backend: Database type only (e.g., postgres, mysql, sqlite)
• Enabled features: Boolean flags for optional features (e.g., has_webserver_auth, has_dag_serialization)

Aggregated Statistics (collected weekly with differential privacy)

Note: All counts below have Gaussian noise added (±10% standard deviation) to prevent exact fingerprinting while maintaining statistical usefulness.

• DAG count tier: Bucketed as 1-10, 11-50, 51-100, 101-500, 501-1000, 1000+
• Task count tier: Bucketed as 1-50, 51-250, 251-1000, 1001-5000, 5000+
• DAG run count (7d): Rounded to nearest power of 10 (e.g., ~1000, ~10000)
• Task instance count (7d): Rounded to nearest power of 10
• Active hours per week: Bucketed as <10, 10-50, 50-100, 100-168 (helps understand CI vs production usage)

Operator Usage (aggregated with noise)

• Top 10 operator only: Count of usage with ±5 noise added to each count
• Percentages only: Instead of raw counts for privacy (e.g., "PythonOperator: 45% of tasks")
• No parameters or task names

Technical Metadata

• Timestamp: UTC timestamp rounded to the nearest hour
• Session ID: Daily rotating hash (not persistent across days)
• IP address: Used only for geolocation to country level, then immediately discarded (last two octets zeroed before logging)

Data NOT Collected

The following data will explicitly not be collected:

• DAG names, descriptions, or any DAG content
• Task names, parameters, or configurations
• Variable names or values
• Connection names, URIs, or credentials
• Log contents or error messages containing user data
• User names, emails, or authentication information
• Full IP addresses (only country-level geolocation)
• Hostnames or deployment identifiers
• Installation UUIDs that persist across reinstalls
• Code or custom operator implementations
• File paths or directory structures
• Exact counts that could fingerprint individual installations
• Any correlation between daily sessions

Consent Mechanism

Default Behavior: Opt-In

Following community feedback from the Scarf incident, telemetry is opt-in by default. While this typically yields lower participation rates (3-10%), it prioritizes user trust and explicit consent, which is critical for rebuilding community confidence after the previous telemetry issues.

First-Time Installation

When Airflow is started for the first time (or after upgrade to a version with this AIP):

CLI Installation: A 30-second interactive prompt appears when a TTY is detected:

╔══════════════════════════════════════════════════════════════════╗
║              Apache Airflow Anonymous Telemetry                  ║
╠══════════════════════════════════════════════════════════════════╣
║                                                                  ║
║  Help improve Airflow by sharing anonymous usage data.           ║
║                                                                  ║
║  We collect: Feature usage, performance metrics, errors          ║
║  We DON'T collect: Personal data, file contents, IPs             ║
║                                                                  ║
║  Privacy protections:                                            ║
║  • Differential privacy (noise added to all counts)              ║
║  • Daily rotating session IDs (no long-term tracking)            ║
║  • Open source implementation you can inspect                    ║
║                                                                  ║
║  View details: airflow.apache.org/docs/telemetry                 ║
║  Debug mode: airflow telemetry --debug                           ║
║                                                                  ║
║  Enable telemetry? [y/N] (auto-decline in 10 seconds)            ║
╚══════════════════════════════════════════════════════════════════╝

User Options:

• Press Y or y: Enable telemetry
• Press N or n or wait 10 seconds: Disable telemetry (default)
• All choices are stored in database for persistence

Non-interactive Mode: If stdin is not a TTY or --non-interactive flag is present, telemetry defaults to disabled (unset)

Web UI Admin Screen

If no explicit choice has been recorded, the Admin UI displays a dismissible banner:

╭────────────────────────────────────────────────────────────────────╮
│ Help Improve Airflow                                               │
│                                                                    │
│ Share anonymous usage data to help us prioritize features.         │
│ [Learn More] [Enable Telemetry] [No Thanks]                        │
╰────────────────────────────────────────────────────────────────────╯

Telemetry Settings Page (Admin > Telemetry Settings):

• Current status (Enabled/Disabled)
• Last transmission timestamp
• Preview of next data payload (viewable as JSON before sending)
• Enable/Disable toggle with immediate effect
• Link to full telemetry documentation
• Export of all telemetry data sent in past 90 days (for GDPR compliance)
• Telemetry test button (send test payload and show response)

Configuration

Database Storage

Telemetry preference is stored in the airflow_settings table:

INSERT INTO airflow_settings (key, value) VALUES 
  ('telemetry.enabled', 'true'),
  ('telemetry.daily_session_salt', 'random-daily-salt'),
  ('telemetry.consent_timestamp', '2025-10-08T12:34:56Z'),
  ('telemetry.last_sent_timestamp', '2025-10-08T13:00:00Z');

Note: No persistent installation UUID is stored to prevent long-term tracking.

Configuration File

Users can override database settings via airflow.cfg:

[telemetry]
# Options: true, false, unset
# unset = defer to database setting (default)
enabled = unset

# Optional: Override collection endpoint
# Default: https://analytics.apache.org/
endpoint = https://analytics.apache.org/

# Collection interval in seconds (default: 86400 = daily)
collection_interval = 86400

# Debug mode - print payloads to stdout instead of sending
debug = false

Environment Variables

# Disable telemetry
export AIRFLOW__TELEMETRY__ENABLED=false

# Enable debug mode to see payloads
export AIRFLOW__TELEMETRY__DEBUG=true

Configuration Precedence:

1. Environment variable AIRFLOW__TELEMETRY__ENABLED
2. Configuration file [telemetry].enabled
3. Database setting
4. Default (enabled with easy opt-out)

Five Ways to Control Telemetry

Following best practices, we provide multiple easy methods to enable or disable telemetry:

1. Environment variable: export AIRFLOW__TELEMETRY__ENABLED=true or false
2. Config file: Set enabled = true or false in [telemetry] section
3. CLI command: airflow config set telemetry.enabled true or false
4. Admin UI: Toggle in Telemetry Settings page
5. Database: Update airflow_settings table directly

Data Transmission

Transmission Schedule

• Frequency: Once per day (configurable via collection_interval)
• Time: Randomized within a 1-hour window to avoid thundering herd
• Retry logic: Up to 3 retries with exponential backoff on failure
• Timeout: 5-second timeout per request (fails silently if exceeded)
• Graceful degradation: Failures are logged at DEBUG level only and never impact Airflow functionality

Transmission Method

• Protocol: HTTPS POST to Matomo tracking API
• Format: JSON payload
• User-Agent: Apache-Airflow/{version} Telemetry/1.0
• IP Anonymization: Last two octets zeroed before Matomo processing (dependent on ASF Infra)
• No cookies: No tracking cookies or persistent identifiers
• Daily session ID: Rotates daily using hash(date + random_salt) - cannot correlate across days

Differential Privacy Implementation

All numeric counts have Gaussian noise added before transmission:

def add_noise(value: int, sensitivity: float = 0.1) -> int:
    """Add Gaussian noise to protect individual values while maintaining aggregate accuracy.
    
    sensitivity: Standard deviation as percentage of value (default 10%)
    """
    noise = random.gauss(0, value * sensitivity)
    return max(0, int(value + noise))

# Example: Report "approximately 47" not "exactly 47"
dag_count = add_noise(actual_dag_count, sensitivity=0.1)

Local Aggregation

Data is aggregated locally before transmission to minimize information leakage:

# Bad: Send every DAG run
for dag_run in dag_runs:
    send_telemetry(dag_run)

# Good: Aggregate to daily totals with noise
daily_stats = {
    'dag_runs_today': add_noise(len(dag_runs)),
    'success_rate': round(success_count / total_count, 2)
}
send_telemetry(daily_stats)

Sample Payload

{
  "telemetry_version": "1.0",
  "timestamp": "2025-10-08T12:00:00Z",
  "daily_session_id": "a7b9c2e4",
  "airflow_version": "3.2",
  "python_version": "3.11",
  "deployment_type": "kubernetes",
  "install_method": "helm",
  "os": "linux",
  "architecture": "x86_64",
  "database_backend": "postgres",
  "executor": "KubernetesExecutor",
  "providers": [
    {"name": "apache-airflow-providers-amazon", "version": "8.x"},
    {"name": "apache-airflow-providers-google", "version": "10.x"}
  ],
  "enabled_features": {
    "webserver_auth": true,
    "dag_serialization": true
  },
  "usage_stats": {
    "dag_count_tier": "51-100",
    "task_count_tier": "251-1000",
    "dag_runs_7d": "~2000",
    "task_instances_7d": "~25000",
    "active_hours_per_week": "100-168"
  },
  "operator_usage_top10": {
    "PythonOperator": "45%",
    "BashOperator": "22%",
    "S3ToRedshiftOperator": "8%"
  }
}

Transparency Features

Debug Mode

Users can inspect telemetry data before it's sent:

# See exactly what would be sent (without actually sending)
$ airflow telemetry --debug

Telemetry Debug Mode
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Telemetry Status: ENABLED
Endpoint: https://analytics.apache.org/
Next transmission: 2025-10-08 14:23:15 UTC

Payload Preview:
{
  "telemetry_version": "1.0",
  "airflow_version": "3.2",
  ...
}

This payload will be sent in 2 hours
Disable with: airflow config set telemetry.enabled false

Version Status

$ airflow version
Apache Airflow: 3.2.0
Python: 3.11.5
Telemetry: ENABLED (disable: airflow config set telemetry.enabled false)

$ airflow version --no-telemetry
Apache Airflow: 3.2.0
Python: 3.11.5
Running with telemetry disabled

Governance and Changes

Process for Modifying Collected Data

Any changes to the data collected (additions or removals) require:

1. Discussion on Dev List: Proposal explaining the change and justification
2. Dev List Vote: Lazy consensus vote on dev@airflow.apache.org (72-hour voting period)
3. Documentation Update: Update telemetry documentation with exact fields
4. Release Notes: Prominent mention in release notes under "Telemetry Changes" section
5. In-App Notification: Users with telemetry enabled see a one-time notification in Web UI about data collection changes with option to review and opt-out
6. Cli Notification: Users with telemetry enabled see a one-time cli notification with an option to review and opt-out

Telemetry Schema Versioning

• Telemetry payloads include a telemetry_version field
• Breaking changes increment the major version
• Additive changes increment the minor version
• Older Airflow versions continue sending their schema version
• Backend supports multiple schema versions simultaneously

Transparency and Data Access

Public Dashboard

A public dashboard will be created showing:

• Aggregate statistics (total installations, version distribution)
• Provider popularity
• Executor type distribution
• Database backend distribution
• Deployment type breakdown
• Geographic distribution (country level only)
• Python version adoption over time

Dashboard URL: https://analytics.apache.org/ (to be created)

Raw Data Access

• Aggregated, anonymized data will be made available as quarterly CSV exports
• Individual installation data will never be published
• Large deployment users (AWS, Google, Astronomer, etc.) may request access to aggregate insights for comparison purposes if they share telemetry data

Open Source Implementation

All telemetry code will be:

• Located in a single module: airflow/telemetry/ for easy inspection
• Thoroughly commented explaining privacy decisions
• Linked prominently in documentation
• Reviewable by privacy advocates before release

Security Considerations

• No Authentication Required: Telemetry endpoint is unauthenticated (prevents tracking via auth tokens) - access is through greylisting if possible (check with ASF Infra)
• Rate Limiting: Backend needs to implement rate limiting per daily session ID to prevent abuse (dependent on ASF)
• Schema Validation: All payloads are validated against JSON schema before processing
• Data Retention: Raw telemetry data retained for 90 days, then deleted; aggregates retained indefinitely
• ASF Infrastructure: Hosted on ASF infrastructure with ASF security policies
• HTTPS Only: All transmissions over TLS 1.2+
• No External Dependencies: Telemetry collection uses only Python standard library (except HTTP client)
• Silent Failures: Network errors, timeouts, and transmission failures never raise exceptions or log above DEBUG level

Our Telemetry Promise

We publicly commit to:

1. We collect the minimum data necessary to improve Airflow
2. We never collect personal information or correlate sessions across days
3. We delete raw data after 90 days and only keep aggregates
4. We will never sell or share individual-level data
5. You can opt out anytime without feature degradation
6. We add statistical noise to all counts to prevent fingerprinting
7. We publish all telemetry code for community review

Backward Compatibility

Removal of Scarf

• Scarf telemetry was removed in Airflow 2.10.2 following community feedback
• Airflow currently has no telemetry capability since 2.10.2
• This AIP proposes a completely new, privacy-first implementation
• No migration of previous telemetry data to new system

Upgrade Experience

Users upgrading from Airflow 3.1.x or earlier to 3.2.0:

• On first 3.2.0 startup, see the telemetry consent prompt
• Release notes prominently explain new telemetry system and lessons learned from Scarf
• Admin UI shows telemetry status banner (dismissible for 30 days)
• Clear communication about five easy ways to opt out
• Emphasis on differential privacy and transparency improvements

Documentation

New documentation will be added:

Main Documentation Page: /docs/apache-airflow/telemetry.rst

• What data is collected (comprehensive list with examples)
• How we protect privacy (differential privacy, noise, bucketing)
• Five ways to enable or disable telemetry
• Where data is sent and who controls it
• How data is used
• Our telemetry promise
• Debug mode instructions
• FAQ addressing common concerns
• Why opt-in and how we handle low participation rates

Admin Guide: /docs/apache-airflow/administration-and-deployment/telemetry.rst

• Configuration options
• Enterprise deployment considerations
• Corporate compliance guidance
• Troubleshooting
• Alternative endpoint configuration

Release Notes

Dedicated section in every release with telemetry changes

Alternatives Considered

Alternative 1: No Telemetry

Rationale for rejection: Development prioritization would rely solely on GitHub issues and surveys, which don't represent actual usage patterns. This leads to poor prioritization decisions and wasted effort.

Alternative 2: Opt-Out by Default

Rationale for rejection: While opt-out would provide better data coverage (20-30% vs 3-10%), it violates the privacy-first principle that's critical for rebuilding trust after the Scarf incident. Given Airflow's history, explicit opt-in is the only approach that respects user autonomy and rebuilds community confidence. We accept lower participation rates as the cost of maintaining trust.

Alternative 3: Exact Counts Without Noise

Rationale for rejection: Exact counts could fingerprint individual installations and track them across sessions. Differential privacy with noise prevents this while maintaining statistical usefulness.

Alternative 4: Persistent Installation UUID

Rationale for rejection: A persistent UUID enables long-term tracking of individual installations. Daily rotating session IDs prevent correlation across days while still allowing daily aggregation.

Alternative 5: Self-Hosted Telemetry Endpoint Only

Rationale for rejection: Users can configure this, but default should be ASF-hosted for trust and convenience. Most users won't self-host.

Alternative 6: Third-Party Service (e.g., PostHog, Segment)

Rationale for rejection: Using ASF infrastructure keeps data under project control and avoids third-party dependencies. Builds more trust with community.

Addressing Common Concerns

"Why not make it opt-out?"

Given typical opt-in participation rates of 3-10%, opt-out would give us much better data coverage. However, after the Scarf incident, explicit user consent is critical for rebuilding trust. We prioritize community trust over data coverage and accept that we'll need creative approaches (surveys, optional extended telemetry programs) to supplement the lower opt-in rates.

"How do I verify what you're sending?"

Three ways:

1. Run airflow telemetry --debug to see the exact payload before transmission
2. Inspect network traffic (single HTTPS POST daily)
3. Review our open source implementation at https://github.com/apache/airflow/tree/main/airflow/telemetry

"Will you sell this data?"

We legally commit in our telemetry promise to never selling individual-level data. We may share aggregate statistics like "40% of users use feature X" in our reports, but never individual installation data.

"Can you track me across reinstalls?"

No. We use daily rotating session IDs instead of persistent UUIDs. Each day gets a new random session ID that cannot be correlated with previous days.

"What about CI/CD environments?"

CI/CD environments should set AIRFLOW__TELEMETRY__ENABLED=false in their environment. The data would also naturally be filtered out on our end due to unusual patterns (hundreds of installations per day from same source).

Risks and Mitigations

References

• GDPR Compliance Guidelines
• ASF Privacy Policy
• Telemetry Best Practices for Open Source
• Airflow 2.10.0 Release Notes - Scarf
• Previous discussion: [GitHub Issue #XXXXX]

Open Questions

1. Should we have different default behavior for development vs production environments (detected by database type or explicit flag)?
2. What is the process for enterprises to share aggregate data without enabling raw telemetry?
3. Should we offer a "telemetry lite" mode with even less data for highly privacy-sensitive users?
4. What specific noise levels (sensitivity parameters) provide the best balance between privacy and utility?
5. Should we implement a "trust score" showing telemetry participation rate to encourage adoption?

Conclusion

This AIP proposes a privacy-first, transparent, and community-governed telemetry system that rebuilds trust while providing the data Airflow maintainers need to make informed development decisions.

Key improvements over previous attempts:

• Explicit opt-in: Users must actively consent, rebuilding trust after Scarf incident
• Differential privacy: Noise added to all counts prevents fingerprinting
• No long-term tracking: Daily rotating session IDs prevent correlation across days
• Five control methods: Makes enabling or disabling telemetry easy
• Debug mode: Complete transparency into what's being sent
• Open source implementation: All code reviewable by privacy advocates
• Public legal commitment: Our telemetry promise is binding

By learning from the Scarf incident, implementing industry best practices for privacy protection, and prioritizing explicit user consent over data coverage, we can rebuild trust while still gathering valuable (though more limited) insights for project sustainability.

Discussion: https://github.com/apache/airflow/discussions/XXXXX
Vote Thread: TBD after discussion period