model-deployment
/model-deployment
Wdraża modele w trybie wsadowym, online lub na urządzeniach brzegowych z testami shadow, bramkowanymi wdrożeniami i wycofywaniem.
- Autor
- Amey-Thakur ↗
- Repozytorium
- AI-SKILLS ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 05.09.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 05.09.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Wdraża modele w trybie wsadowym, online lub na urządzeniach brzegowych z testami shadow, bramkowanymi wdrożeniami i wycofywaniem.
Oryginalny opis (EN)
Ship models through batch, online, or edge patterns with shadow testing, gated rollouts, and rollback. Use when moving a model to production or designing the serving architecture.
Jak działa
- Method
- Boundaries
Podgląd SKILL.md
--- name: model-deployment description: Ship models through batch, online, or edge patterns with shadow testing, gated rollouts, and rollback. Use when moving a model to production or designing the serving architecture. --- # Model deployment A model in production is software with an extra failure mode: it can be healthy and wrong. Deployment therefore adds statistical gates (shadow, canary against metrics) on top of ordinary release engineering. ## Method 1. **Choose the serving pattern by decision latency.** Batch scoring (nightly scores into a table) when decisions tolerate staleness: cheapest, testable like any pipeline (see data-pipeline-design). Online service (REST/gRPC) for request-time decisions, with p99 latency and throughput budgets set before building. Edge/in-process for offline-capable or ultra-low-latency paths, paying in update latency and fleet fragmentation (see mobile-release-strategy). 2. **Package the whole prediction function.** Model weights plus preprocessing, feature transformations, and postprocessing ship as one versioned artifact from the registry (see experiment-tracking); a model served without its exact pipeline is training/serving skew by construction (see feature-engineering). Containerize with pinned dependencies; validate the artifact answers golden requests correctly in CI before any environment sees it. 3. **Shadow before serving.** Run the candidate on live traffic without acting on outputs: compare latency, error rate, prediction distribution vs incumbent, and (where labels arrive fast) accuracy. Shadow catches the engineering skew and distribution surprises offline evaluation cannot (see model-evaluation boundaries), at zero user risk. 4. **Roll out with statistical gates.** Canary a small traffic slice; gates are business guardrails and prediction-health metrics (see drift-monitoring), not just HTTP 200s; where the model's value claim matters, the canary is a real experiment (see ab-test-design). Expand on evidence; halt criteria numeric and pre-agreed (the mobile-release-strategy discipline, applied to models). 5. **Keep rollback one switch away.** Previous model version stays warm and loadable; rollback is a config flip, tested in drills, not a retraining project. Model versions are immutable and addressable; "we retrained over it" means you have no rollback (see rollback-strategy). 6. **Instrument the model-specific telemetry.** Log inputs (sampled, PII-scrubbed), predictions, scores, model version per request; dashboards on prediction distribution, feature nulls, latency by version. This telemetry is what drift-monitoring and incident debugging consume; without it the model is a black box that pages you in business metrics. ## Boundaries - Retraining cadence and triggers are the drift-monitoring contract; deployment provides the safe path each retrain travels, not the schedule. - Multi-armed bandits and online learning change the statistical machinery entirely; this skill covers the fixed-model release loop. - LLM serving adds token streaming, caching, and cost tiers on top (see llm-cost-latency, inference-serving-optimization); the gating and rollback discipline transfers unchanged.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · Amey-Thakur/AI-SKILLS | 90 | MIT |
| secondsky/claude-skills | 100 | MIT |
| seb1n/awesome-ai-agent-skills | 100 | MIT |
| Harrypatria/data-mining-ml | 90 | MIT |
| awslabs/agent-plugins | 90 | Apache-2.0 |
| 420company/artemis | 85 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki b3b47535d5c3 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100model-serving-and-scaling
Prowadzi wdrożenie modeli ML i LLM do inferencji online: autoskalowanie, budżety opóźnień, batching, wdrożenia canary i shadow oraz kontrolę kosztów, a także wybór między serwowaniem wsadowym, online i strumieniowym. - 90ml-model-serving
Wdraża modele do wnioskowania: TorchServe, BentoML, Ray Serve, KServe, wdrożenia canary, autoskalowanie, wersjonowanie i wycofywanie. - 90testing-in-production
Weryfikuje zmiany na prawdziwym ruchu bez ryzyka dla użytkowników przez canary, shadow traffic i feature gates z automatycznym wycofaniem, gdy samo staging nie wystarcza. - 100vercel-deploy-integration
Zarządza produkcyjnymi wdrożeniami Vercel: promocja, rollback, strategie wieloregionowe i wzorzec blue-green. - 100ship
Wdraża stronę lub aplikację w infrastrukturze online: jeden kod e-mail, adres HTTPS, monitoring i potrzebne usługi. - 100mlops-engineer
Obsługuje MLOps: śledzenie eksperymentów w MLflow lub Weights & Biases, magazyny cech z wykrywaniem wycieków, serwowanie przez FastAPI i Kubernetes, monitorowanie dryfu oraz wdrożenia A/B w trybie cienia i canary.