ZenML platform feasibility for AMMM3
Assessed: 18 September 2026. Scope: current first-party documentation and package metadata. This note does not report a new installation, pipeline execution or deployment. The separate AMMM3 repository assessment and historical trial evidence must establish the local integration status.
Conclusion
ZenML is a feasible orchestration and artefact-tracking layer for AMMM3 batch workloads. Existing Python code can be wrapped in steps; ZenML does not require a particular statistical modelling framework. This establishes architectural fit, not qualification of AMMM3 sampling, production capacity or causal inference. Official repository
The OSS server stores metadata and exposes the API/dashboard. An orchestrator runs the workload and an artefact store persists its outputs. Deploying the server alone therefore does not supply an AMMM3 execution environment or compute capacity. Architecture, stack components
Integration constraints
| Concern | Verified platform capability | Implication for AMMM3 |
|---|---|---|
| Versions | PyPI returned ZenML 0.96.4, Python >=3.10,<3.15, Pydantic >=2.0,<=2.12.5. | Pin the tested ZenML version. Compare the actual AMMM3 environment with this range; an environment conflict does not establish an intrinsic package incompatibility. |
| Runtime isolation | Container settings support custom images, explicit requirements and per-step environments. The image executing a ZenML step must contain ZenML. | A container by itself does not solve a conflict between ZenML and AMMM3 within the same interpreter. Use compatible resolved dependencies, or invoke AMMM3 through its own interpreter/subprocess with a file-based contract. The subprocess arrangement is an integration design recommendation, not a special ZenML feature. |
| File outputs | Built-in PathMaterializer copies files and archives directories to the artefact store, then reconstructs local paths for consumers. Custom materializers are supported. | Return a persisted AMMM3 output directory initially. No custom model materializer is necessary solely to transport files. Avoid generic cloudpickle as the durable saved-model contract. |
| Long sampling jobs | The Kubernetes orchestrator supports asynchronous execution, CPU/memory settings and configurable deadlines. Its synchronous wait timeout defaults to unlimited. | Long batch fits are compatible with the execution model. Actual wall time, memory, chain/process counts and failure recovery still need workload-specific verification. |
| Caching | Caching is enabled by default; filesystem and external API changes are not detected automatically. Cache dependencies can be configured. | Disable fitting/input-loader caching initially. Before enabling it, include data identity, AMMM3 code/version, complete configuration and random seed in the cache contract. |
| Retries | Step retries support a retry count, delay and backoff. | Treat a retry as another execution attempt. Do not assume interrupted MCMC resumes its sampler state. Use unique attempt directories and publish completed outputs only; verify cancellation and subprocess cleanup. |
Sources: PyPI metadata, upstream dependency contract, containerisation, materializers, Kubernetes orchestrator, caching, retries and execution controls.
Recommended initial boundary
Use one coarse batch step around the existing AMMM3 pipeline, with explicit input configuration and a persisted output bundle. Keep fitting and save/load semantics owned by AMMM3. Split into smaller steps only where a durable intermediate contract already exists. This is an architectural recommendation based on ZenML’s support for existing Python functions and file artefacts, not a verified AMMM3 implementation in this assessment. Python steps, file artefacts
Start with a single worker host and persistent artefact storage. The local Docker orchestrator is an available isolated execution option; moving to Kubernetes is a separate infrastructure choice, not a feasibility prerequisite. For model fitting, use batch execution. ZenML’s HTTP pipeline deployments target immediate request-response workflows and should not be assumed to provide a complete asynchronous fitting service. Local Docker, HTTP deployments
Local AMMM3 evidence
The accompanying repository inspection identified an existing pipeline entry point (run_pipeline, line 231) and model persistence methods (save, line 152; load, line 197). These provide the initial execution and artefact boundaries without changing the statistical library.
The inspected local environment contains AMMM 3.3.2, Pydantic 2.13.4, PyMC 5.28.5 and PyTensor 2.38.3; ZenML is not installed there. Its Pydantic version exceeds ZenML 0.96.4’s upper bound. AMMM3’s declared Pydantic requirement is only >=2.1.0, so this is a conflict with the current resolved environment, not proof that the package requirements cannot coexist. Preserve the existing environment while evaluating the separate worker arrangement.
The 16 September trial record in the separate ammm-platform checkout (.planning/sds/zenml-trial-2026-09-16.md) records an AMMM3 3.3.2 synthetic fit/save/load flow, artefact hashes and retrieval after ZenML 0.96.4 server recreation. Its wrapper used separate interpreters, disabled caching and PathMaterializer. This is prior bounded execution evidence. It was not rerun for this assessment and does not establish present service availability, browser readiness, production capacity or statistical qualification.
Before calling the result production-ready, verify one bounded fit/save/load flow on the intended worker image, retrieval after server/worker restart, failure and cancellation propagation, output integrity, and measured CPU/RAM/disk use. Those checks would establish operational evidence only; model diagnostics and statistical qualification remain separate.
Verification performed
- Read the linked official ZenML documentation and upstream source metadata.
- Queried the live PyPI JSON API with Python’s standard library and extracted the release, Python requirement and Pydantic constraint stated above.
- Did not install dependencies, build images, run sampling, alter a server, or exercise an orchestrator.
- Online documentation and the upstream
mainbranch are mutable. Validate configuration against the pinned release when implementing.