Remove pod-affinity rules; rely on RWO PVC for co-location

[t0mdavid-m]

Summary

Simplify pod co-location by removing explicit pod-affinity rules and relying instead on Kubernetes' native VolumeBinding scheduler plugin, which automatically pins pods that mount the same ReadWriteOnce PVC to the same node.

Changes

• Remove pod-affinity rules from streamlit-deployment.yaml, rq-worker-deployment.yaml, and cleanup-cronjob.yaml
• Remove volume-group: workspaces labels from all three deployments (no longer needed as affinity selectors)
• Update documentation to explain the new co-location mechanism:
  • Pod co-location is now enforced by the shared RWO PVC mount, not by explicit affinity rules
  • Clarify that NodeSelector picks the eligible node set, and the RWO mount picks the specific node within that set
  • Emphasize that co-location applies within a fork (shared PVC), not across forks (each fork has its own PVC)
  • Update deployment descriptions to reference the RWO mount instead of the affinity label
• Update skill guide to remove references to the volume-group label and pod-affinity rule

Implementation Details

The Kubernetes scheduler's VolumeBinding plugin automatically ensures that once a ReadWriteOnce PVC is attached to a node, all subsequent pods mounting that PVC are scheduled to the same node. This eliminates the need for manual pod-affinity configuration while achieving the same co-location guarantee. The change reduces manifest complexity and makes the scheduling constraint more explicit and maintainable.

https://claude.ai/code/session_01HLxsvLLznn5WV42iGHBxiP

Summary by CodeRabbit

• Documentation

  • Updated Kubernetes deployment documentation to clarify current pod co-location mechanisms.

• Chores

  • Simplified Kubernetes configuration by removing deprecated pod affinity constraints from multiple deployment manifests.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR updates Kubernetes deployment manifests and documentation to replace pod-affinity scheduling with storage-based pod co-location. The volume-group labels and associated podAffinity rules are removed from four K8s manifests, with documentation clarifying that ReadWriteOnce PVC attachment now enforces node pinning instead.

Changes

Cohort / File(s)	Summary
Documentation .claude/skills/configure-k8s-deployment.md, docs/kubernetes-deployment.md	Updated co-location mechanism explanation from hostname-based pod affinity to RWO PVC-based storage enforcement. Removed references to volume-group labeling and podAffinity rules; clarified that shared per-fork PVC mounts ensure workspace pod co-location within Kubernetes' VolumeBinding.
Kubernetes Manifests k8s/base/cleanup-cronjob.yaml, k8s/base/rq-worker-deployment.yaml, k8s/base/streamlit-deployment.yaml	Removed volume-group: workspaces pod labels and podAffinity scheduling rules that enforced hostname-level co-scheduling. Container specs, volumes, environment variables, and job scheduling behavior remain unchanged.

Possibly related PRs

• Add Kubernetes deployment docs and refactor Claude skills #362: Both PRs directly modify the same Kubernetes deployment documentation and base manifests around workspace co-location; the earlier PR documented pod-affinity/volume-group rationale, while this PR removes those mechanisms in favor of PVC-based co-location.
• Refactor K8s deployment to use memory-tier components #375: Both PRs modify the same streamlit-deployment.yaml and rq-worker-deployment.yaml manifests; while this PR removes volume-group labels and podAffinity rules, the earlier PR modified resource specifications and introduced nodeSelector patches.

Poem

🐰 Pod affinity labels fade away,
PVC storage saves the day!
Kubernetes binds each fork in place,
Through shared volumes, not label race.
Simpler schedules, cleaner state,
Co-location feels just great!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately and concisely summarizes the main change: removing pod-affinity rules and relying on RWO PVC for co-location, which directly aligns with all modified files and documentation updates.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/fix-pod-affinity-labels-YvnlN

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

docs/kubernetes-deployment.md (1)

93-93: ⚠️ Potential issue | 🟠 Major

Remove stale “Pod affinity exists …” statement; it now conflicts with this PR.

Line 93 says pod affinity exists for warm WebSocket/cache behavior, but Line 83 states there is no pod-affinity rule. Please reconcile this to a single source of truth (RWO-PVC-based co-location only).

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@docs/kubernetes-deployment.md` at line 93, Remove or update the stale
sentence "Pod affinity exists to keep the WebSocket warm and reuse Streamlit's
in-process script cache" so the doc consistently states that there is no
pod-affinity rule and co-location is achieved only via the RWO-PVC (shared
workspace PVC) and Redis; ensure the paragraph about sticky `stroute` cookies
remains and replace the pod-affinity claim with a brief note that pod affinity
is not used and RWO-PVC-based co-location is the source of shared state.

🧹 Nitpick comments (1)

.claude/skills/configure-k8s-deployment.md (1)

32-33: Broaden the recon check to all PVC consumers, not just Streamlit.

This check currently validates claimName only in k8s/base/streamlit-deployment.yaml; please also require matching claimName: workspaces-pvc in k8s/base/rq-worker-deployment.yaml (and cleanup job if present) so drift cannot silently break co-location assumptions.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In @.claude/skills/configure-k8s-deployment.md around lines 32 - 33, The recon
check currently only validates claimName in k8s/base/streamlit-deployment.yaml;
extend it to validate the same claimName: workspaces-pvc in all PVC-consuming
manifests (specifically k8s/base/rq-worker-deployment.yaml and any cleanup Job
manifest), i.e., update the check logic that looks for claimName to scan both
Deployment and Job resources and assert claimName === "workspaces-pvc" for each
volume/volumeMount reference; also update tests/assertions for the recon check
to include examples of rq-worker-deployment.yaml and the cleanup job so drift
cannot silently break co-location assumptions.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@docs/kubernetes-deployment.md`:
- Line 93: Remove or update the stale sentence "Pod affinity exists to keep the
WebSocket warm and reuse Streamlit's in-process script cache" so the doc
consistently states that there is no pod-affinity rule and co-location is
achieved only via the RWO-PVC (shared workspace PVC) and Redis; ensure the
paragraph about sticky `stroute` cookies remains and replace the pod-affinity
claim with a brief note that pod affinity is not used and RWO-PVC-based
co-location is the source of shared state.

---

Nitpick comments:
In @.claude/skills/configure-k8s-deployment.md:
- Around line 32-33: The recon check currently only validates claimName in
k8s/base/streamlit-deployment.yaml; extend it to validate the same claimName:
workspaces-pvc in all PVC-consuming manifests (specifically
k8s/base/rq-worker-deployment.yaml and any cleanup Job manifest), i.e., update
the check logic that looks for claimName to scan both Deployment and Job
resources and assert claimName === "workspaces-pvc" for each volume/volumeMount
reference; also update tests/assertions for the recon check to include examples
of rq-worker-deployment.yaml and the cleanup job so drift cannot silently break
co-location assumptions.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 857f910e-46dd-49cc-a173-ca22b453b46e

📥 Commits

Reviewing files that changed from the base of the PR and between 8e096ac and 8d76292.

📒 Files selected for processing (5)

• .claude/skills/configure-k8s-deployment.md
• docs/kubernetes-deployment.md
• k8s/base/cleanup-cronjob.yaml
• k8s/base/rq-worker-deployment.yaml
• k8s/base/streamlit-deployment.yaml

💤 Files with no reviewable changes (3)

• k8s/base/streamlit-deployment.yaml
• k8s/base/rq-worker-deployment.yaml
• k8s/base/cleanup-cronjob.yaml