Supported

Living knowledge views

Refresh decisions

Source change	Dependency relation	View action
One evidence revision changes	Summary reads that revision	Mark summary stale
Entity label changes	Index stores entity text	Rebuild affected index rows
Unrelated document changes	No dependency path	Reuse view unchanged
Parser version changes	All regions derived by parser	Rebuild matching parser outputs

How it works

A MaterializedView names a transform version and source pattern. Each ViewMaterialization records its input revisions and output artifact ID. plan_view_refresh compares those records with changed revisions through the shared dependency planner. It returns refresh tasks and reused artifact IDs. Execution and storage remain caller-owned.

Plan an incremental refresh with recorded dependencies

from mari_components.platform import MaterializedView, plan_view_refresh

view = MaterializedView(
    view_id="project-summary",
    transform="summarize-community@2",
    source_pattern="project:mari/**",
)

plan = plan_view_refresh(
    view=view,
    materializations=prior_builds,
    changed_revisions={"github/readme": "git:91df"},
)

for task in plan.tasks:
    rebuild_view(task.artifact_id, reason=task.reason)  # application callback

Dependencies drive recomputation. Generated summaries remain evidence-bound artifacts. Their provenance marks them as derived values.

Refresh contract

The view adapter accepts revision deltas for existing materializations. It detects changed inputs, matching source additions, and transform-version changes. Include configuration changes in the transform identifier. Partition legacy string source IDs by tenant and space before calling this adapter.

Use plan_dependency_updates directly for source deletion, explicit scoped identities, recipe configuration, and multi-stage graphs. Supply the complete current source snapshot and derivations. Completed output receipts allow unchanged output fingerprints to stop downstream work. New view creation and removal of retired projections are application operations.

Measures

Measure	Meaning
Rebuild precision	Rebuilt artifacts with a dependency on the change
Rebuild recall	Changed derivatives that were correctly invalidated
Reuse ratio	Safe materializations reused across the refresh
Equivalence	Incremental output equals a clean full rebuild

Papers and implementations

Differential DataflowDBSPSelf-adjusting computationSemiont projections

Mari implements batch dependency comparison and rebuild planning. Streaming execution and database integration remain application adapters.