Production Engineering: Reproducibility, CI/CD for ML, Versioning ve Deployment Patterns

Ders Haritası#

1. ML CI/CD ile classical SDE CI/CD farkı
2. Reproducibility temelleri — pinned dependencies, lock files
3. Code + data + model versioning
4. MLflow experiment tracking
5. DVC data versioning
6. HuggingFace Hub model registry
7. Eval harness CI'a integration
8. Canary deployment strategies
9. A/B testing for ML
10. Drift monitoring + auto-rollback
11. KVKK uyumlu deployment
12. Part I kapanışı + Part II'ye geçiş

1. ML CI/CD ile Classical SDE CI/CD Farkı#

Klasik SDE CI/CD#

Code → Lint → Unit test → Build → Deploy

ML CI/CD#

Code change → ?
Data change → ?
Model change → ?
Prompt change → ?

Trigger matrix#

Modern öneri#

• Small change → fast tests (5 min)
• Big change → full eval (1-3 hour)

2. Reproducibility Temelleri#

Katman 1: Code#

• Git commit hash her artifact ile birleşik save
• Branch policy: main locked, PR gerekli
• Tag releases:
  v1.2.3
  ile checkpoint

Katman 2: Dependencies#

• Pinned versions:
  requirements.txt
  exact versions
• Lock file:
  poetry.lock
  ,
  uv.lock
  ,
  requirements-lock.txt
• Container: Dockerfile pin everything (CUDA version, PyTorch, OS)

# requirements-lock.txt
torch==2.5.1+cu124
transformers==4.46.2
deepspeed==0.15.0

Katman 3: Data#

• Dataset version: HuggingFace dataset hash
• Snapshot date: data fetched on YYYY-MM-DD
• DVC tracking: git-like for big data (aşağıda)

Katman 4: Hardware/Random#

• Seed set: numpy, torch, cuda all
• Deterministic mode: optional but available
• Hardware spec: documented (GPU model, driver, CUDA)

Production reproducibility report#

{
  "run_id": "uuid",
  "timestamp": "2026-05-12T10:00:00Z",
  "git_commit": "abc123def",
  "git_branch": "main",
  "dataset_hash": "sha256:...",
  "config_hash": "sha256:...",
  "hardware": "8x H100 80GB",
  "torch_version": "2.5.1+cu124",
  "deepspeed_version": "0.15.0",
  "final_loss": 1.85,
  "checkpoint_uri": "s3://...",
  "eval_results": {...}
}

3. Code + Data + Model Versioning#

Code: Git (standart)#

src/
  train.py
  eval.py
configs/
  llama_3_8b.yaml
  llama_3_70b.yaml
.gitignore  # data, checkpoints

Data: DVC veya HuggingFace Datasets#

dvc init
dvc remote add origin s3://bucket/dvc-store
dvc add data/turkish_corpus_v1.parquet
git add data/turkish_corpus_v1.parquet.dvc
git commit -m "Add Turkish corpus v1"
dvc push  # Upload to S3

from datasets import Dataset
ds = Dataset.from_parquet("...")
ds.push_to_hub("sukruyusufkaya/turkce-egitim-v1")

Model: HuggingFace Hub + MLflow#

model.push_to_hub("sukruyusufkaya/llama3-tr-v1.0")

import mlflow

with mlflow.start_run():
    mlflow.log_param("lr", 3e-4)
    mlflow.log_metric("loss", 1.85)
    mlflow.pytorch.log_model(model, "model")

Prompt versioning#

# prompts/customer_support_v3.yaml
version: 3
system_prompt: |
  Sen Trendyol müşteri destek asistanısın...
created: 2026-04-15
author: ali.veli
changelog: |
  v3: Türkçe argo handling iyileştirildi

4. MLflow — Experiment Tracking#

Setup#

pip install mlflow
mlflow server --host 0.0.0.0 --port 5000  # production deployment

Usage#

import mlflow

mlflow.set_tracking_uri("http://mlflow-server:5000")
mlflow.set_experiment("llama3-tr-finetune")

with mlflow.start_run(run_name="v1.0-baseline"):
    # Params
    mlflow.log_params({
        "model": "Llama-3.1-8B",
        "lr": 3e-4,
        "batch_size": 32,
        "epochs": 3,
    })

    # Metrics (each step)
    for epoch in range(3):
        for step, batch in enumerate(train_loader):
            loss = train_step(batch)
            mlflow.log_metric("loss", loss, step=step + epoch*len(train_loader))

    # Artifact: model
    mlflow.pytorch.log_model(
        model,
        artifact_path="model",
        registered_model_name="llama3-tr",
    )

    # Custom artifact: eval result
    with open("eval_results.json", "w") as f:
        json.dump(eval_results, f)
    mlflow.log_artifact("eval_results.json")

Features#

• UI: experiments comparison, plot metrics, model artifacts
• Model Registry: model lifecycle (Staging, Production, Archived)
• API: programmatic access
• Deployment integration:
  mlflow.pytorch.load_model()
  direkt prod

Türk şirketler#

5. DVC — Data Version Control#

Workflow#

# Init
git init
dvc init

# Add remote
dvc remote add -d myremote s3://mybucket/dvc-store

# Add data
dvc add data/turkish_corpus_50gb.parquet
# DVC creates: data/turkish_corpus_50gb.parquet.dvc (small, git-trackable)
# Actual file in .dvc/cache/

git add data/turkish_corpus_50gb.parquet.dvc .gitignore
git commit -m "Add corpus v1"
dvc push  # Upload to S3

# Update data
dvc add data/turkish_corpus_50gb.parquet  # detect changes
git commit -am "Update corpus to v2"
dvc push

# Time travel
git checkout v1-tag
dvc pull  # download v1 data

Pipelines#

# dvc.yaml
stages:
  preprocess:
    cmd: python preprocess.py
    deps:
      - data/raw.parquet
    outs:
      - data/processed.parquet

  train:
    cmd: python train.py
    deps:
      - data/processed.parquet
      - src/train.py
    outs:
      - models/checkpoint.pt

Benefits#

• Git-like data: branch, diff, merge
• Lazy fetch: sadece gerekli data download
• Pipeline caching: incremental updates
• Reproducibility: code + data history aligned

Türk perspektif#

6. HuggingFace Hub — Model Registry#

Push#

from transformers import AutoModelForCausalLM

# Fine-tuned model'i hub'a push
model.push_to_hub(
    "sukruyusufkaya/llama3-tr-customer-support-v1",
    use_auth_token=True,
    commit_message="Initial release: v1.0 fine-tuned on customer support data",
)

Versioning#

model = AutoModelForCausalLM.from_pretrained(
    "sukruyusufkaya/llama3-tr-customer-support-v1",
    revision="abc123def",  # specific git hash
)

Model card#

---
license: apache-2.0
language:
  - tr
base_model: meta-llama/Llama-3.1-8B
tags:
  - turkish
  - customer-support
---

# Llama 3 Türkçe Customer Support v1.0

## Training data
50K customer-support conversations, Türkçe.

## Eval metrics
- TR-MMLU: 62%
- Customer-support accuracy: 87%

## Limitations
- ...

Private models#

model.push_to_hub(repo_id, private=True)  # team-only

Comparison to MLflow#

• MLflow: experiment tracking + on-prem registry
• HF Hub: artifact-focused, public sharing, integrated with transformers/datasets

7. Eval Harness CI'a Integration#

Eval suite tasarımı#

# tests/eval/test_model.py
import pytest
from my_model import MyModel
from eval_harness import EvalSuite

@pytest.mark.eval
def test_model_accuracy():
    model = MyModel.from_pretrained("path/to/checkpoint")
    suite = EvalSuite(
        benchmarks=["mmlu", "gsm8k", "humaneval"],
        languages=["en", "tr"],
    )
    results = suite.run(model)

    # Thresholds (regression prevention)
    assert results["mmlu"] > 0.60, f"MMLU regression: {results['mmlu']}"
    assert results["gsm8k"] > 0.45, f"GSM8K regression: {results['gsm8k']}"

CI integration#

# .github/workflows/ml-ci.yml
name: ML CI
on:
  pull_request:
    paths:
      - 'src/**'
      - 'configs/**'

jobs:
  fast-eval:
    runs-on: gpu-runner
    steps:
      - uses: actions/checkout@v4
      - run: pip install -r requirements.txt
      - run: pytest tests/eval/fast/ -m eval
        # 5-10 min, basic regression check

  full-eval:
    runs-on: gpu-runner
    if: github.event.pull_request.labels.*.name == 'full-eval'
    steps:
      - run: pytest tests/eval/full/ -m eval
        # 1-3 hour, full benchmark suite

Türkçe için custom eval#

# Türkçe-specific eval
turkish_benchmarks = ["tr-mmlu", "tr-gsm", "trendyol-customer-support-test"]
suite = EvalSuite(benchmarks=turkish_benchmarks)

Benchmark regression matrix#

• Last 30 PR
• Trend up/flat/down
• Regression alert (>%2 drop)

8. Canary Deployment — Gradual Rollout#

Strategy#

Day 0: %0 new model, %100 old (current state)
Day 1: %5 new, %95 old (canary)
Day 2: %25 new, %75 old (if metrics OK)
Day 3: %50 new, %50 old
Day 4: %100 new (full rollout)

Implementation patterns#

Pattern 1: Feature flag

def get_model(user_id):
    if feature_flag.is_enabled("new_model_canary", user_id):
        return new_model
    return old_model

Pattern 2: Load balancer routing

# kubernetes
spec:
  rules:
    - host: api.example.com
      http:
        paths:
          - backend:
              service: new-model-service
              weight: 5  # %5
          - backend:
              service: old-model-service
              weight: 95

Pattern 3: Sticky sessions

Metrics to monitor#

• Error rate: yeni model crash mi?
• Latency p99: yavaş mı?
• User satisfaction: thumbs down up?
• Cost: token consumption?
• Quality: hallucination, accuracy

Auto-rollback#

if error_rate > 2% AND error_rate_baseline < 0.5%:
    rollback_to_old_model()
    alert_team()

9. A/B Testing for ML#

Setup#

• Variant A: old model
• Variant B: new model
• Random assignment: %50-50 (or different split)
• Metric: success criterion (accuracy, NPS, conversion)

Statistical considerations#

Sample size

N = ceil(2 × (Z_α + Z_β)^2 × σ^2 / Δ^2)

α=0.05, β=0.2, σ=0.1 (std), Δ=0.02 (effect)
N ≈ 200 per variant

P-value

Bayesian alternative

Pratik framework#

from ab_testing import Experiment

exp = Experiment(
    name="prompt-v3-vs-v2",
    variants={"control": prompt_v2, "treatment": prompt_v3},
    metric="user_satisfaction",
    target_n=1000,
)

exp.start()
# Real traffic split

Common pitfalls#

1. Peeking: sample size'a ulaşmadan karar verme
2. Multiple testing: birçok metric → false positive
3. Novelty effect: yeni şey ilk gün ilginç, sonra erozyon
4. Network effects: variants etkileşiyor mu?

Türk perspektif#

10. Drift Monitoring + Auto-Rollback#

Drift types#

1. Data drift

2. Concept drift

3. Performance drift

Monitoring stack#

Input → [model] → Output
   ↓                ↓
   Drift detector  Performance monitor
   ↓                ↓
   Alert + retrain trigger

Auto-rollback triggers#

• Error rate > 2%: immediate rollback
• Latency p99 > 5s: rollback
• User feedback negative > 30%: rollback within 1 hour
• Hallucination rate > 5%: rollback (Modül 56)

Production tooling#

• Langfuse (Modül 48): LLM-specific drift detection
• Evidently AI: ML model drift
• Arize: production ML observability
• Prometheus + custom rules: low-level monitoring

11. KVKK Uyumlu Deployment#

Data residency#

Production stack:
- LLM training: cloud GPU (AWS Frankfurt, GCP Belgium) — uygun coğrafya
- LLM inference: Türkiye'de hosted (Aselsan, BiSU, on-prem)
- Customer data: Türkiye DB

Audit trail#

• User ID (anonymized OK)
• Timestamp
• Input (PII-stripped veya encrypted)
• Output
• Model version
• Cost
• Latency

Right to be forgotten#

• Training data'dan removable mı?
• Model memorize ettiyse (Modül 4.8) — challenging
• Output cache'lerden silmek mümkün

Consent management#

• LLM'le konuşma: explicit consent (T&C)
• Data fine-tuning'e gidiyorsa: explicit + ayrı consent
• Audit log: consent timestamp + version

Production checklist#

• Data residency Turkey
• Audit logs PII-redacted
• Right-to-forget pipeline
• Consent management system
• Encryption at rest + in transit
• Access control (RBAC)
• Incident response plan
• DPO (Veri Sorumlusu) assignment
• Periodic privacy audit

12. Part I Kapanışı + Part II'ye Geçiş#

🎉 Part I Tamamlandı!#

Ne öğrendin?#

• Matematik: lineer cebir, calculus, olasılık, bilgi teorisi, optimization
• Programming: PyTorch internals, custom autograd, Triton kernels
• Sistem: distributed training, memory profiling, debugging
• Kavramsal: LLM'in olasılıksal modeli, scaling laws, emergent capabilities
• Tarih: 70 yıllık AI evrim, paradigma değişimleri
• Production: CI/CD, versioning, deployment, KVKK

Sıradaki: Part II — Transformer Mimarisinin İskeleti#

• Modül 6: Tokenization Mikro-Cerrahisi
• Modül 7: Embedding'ler — Anlamın Geometrisi
• Modül 8: Attention'ın Matematiksel İnşası
• Modül 9: Position Encoding — RoPE, ALiBi
• Modül 10: Transformer Bloğunu Sıfırdan
• Modül 11: Modern Mimari Aileleri (Llama, Qwen, DeepSeek)
• Modül 12: Mixture of Experts (MoE)
• Modül 13: Transformer Alternatifleri (Mamba, RWKV)

Senin işin Part I sonrası#

1. Kendi mini LLM'in çalışıyor olmalı (Modül 2 capstone)
2. Production-ready PyTorch stack hâkim
3. Bir paper okuyup anlamalısın (özellikle FlashAttention, Llama, DeepSeek-V3)
4. Yapılandırılmış debug workflow'un olmalı
5. CI/CD pipeline kurabilmelisin

13. Mini Egzersizler#

1. Reproducibility test: 6 ay önce çalışan bir LLM training run'ını rerun. Aynı sonucu nasıl alırsın?
2. CI design: Türk e-ticaret şirketi için fine-tune workflow. CI pipeline tasarla.
3. Canary metrics: 100K req/day LLM service'a yeni model deploy. İlk 24 saatte hangi metric'leri izle?
4. Drift example: Türkçe customer support model 6 ay sonra accuracy %5 düştü. Olası sebep + çözüm?
5. KVKK incident: Kullanıcı PII LLM training'e girdi. İlk 24 saat aksiyon planı?

Bu Derste Neler Öğrendik?#

🎉 Part I Tamamlandı!#

Sıradaki Bölüm#