Hello all ,

*Proposal*:
  I'd like to propose adding operational agent skills to the Apache Flink
ecosystem, which would include structured, action-oriented guides that help
AI coding assistants (and humans) troubleshoot, debug, tune, and deploy
Flink applications.

*What exists today*:
So far there have been several amazing tickets for AI-assisted
contributions:

   - FLINK-39477 Clarify how Flink accepts and supports AI-Assisted
   Contributions and AI Tooling Support
   - FLINK-39714 Backport AGENTS.md to active branches 1.20, 2.0, and 2.2
   - FLINK-39529 Extend planner's Agents.md with more detailed SQL to
   operator lifecycle
   - FLINK-39881 [FKO] Adopt AI-assisted contribution guidelines for the
   Kubernetes Operator
   - FLINK-39539 Extend planner's Agents.md with more detailed information
   for our internal table module CDC processing

For operations on AWS, the aws/agent-toolkit-for-aws repo includes detailed
Flink skills covering checkpoint tuning, first-fault isolation, state
management, but these are specific to AWS Managed Service for Apache Flink
(KPU sizing, CloudWatch metrics, aws CLI commands).

*What's missing*: provider-agnostic operational skills for the many teams
running Flink on self-managed Kubernetes, bare-metal, or non-AWS clouds.
Flink's official docs (docs/ops/) cover the topics well as reference
material, but they aren't structured as concise
symptom→diagnosis→resolution guides optimized for AI agent consumption.

*More Detailed Proposal*
A set of markdown-based operational skills, each one a self-contained guide
following a consistent structure:

   -  Symptoms:  what the operator sees (metrics, logs, behavior)
   -  Diagnosis steps:  how to narrow down the root cause
   -  Resolution patterns:  concrete fixes with commands/config
   -  Common pitfalls:  tribal knowledge that isn't obvious from docs


Example topics:

   1. Restart loop diagnosis:  "recent logs show side effects, not root
   cause; find attempt #0's first exception"
   2. Classloader conflicts:  parent-first vs child-first, shading, version
   mismatches (e.g., NoSuchFieldError from AWS SDK conflicts)
   3. Checkpoint failure triage:  alignment timeout vs state size vs slow
   sink vs backpressure
   4. State migration / skip-state:  when it's safe, how to do it, what
   breaks
   5. Backpressure analysis:  interpreting metrics, identifying
   bottlenecks, tuning
   6. Memory tuning:  TaskManager/JobManager OOM diagnosis, RocksDB memory
   accounting

Where the skills could live:

   -   A skills/ or ops-skills/ directory within apache/flink
   -   A dedicated repo (like apache/flink-agent-skills)
   -   Within the existing docs/ alongside module-level AGENTS.md files


These would be provider-agnostic (no AWS/GCP/Azure specifics) and
complement the existing official documentation rather than replace it.

I'm happy to contribute initial skills based on real operational patterns.
The classloader conflict and restart-loop examples above come from
production issues I've debugged in the past week.

Looking forward to the community's thoughts.

Best regards,
James Pizagno (github repo <https://github.com/jpizagno>)

Reply via email to