mlops

name	mlops
description	MLOps and model deployment.

Activate When

• /godmode:mlops, "deploy model", "model serving"
• "model drift", "retrain", "A/B test models"
• Trained model ready for production deployment

Workflow

1. Model Readiness

Model: <name and version>
Source: EXP-<ID>
Checklist:
  [ ] Evaluation complete (test metrics documented)
  [ ] Bias/fairness check passed
  [ ] Artifacts saved (weights, config, preprocessor)
  [ ] Input/output schema documented
  [ ] Latency benchmarked (< target p99 ms)
  [ ] Size acceptable (< N MB)

IF latency p99 > 100ms: apply optimization. IF model size > 500MB: consider distillation/pruning.

2. Serving Infrastructure

Options:
  TF Serving: TensorFlow models, gRPC/REST
  Triton: multi-framework, ONNX/TensorRT
  SageMaker: managed AWS, auto-scaling
  FastAPI/Ray Serve: custom, flexible

# Check for serving frameworks
pip list | grep -iE "fastapi|ray|triton|sagemaker"
ls model_repository/ serve/ 2>/dev/null

3. Inference Optimization

| Optimization | Latency | Size | Accuracy |
| Baseline FP32 | <ms> | <MB> | <val> |
| FP16 quant | <ms> | <MB> | <val> |
| INT8 quant | <ms> | <MB> | <val> |
| ONNX | <ms> | <MB> | <val> |
| Distillation | <ms> | <MB> | <val> |

IF accuracy drop > 1% from quantization: use FP16 only. IF latency target not met: try TensorRT or distillation.

Batching: static (fixed workload), dynamic (variable traffic, max_queue_delay_ms), adaptive (auto-tune).

4. Model Versioning

| Version | Metric | Status | Traffic |
| v3.1 | F1=0.891 | CHAMPION | 90% |
| v3.2 | F1=0.903 | CANARY | 10% |
| v3.0 | F1=0.879 | ARCHIVED | 0% |
Lifecycle: STAGED->CANARY->CHAMPION->ARCHIVED

5. A/B Testing

Champion: v<N>  Challenger: v<N>
Split: <champion%>/<challenger%>
Routing: random|user-hash|feature-flag
Duration: <minimum days>
Sample size: <minimum per variant>
Success: primary metric >= <threshold> improvement
Guardrails: latency p99, error rate, business KPIs

IF p-value > 0.05 after min samples: no winner. IF guardrail regresses > 2%: stop test, revert.

6. Drift Detection

Feature drift (PSI):
  < 0.1: no drift
  0.1-0.2: moderate — monitor closely
  > 0.2: significant — trigger retraining
Performance:
  < 2% drop: normal variance
  2-5% drop: warning — schedule review
  > 5% drop: alert — trigger retraining

7. Retraining

Trigger: scheduled|drift-based|performance-based
Frequency: daily|weekly|monthly
Data window: last N days
Auto_deploy: false (requires A/B or human gate)
Cooldown: minimum time between retraining runs

8. Monitoring Dashboard

Requests/sec: <current> (avg/peak)
Latency p50/p95/p99: <ms>/<ms>/<ms>
Error rate: <pct>
Primary metric (7d rolling): <val>
Drift status: NONE|LOW|MODERATE|HIGH

Hard Rules

1. NEVER deploy without readiness checklist complete.
2. NEVER 100% traffic to new model on day one. Start shadow->5% canary->ramp.
3. ALWAYS keep previous champion ready for rollback.
4. NEVER auto-promote without validation gate.
5. Rollback must take < 5 minutes.

TSV Logging

Append .godmode/mlops-results.tsv:

timestamp	model	version	action	latency_p99	status

Keep/Discard

KEEP if: metrics improve AND no guardrail regression
  AND pipeline runs end-to-end.
DISCARD if: metrics regress OR pipeline fails.
  Revert and log reason.

Stop Conditions

STOP when FIRST of:
  - Model stable at 100% for 24h
  - Drift monitoring configured
  - Rollback tested < 5 min

Autonomous Operation

On failure: git reset --hard HEAD~1. Never pause.

Error Recovery

Failure	Action
OOM during training	Resume checkpoint, reduce batch
Performance degrades	Check drift, trigger retrain
A/B no difference	Verify sample size, document null