Zenml Repo Workflows

ZenML Repo Workflows

Use this skill when work in the ZenML repository needs more detail than the always-loaded AGENTS.md files provide. The root guide keeps safety rules, universal conventions, and the most common commands in memory. This skill keeps the longer recipes, examples, and subsystem checklists.

Moved Content Index

Former root headings covered here:

• Code Style & Quality Standards
• Commenting policy
• Formatting and Linting
• Python Standards
• Util Function Placement
• Private Methods and API Stability
• FastAPI Agent Profile
• FastAPI Project Structure
• Prefer typing over dynamic attribute checks
• Error Handling & Validation
• Testing Requirements
• Dependencies & Runtime Constraints
• Development Workflow
• Prerequisites
• Documentation Access via MCP
• Environment Variables
• Branch Management
• Making Changes
• Security Guidelines
• Database and Migration Guidelines
• Migration Testing Workflow
• Commit Message Guidelines
• When Implementing Features
• When Fixing Bugs
• Pull Request Guidelines
• Continuous Integration
• Core Concepts
• Important Terminology
• Pipeline Architecture
• Key Abstractions
• Cross-Cutting Architecture Areas
• Common Tasks
• Adding New Integrations
• Modifying Core Functionality
• Expert Tips
• Summary Checklist for PR Reviewers
• Documentation Guidelines
• Structure
• Format and Style
• Content Standards

Development Workflow

Setup

• Install ZenML in development mode with dev dependencies.
• ZenML recommends uv for Python package installation because it resolves dependencies more quickly and reliably than plain pip.
• Useful development environment variables:
  • ZENML_LOGGING_VERBOSITY=DEBUG
  • MLSTACKS_ANALYTICS_OPT_OUT=true
  • AUTO_OPEN_DASHBOARD=false
  • ZENML_ENABLE_RICH_TRACEBACK=false
  • TOKENIZERS_PARALLELISM=false
• Always set the following environment variables when developing:
  • ZENML_ANALYTICS_OPT_IN=false: Disables analytics during development
  • ZENML_DEBUG=true: Uses the development ZenML analytics server to avoid sending analytics to the official ZenML analytics server (IMPORTANT!). This must be set even if ZENML_ANALYTICS_OPT_IN=true because in a client-server setup, the server controls the client-side analytics opt-in status.

Formatting, Linting, and Tests

• Format before committing: bash scripts/format.sh.
• Check quality: bash scripts/lint.sh.
• The lint script runs Ruff, pydoclint on src/zenml tests/harness, yamlfix, zizmor, unused import/variable checks, Ruff formatting checks, and mypy.
• Full mypy is slow on the whole codebase. For focused work, run mypy on the changed file or package.
• Run targeted pytest files or tests for the component you changed. Avoid the full local test suite unless there is a specific reason.
• Some coverage workflows use bash scripts/test-coverage-xml.sh, but that does not run every test.

Branches and PRs

• develop is the primary working branch. PRs should target develop.
• main is only updated during the release process.
• PR titles should be human-readable and concise. Avoid prefixes like feat:.
• PR descriptions should explain what changed, why it was needed, important implementation decisions, and anything reviewers should inspect carefully.
• Every PR needs exactly one release-notes label:
  • release-notes for user-facing features, API changes, important fixes, or changes users should know about.
  • no-release-notes for internal work, CI fixes, refactors, minor docs-only changes, or maintenance work.

Continuous Integration

ZenML uses a two-tier CI setup:

• Fast CI runs on all PRs and covers basic tests, linting, and type checking.
• Full CI includes integration tests and broader coverage.
• The run-slow-ci label triggers full CI. Maintainers add it when required.
• If changes touch integrations or core behavior, mention in the PR that full CI should run.
• For GitHub Actions debugging, check whether logs have expired before deep investigation. If logs are expired, inspect current workflow files and recent runs instead.

Code Style Details

Comments

Use comments to explain intent, trade-offs, invariants, and edge cases. Prefer clear names and small functions over comments that restate the code.

Good comments answer questions such as:

• Why is this constraint here?
• What edge case is being protected?
• Which external API behavior forced this shape?
• What invariant must future edits preserve?

Avoid:

• Change-tracking comments such as "updated from previous version".
• Comments that repeat the line below them.
• Multi-line banner comments grouping classes or functions.

Python

• Use Python 3.10+ compatible code.
• Follow Google Python style for docstrings. Include Args, Returns, Yields, and Raises sections whenever the function contract requires them; do not use a summary-only docstring to omit applicable sections.
• Type hint function parameters and return values.
• Use descriptive names.
• Keep functions manageable. A 50-line target is useful, though not absolute.
• Prefer functional style where it fits the existing code.

Helper Placement

Put a helper on a class when it only makes sense in that class context or when subclasses call it frequently. Use a utility module for generic behavior shared across unrelated modules.

Example: BaseOrchestrator.requires_resources_in_orchestration_environment could be a global helper, but it lives on BaseOrchestrator because orchestrator subclasses call it often and the behavior is part of orchestrator execution decisions.

Useful utility locations:

• src/zenml/utils/
• src/zenml/orchestrators/utils.py
• src/zenml/orchestrators/step_run_utils.py
• src/zenml/orchestrators/publish_utils.py

Private Methods and API Stability

Symbols with a leading underscore are private and should only be called from inside their class or module.

When changing a non-underscore symbol, check whether it is exported from zenml.__init__. Root exports and public methods on those exports are public API and generally require deprecation before breaking changes.

Internal non-underscore code that is not exported at the root can usually be changed without deprecation, but update all internal usages.

Integrations should avoid ZenML private methods because future external integration packages will not be protected by in-repo mypy checks.

Prefer Typing Over Dynamic Attribute Checks

Prefer Protocol, ABCs, Union with isinstance narrowing, or typed adapters over getattr/hasattr for capability checks. Use dynamic attribute checks only when the object is truly untyped, and isolate that behavior in a small typed helper.

FastAPI and Server Work

ZenML OSS FastAPI work expects strong FastAPI, SQLModel, SQLAlchemy 2.0, and Pydantic v2 judgment.

Preferred patterns:

• Extend existing service or repository classes when behavior belongs there.
• Keep shared state in FastAPI dependency injection or the application factory.
• Use synchronous def route handlers in OSS Codex contributions.
• Use descriptive lower_snake_case paths and named exports.
• Co-locate Pydantic request/response models with consuming routes.
• Keep reusable behavior in service or repository modules accessed through dependency injection.
• Start routes, dependencies, and services with guard clauses.
• Avoid redundant else blocks after returns.
• Raise HTTPException with precise status codes for expected failures.
• Let middleware translate unexpected failures into consistent JSON responses with logging and metrics.
• Define inputs and outputs with Pydantic models so validation and docs stay in sync.

Import rule: code outside src/zenml/zen_server/ must not import from zen_server/. Client-side code should use Client and shared models from src/zenml/models/.

Database and Migrations

ZenML uses SQLModel and SQLAlchemy. Avoid raw SQL unless it is genuinely needed.

Database schema changes require Alembic migrations.

Migration rules:

• Create descriptive migrations, e.g. alembic revision -m "Add X to Y table".
• Test upgrade paths with alembic upgrade head.
• Downgrade testing is optional because ZenML generally does not support downgrades.
• Never modify existing migrations that are already on main or develop.
• Consider rolling-deployment compatibility.
• Include schema changes and data migrations when both are needed.
• Run scripts/check-alembic-branches.sh to verify migration consistency.

Migration testing workflow:

1. Check out develop or the relevant old release.
2. Populate the database from that version.
3. Switch to the feature branch.
4. Run alembic upgrade head.

Useful MySQL query for foreign key dependency tracing:

SELECT
  kcu.CONSTRAINT_NAME AS fk_name,
  kcu.TABLE_SCHEMA AS referencing_schema,
  kcu.TABLE_NAME AS referencing_table,
  kcu.COLUMN_NAME AS referencing_column,
  kcu.REFERENCED_TABLE_SCHEMA AS referenced_schema,
  kcu.REFERENCED_TABLE_NAME AS referenced_table,
  kcu.REFERENCED_COLUMN_NAME AS referenced_column,
  rc.UPDATE_RULE,
  rc.DELETE_RULE
FROM INFORMATION_SCHEMA.KEY_COLUMN_USAGE kcu
JOIN INFORMATION_SCHEMA.REFERENTIAL_CONSTRAINTS rc
  ON rc.CONSTRAINT_SCHEMA = kcu.CONSTRAINT_SCHEMA
 AND rc.CONSTRAINT_NAME  = kcu.CONSTRAINT_NAME
WHERE kcu.REFERENCED_TABLE_SCHEMA = '<DB_NAME>'
  AND kcu.REFERENCED_TABLE_NAME IN ('table1', 'table2')
ORDER BY kcu.TABLE_NAME, kcu.CONSTRAINT_NAME, kcu.ORDINAL_POSITION;

Use it before adding cascade behavior, dropping columns/tables, debugging FK failures, or auditing existing delete rules.

Documentation Work

Documentation content lives in docs/book/. Generated docs directories such as docs/mkdocs/ and docs/site/ are not edit targets.

GitBook URLs follow the table of contents, not direct file paths. Pro docs under docs/book/getting-started/zenml-pro/ are served under https://docs.zenml.io/pro/....

When adding or removing docs pages, update the relevant toc.md. Assets belong in a .gitbook folder beside the relevant toc.md.

Before committing docs changes with absolute URLs:

1. Check whether the URL exists.
2. Compare with existing working links in the same section.
3. Inspect the relevant toc.md to understand the public URL.

Link checking:

• scripts/check_broken_links.py and scripts/check_relative_links.py check relative markdown links.
• Lychee covers local files, images, HTML links, and external URLs.
• Local full docs check: lychee --no-progress 'docs/book/**/*.md'.
• Offline local links only: lychee --offline --no-progress 'docs/book/**/*.md'.

Domain Models

Models live under src/zenml/models.

Core concepts:

• Requests, Updates, and typed Responses use Body, Metadata, and Resources.
• Filters provide typed operations, scoping, sorting, and pagination.
• Domain models must stay aligned with ORM schemas.
• Canonical Cursor rules live in .cursor/rules/zenml-domain-models.mdc and .cursor/rules/zenml-orm-schema.mdc.

Common base patterns:

• BaseRequest -> {Entity}Request
• BaseUpdate -> {Entity}Update
• BaseResponse[Body, Metadata, Resources]
• BaseFilter, UserScopedFilter, ProjectScopedFilter, TaggableFilter

Adding a new entity usually means implementing Request, ResponseBody, ResponseMetadata, ResponseResources, Response, and Filter classes. Choose the narrowest scope that matches ownership semantics.

Filter Field Coupling

When adding a field to a filter model, update three locations:

1. The filter model field.
2. The corresponding Client list method signature.
3. The filter model instantiation inside that client method.

If the field should not become a CLI option, add it to CLI_EXCLUDE_FIELDS.

Test the new filter through the CLI, for example:

zenml pipeline runs list --new_field=some_value

Relationship-backed filter fields can also require custom ORM join behavior in the store layer.

Model Compatibility

Usually safe:

• Adding optional properties.

Risky or breaking:

• Deleting properties.
• Renaming properties.
• Making required fields optional in a way older code cannot tolerate.
• Changing property types.

Safe evolution pattern: add optional fields, deprecate before removal, use defaults when adding required behavior, and consider versioned responses for major changes.

CLI Work

Important files:

• src/zenml/cli/utils.py for list_options.
• src/zenml/cli/pipeline.py for pipeline, run, legacy schedule, and replay/resume commands.
• src/zenml/cli/trigger.py for native trigger commands.
• src/zenml/cli/resource_pool.py and resource_request.py for resource pool and queue inspection commands.
• src/zenml/cli/stack.py for stack management.
• src/zenml/cli/base.py for core CLI setup and common decorators.

List commands use @list_options(FilterModel) to generate options from filter model fields, then pass them to explicit client method parameters. If a field is missing from the client method, the CLI can raise:

TypeError: list_pipeline_runs() got an unexpected keyword argument 'new_field'

Scheduling command families:

• zenml pipeline schedule ... for legacy schedule records.
• zenml trigger schedule ... for native schedule triggers.
• zenml trigger platform-event ... for platform event triggers.

Resource pools use both management commands and request/queue inspection commands. Check ResourcePool*, ResourcePoolSubjectPolicy*, and ResourceRequest* models when editing this area.

Import rules:

• Do not import integration libraries at module level in CLI files.
• Do not import from zen_server/ in CLI code.
• Do not import SQL schemas directly from zen_stores/; use Client.

Integrations

The most important integration rule: flavor files must not import integration libraries at module level.

Flavor files live at src/zenml/integrations/*/flavors/*.py. They are imported for component discovery, so optional third-party libraries must only be imported inside methods or under TYPE_CHECKING.

Correct pattern:

from typing import TYPE_CHECKING, Type

from zenml.orchestrators import BaseOrchestratorConfig
from zenml.orchestrators.base_orchestrator import BaseOrchestratorFlavor

if TYPE_CHECKING:
    from zenml.integrations.aws.orchestrators import SagemakerOrchestrator


class SagemakerOrchestratorFlavor(BaseOrchestratorFlavor):
    @property
    def implementation_class(self) -> Type["SagemakerOrchestrator"]:
        from zenml.integrations.aws.orchestrators import SagemakerOrchestrator

        return SagemakerOrchestrator

Integration package shape:

src/zenml/integrations/<name>/
├── __init__.py
├── flavors/
├── orchestrators/
├── step_operators/
├── materializers/
└── ...

Step operators should implement the async-first lifecycle:

• submit(...)
• get_status(...)
• wait(...)
• cancel(...)

StepLauncher calls submit() plus wait() by default and falls back to legacy launch() only for backwards compatibility. Store backend job IDs in run metadata immediately after submission.

Dependency updates:

• Prefer inclusive lower bounds and exclusive upper bounds.
• Check Python version compatibility.
• Test locally with realistic integration scenarios.
• Treat dropping support for an old version as breaking.
• Mention breaking dependency changes in release notes.

Orchestrators

Key files:

• base_orchestrator.py
• containerized_orchestrator.py
• utils.py
• step_launcher.py
• step_runner.py
• cache_utils.py, input_utils.py, publish_utils.py

Main submission methods:

• submit_pipeline(...) for static pipelines.
• submit_dynamic_pipeline(...) for dynamic pipelines.
• Isolated-step APIs for dynamic pipeline support:
  • submit_isolated_step(...)
  • get_isolated_step_status(...)
  • wait_for_isolated_step(...)
  • stop_isolated_step(...)

BaseOrchestrator.run(...) already prunes steps skipped by replay or client-side caching before submission. Integration orchestrators should not re-implement that pruning.

get_orchestrator_run_id

This method must return an ID that is unique per backend run and stable for all steps in the same ZenML pipeline run.

Static pipelines often start directly with the first step. The first step uses the orchestrator run ID to create the ZenML run, and downstream steps use the same ID to find it.

Dynamic pipelines have an initial orchestration container. Use an ID that is stable for retries of that orchestration environment.

Kubernetes is special because it has an orchestration container even for static pipelines. Prefer the configured Kubernetes run ID for static pipelines and the parent Kubernetes job name for dynamic pipelines, falling back only when the job lookup fails.

Implementation checklist:

• Inherit from ContainerizedOrchestrator if steps run in containers.
• Implement get_orchestrator_run_id() correctly.
• Implement submit_pipeline() for static pipelines.
• Implement submit_dynamic_pipeline() when supporting dynamic pipelines.
• Add isolated-step methods when needed.
• Handle scheduling hooks when the backend supports scheduling.
• Respect CPU, memory, GPU, and other resource settings.
• Return SubmissionResult with wait_for_completion for synchronous execution.
• Use self.get_image(deployment, step_name).
• Use orchestrator_utils.get_step_entrypoint_command(...).

ORM Schemas and Zen Stores

ORM schemas live under src/zenml/zen_stores/schemas.

Rules:

• Use SQLModel declarative table classes.
• Prefer BaseSchema or NamedSchema.
• Keep fields aligned with domain models.
• Schema changes usually require migrations.
• Export new schemas from src/zenml/zen_stores/schemas/__init__.py.
• General string column limit is about 250 characters because of MySQL.

Foreign keys:

• Use schema_utils.build_foreign_key_field.
• Define relationships on both sides with matching back_populates.
• For many-to-many relationships, use a link model with composite primary keys.

Conversions:

• Provide to_model.
• Provide from_request or from_model where appropriate.
• Provide update or update_from_model.
• Include related resources only when requested by flags.
• Refresh updated on mutations.

Eager-loading rules:

• Collections use selectinload; joinedload on collections multiplies rows.
• Single-valued relationships use joinedload for get paths when the related row usually exists.
• Keep selectinload for nullable FKs that are usually empty.

Import rule: code outside zen_stores/ should not import SQL-related code directly from this directory. Use Client, or client.zen_store when lower level access is truly needed.

Cross-Area Change Checklist

Some features span many layers. When touching these, trace the whole path:

• Dynamic pipelines and nested child pipelines:
  • src/zenml/execution/pipeline/dynamic/
  • src/zenml/pipelines/dynamic/
  • orchestrators and step_launcher.py
  • pipeline run models, schemas, migrations, tests, and docs
• Triggers:
  • src/zenml/triggers/
  • trigger models, schemas, CLI, client, server endpoints, docs
• Resource pools and requests:
  • src/zenml/models/v2/core/resource_pool*.py
  • resource_request.py
  • src/zenml/zen_stores/resource_pools/
  • CLI, store behavior, Pro/backend scheduling behavior
• Container engines:
  • use src/zenml/container_engines/ instead of direct Docker-specific calls

Review Checklist

• Integration PRs: no library imports in flavor files.
• Orchestrator PRs: get_orchestrator_run_id is unique per run and stable for all steps in that run.
• Filter model changes: matching client method signature and body are updated.
• Private method changes: all internal usages are updated.
• Import checks: no zen_server imports outside zen_server.
• Import checks: no direct SQL imports outside zen_stores.
• Model changes: adding properties is usually OK; deleting, renaming, or incompatible type changes are risky.
• Dependency bumps: dropping old version support is breaking.
• Scheduling changes: update both legacy schedule and trigger stacks where relevant.
• Step operator changes: check BaseStepOperator, StepLauncher, and at least one concrete integration.