Skip to content

Calibration distributed patterns - #416

Open
jonathanhhb wants to merge 3 commits into
laser-base:mainfrom
jonathanhhb:calibration-distributed-patterns
Open

Calibration distributed patterns#416
jonathanhhb wants to merge 3 commits into
laser-base:mainfrom
jonathanhhb:calibration-distributed-patterns

Conversation

@jonathanhhb

Copy link
Copy Markdown
Collaborator

No description provided.

Jonathan Bloedow and others added 2 commits June 16, 2026 12:34
Adds a 7-subsection block after Next Steps covering patterns,
anti-patterns, and deployment heuristics for distributed Optuna
calibration of LASER models on Kubernetes with a shared MySQL
backend. Drawn from one extended deployment cycle but written
project-neutrally — no references to specific projects, clusters,
node-pool names, or IP addresses.

Subsections:
  - Kubernetes Deployment Patterns (nodeSelector + toleration;
    enableServiceLinks: false; verify placement on first deploy)
  - Container Build Patterns (wheel-based install; -march=x86-64-v2
    for cluster images; PyMySQL; staged data assets; shell-glob
    extras trap)
  - Shared Optuna Backend Patterns (which API calls are safe at
    scale on a long-history shared backend; local-vs-cluster
    Optuna version skew)
  - Probe-and-Diagnose Workflow (STAGE_MODULE env-var scaffold;
    in-cluster probe Pod pattern; automated ditch-detection
    heuristics)
  - Calibration Objective Design (don't calibrate to pooled
    summary stats alone; constrain physically meaningful
    quantities; add mechanism only after demonstrated failure)
  - Identifiability Workflow (1D sweep before widening; S/N
    interpretation; edge minima caveats)
  - Pre-Deployment Checklist (one-line questions mapped to each
    pattern above)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Two new subsections at the top of "Distributed Calibration Patterns",
addressing users new to running on a managed Kubernetes cluster.
Neither subsection re-documents Kubernetes or any specific cloud
provider; they scope what to ask the cluster admin for and how to
catch deployment bugs before reaching the cluster.

Assumptions and Cluster Handoff:
  Lists what the patterns section assumes you already have (working
  local calibration, working local Docker, kubectl, KUBECONFIG,
  cluster network access) and what it deliberately does not require
  you to learn first (Kubernetes internals, vendor-specific admin
  concepts). Provides a one-page list of questions to put to the
  cluster admin in a single 15-minute handoff: kubeconfig, registry
  URL + imagePullSecret, simulation-pool nodeSelector + toleration,
  MySQL Secret name + keys, network restrictions, namespace
  quotas, autoscaling rules.

Validate Locally with Docker Before Pushing to the Cluster:
  Codifies the rule that the cluster is a scaling environment,
  not a debugging environment. Provides a docker-compose template
  for a three-container local stand (MySQL + study-creator +
  worker) that runs the exact image that would be pushed to the
  cluster. Lists which classes of failure the local stand catches
  (build context, wheel-build errors, shell-glob extras trap,
  data-asset paths, native-extension crashes, MySQL URL bugs,
  Optuna round-trip failures, stage-module interface mismatches,
  trial completion failures) and which it does not (node-pool
  placement, shared-MySQL-at-scale issues, Optuna version skew,
  Kubernetes Pod-spec quirks).

Pre-deployment checklist updated with two new items at the top
covering cluster handoff and local validation, before the original
node-pool / service-links / driver / native-extensions / build /
data-paths / Optuna-API / version-parity / scaffold / objective /
identifiability items.

Section intro paragraph updated: "split into nine categories"
(was seven), with the first two flagged as the cheapest investments.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings June 16, 2026 19:35

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds a new documentation section to guide running distributed Optuna-based calibrations of LASER models on Kubernetes with a shared MySQL backend, capturing operational patterns, anti-patterns, and workflow heuristics.

Changes:

  • Added “Distributed Calibration Patterns” guidance covering cluster handoff, local Docker validation, Kubernetes scheduling, build practices, shared-DB scalability pitfalls, and probe workflows.
  • Added objective-design and identifiability workflow recommendations plus a pre-deployment checklist.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread docs/calibration.rst Outdated
Comment thread docs/calibration.rst
Comment thread docs/calibration.rst Outdated
Comment thread docs/calibration.rst
Comment thread docs/calibration.rst Outdated
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants