Internal Documentation

This page describes internals that are useful for maintainers. Users should prefer the public API docs unless they are extending the package.

Runtime Model

Every operation call creates an internal node with:

• Parent dependencies
• Operation identity
• Serializable configuration
• User-facing name and metadata
• Cached output after execution

The public Graph object is a wrapper around this lazy node chain.

Execution

Graph.compute() creates an execution runtime and evaluates dependencies recursively. Within one compute call, shared upstream nodes are executed once. The runtime validates domain compatibility before multi-input composition, applies operation functions to pandas frames, then wraps results back into Panel objects.

Dynamic universe masks are reapplied after derived computations so inactive cells cannot leak into later operations.

Hashing And Caching

Panels and operation specs participate in deterministic cache keys. The current cache is in-memory for one execution run. Persistent caches, incremental invalidation, and parallel scheduling are future extensions.

Operation Registry

Built-in operation modules define transformer and composer functions using the same decorators available to users. Generated reference docs are produced from the exported operation catalog:

uv run python scripts/generate_operator_reference.py

Regenerate the reference after adding, removing, or changing operation metadata.

Module Boundaries

• panel owns domain alignment and immutable data containers.
• graph owns lazy graph handles and user-facing compute/output access.
• execution owns dependency evaluation.
• transformer owns unary frame operations.
• composer owns multi-input frame operations.
• registry and _operation own operation metadata and decorators.