English | Français
amlops compiles a short, variability-aware description of an MLOps pipeline into
Infrastructure- and Platform-as-Code (Terraform, Argo Workflows / Argo Events,
Kubernetes, CI), places its steps across several cloud providers under cost /
carbon / latency / sovereignty objectives, and adapts it at run time (drift
detection, executable coordination contracts, re-placement).
It is the GetCaaS contribution to the ANR project AdaptiveMLOps (ANR-24-IAS2-0004) and builds on the consortium's results: the six variability categories and requirements R1–R8 of Toward Adaptive MLOps: Variability Mapping and Modeling (GdR GPL 2025, hal-05127859) and the 38 activities and coordination contracts of SkeltyMLOps (MLOps@ECAI 2025, hal-05337700; FGCS 185 (2026) 108700).
Status: research prototype, v0.1. The provider catalogue (
src/amlops/knowledge/providers.yaml) contains illustrative prices, power draws and grid intensities, and the dynamic experiments use synthetic traces. Generated Terraform/Kubernetes artefacts are parsed and unit-tested but have not been deployed. The feature model and rules are a proposal awaiting consortium review.
| Title | MLOps Adaptatif (AdaptiveMLOps) |
| Funding | French National Research Agency (ANR), grant ANR-24-IAS2-0004 |
| Call | AAP 2024 Thématiques Spécifiques en Intelligence Artificielle (TSIA): Machine Learning Operations, Software Engineering for AI |
| ANR contribution | €489,646 |
| Start / duration | September 2024 / 48 months |
| Coordinator | Sylvain Vauttier (EuroMov Digital Health in Motion) |
Objective. MLOps extends DevOps principles to data science and machine learning so that AI models are trained, deployed and operated as ordinary software components. A key issue is the continuous training of models to adapt them to changes observed in production data (data and concept drift). AdaptiveMLOps studies how domain-engineering concepts (feature models, software product lines) can capture the commonalities of MLOps processes and document best practices, and uses this knowledge to guide the design of new pipelines through a model-driven approach: a Domain Specific Language that is (i) generic and extensible, (ii) abstract enough for non-expert users, (iii) open to fine-tuning by experts, and (iv) pivotal to generate Platform-as-Code / Infrastructure-as-Code. The project targets automatic and dynamic (re)deployment of pipeline components hosted by several providers, optimising efficiency, cost and environmental footprint. Proposals are prototyped and validated through proofs of concept on the industrial partner's cloud platform.
Consortium.
| Partner | Role | Website |
|---|---|---|
| EuroMov Digital Health in Motion (EuroMov DHM), Université de Montpellier & IMT Mines Alès | Coordinator | https://dhm.euromov.eu/ |
| LIRMM, Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier (Université de Montpellier, CNRS) | Academic partner | https://www.lirmm.fr/ |
| GetCaaS | Industrial partner (this repository) | https://www.getcaas.io/ |
Official links.
- ANR project page: https://anr.fr/Projet-ANR-24-IAS2-0004
- Project website: https://adaptivemlops.wp.imt.fr/
- IMT Mines Alès: https://www.imt-mines-ales.fr/, Université de Montpellier: https://www.umontpellier.fr/, CNRS: https://www.cnrs.fr/
Consortium publications this work builds on.
- C. El Hatimi et al., Toward Adaptive MLOps: Variability Mapping and Modeling, GdR GPL national days, Pau, 2025. https://hal.science/hal-05127859
- C. Daoud et al., SkeltyMLOps: Orchestrating Collaborative MLOps Activities, MLOps25 @ ECAI 2025, CEUR-WS vol. 4109. https://hal.science/hal-05337700
- C. Daoud et al., A reference architecture for an orchestrated collaborative MLOps, Future Generation Computer Systems 185 (2026) 108700. https://doi.org/10.1016/j.future.2026.108700
flowchart LR
M["DSL model<br/>(5 lines for a non-expert)"] --> C["Feature model<br/>completion and advice"]
C --> D["Derivation<br/>(process line)"]
D --> P["Multi-objective placement<br/>cost · carbon · latency ·<br/>sovereignty"]
P --> G["Generation<br/>Terraform · Argo · CI"]
G --> R[["Multi-provider<br/>deployment"]]
R -. "drift, prices, outages" .-> A["MAPE-K adaptation<br/>contracts · re-placement"]
A -.-> P
Detailed scientific and software architecture, with diagrams of the feature model, the process line, the placement problem, the adaptation loop and the coordination contracts: docs/ARCHITECTURE.md.
Planned work for the second half of the project (M25 to M48), by axis and priority: ROADMAP.md.
pip install -e ".[dev,experiments]"
amlops knowledge # what the knowledge base contains
amlops validate examples/churn_novice.amlops.yaml # best-practice findings
amlops place examples/churn_novice.amlops.yaml --pareto
amlops generate examples/churn_novice.amlops.yaml -o out/churn
amlops simulate examples/predictive_maintenance_expert.amlops.yaml
amlops analyse # #configurations, core/dead/false-optional features
amlops export-uvl -o model.uvl # feature model in UVL (FeatureIDE, flamapy)
amlops schema -o amlops.schema.json # JSON Schema of the DSL (editor completion)
pytest -q # 70 tests
python experiments/run_all.py # regenerates all paper numbers (E1 to E7)
python experiments/sovereignty.py # E8: price of sovereignty on public data (data/public-2026-09-29)
python experiments/case_study.py # E9: industrial case, anonymised aggregates (data/case-edu-llm-2026)A non-expert model is five lines:
amlops: "0.1"
pipeline: churn-prediction
profile: tabular-classification-continuous
data: {source: "s3://datalake/crm/churn.parquet", volume_gb: 40}
objectives: {cost: 0.6, carbon: 0.4}Experts refine features, steps, triggers, placement and parameters (see
examples/predictive_maintenance_expert.amlops.yaml and docs/DSL.md).
| Project requirement | Where |
|---|---|
| Feature models / product lines to capture commonalities | knowledge/mlops_feature_model.yaml, variability/ |
| Documented best practices guiding design | knowledge/best_practices.yaml, variability/advisor.py |
| (i) generic, extensible DSL | step-kind & generator registries, amlops.step_kinds entry points |
| (ii) abstract, ready-to-use by non-experts | profiles, examples/churn_novice.amlops.yaml |
| (iii) parameterisation by experts | features, steps, overrides |
| (iv) pivotal model → PaC/IaC | dsl/derivation.py, generators/ (+ trace.json) |
| Continuous training on drift | adaptation/drift.py, contracts, Argo sensors |
| Dynamic multi-provider (re)deployment, cost / footprint | placement/optimizer.py, placement/dynamic.py |
src/amlops/
knowledge/ feature model, rules, profiles, 38 SkeltyMLOps activities, provider catalogue
variability/ feature-model semantics, completion, advisor, exact #SAT analysis, UVL export
dsl/ metamodel, YAML parser, registry, derivation (process-line base model)
placement/ exact multi-objective placement, Pareto front, dynamic policies simulator
generators/ Terraform, Kubernetes/Argo, GitHub Actions, traceability
adaptation/ PSI/KS drift detection, executable coordination contracts, MAPE-K loop
examples/ four illustrative case studies (+ generated output of one)
experiments/ run_all.py (E1 to E7), export_results_tex.py and results/*.json
schema/ JSON Schema of the DSL and UVL export of the feature model
paper/ LaTeX sources (main.tex, EN; main_fr.tex, FR); generated/; figures/
scripts/ git hooks (run `sh scripts/install-hooks.sh` after cloning)
Apache-2.0 © 2026 GetCaaS SARL. See CITATION.cff. The accompanying paper
(paper/main.tex) is a draft for consortium review: do not circulate before the
publication review foreseen by the consortium agreement.
This work is supported by the French National Research Agency (ANR) under grant ANR-24-IAS2-0004 (AdaptiveMLOps).