Review contracts as files and retain delivery evidence

A reproducible report needs more than a saved total: retain the input release identifiers, the metric definition and the checks made against those inputs. dataraft.metrics::dr_report_release() in dataraft.metrics freezes that evidence for lake-backed reports. The compiler-free example below first establishes a checked, versioned delivery; it does not claim that an RDS delivery is itself a report release.

library(dataraft)
contract <- dr_contract("orders",
columns = c(id = "integer", amount = "numeric"),
key = "id") |>
  dataraft.core::dr_contract_meta(owner = "Reporting",
governance = list(steward = "Data operations", retention = "P7Y",
    tags = c("Finance.Reporting")),
column_metadata = list(amount = list(description = "Order value in EUR")),
producer = "Reporting")
file <- tempfile(fileext = ".json")
dataraft.adapters::dr_contract_odcs(contract, file)
reviewed <- dataraft.adapters::dr_contract_from_odcs(file)
reviewed

This is ODCS 3.2.0, validated against the upstream JSON Schema in CI. Exported metadata includes roles, SLAs and classifications when supplied. Policies are metadata, not automatic enforcement: setting retention does not schedule deletion and identifying a steward does not grant access.

The importer executes flat R-compatible types, required columns, primary keys, nullValues and rowCount library checks with simple numeric comparisons, and DataRaft custom predicates. It rejects unsupported executable constraints, relationships and engines. It does not execute SQL or R strings from contracts. DataRaft predicates allow column names, scalar literals, arithmetic, comparisons, logical operators, is.na() and literal membership via %in% and c(). Other engines must implement that custom rule type before executing it.

root <- tempfile("orders-")
product <- dr_product("orders", data.frame(id = 1:3, amount = c(25, -75, 50))) |>
  dr_add_contract(reviewed) |>
  dr_add_quality(~ amount >= 0, action = "quarantine", dimension = "conformity") |>
  dr_set_target(dataraft.adapters::dr_target_rds(root))
result <- dr_run(product)
dr_collect(result)
dataraft.core::dr_quarantine_rows(result)
dr_quality_report(result)

Quarantine is for native table row predicates. It removes all rejected rows before publication, even if a threshold would otherwise tolerate them. Remaining rows must still pass schema, type, key and required-column checks. Exceptions block execution. Quarantine rows stay local; decide explicitly where sensitive rejected rows may be stored. Model-level or aggregate checks cannot automatically partition member tables.

For warning and blocking rules, threshold = .02 permits up to two percent failures. Existing severity and max_failure remain supported.

before <- dataraft.core::dr_profile_snapshot(data.frame(id = 1:3, amount = c(25, 75, 50)), "input-v1")
after <- dataraft.core::dr_profile_snapshot(data.frame(id = 1:3, amount = c(25, NA, 50)), "input-v2")
dataraft.core::dr_profile_compare(before, after)
profile_path <- tempfile("profiles-")
profile_release <- dr_product("orders.profile", after) |>
  dr_set_target(dataraft.adapters::dr_target_rds(profile_path)) |>
  dr_run()
profile_release$outputs$version

Profiles carry an observation timestamp and explicit input version. Store each profile as a versioned release and compare a chosen baseline, rather than silently replacing it. Missingness changes are deterministic alerts, not a calibrated statistical anomaly detector. Exact profiles can still expose sensitive counts.

recipe <- dataraft.core::dr_recipe() |>
  dataraft.core::dr_step_mutate(net = amount / 1.19) |>
  dataraft.core::dr_step_select(id, net)
dataraft.core::dr_column_lineage(recipe, c("id", "amount"))

The direct-value map handles simple mutation, selection and rename. An opaque function or dynamic selector marks the analysis incomplete. Successful single-input runs publish complete maps through OpenLineage column facets. This does not infer row influence, joins, historical graphs or arbitrary R side effects.

For a file-first project, place project.yaml, an ODCS contract and a CSV source in one directory. dataraft.adapters::dr_project_yaml("project.yaml") |> dr_run() reads this manifest:

id: orders
contract: orders.odcs.json
source: delivery.csv
target: releases/orders

All paths stay inside the project directory. Review the files in Git. Add custom R transformations explicitly in trusted project code when the declarative grammar is insufficient.

Iceberg is an experimental alternative: dataraft.adapters::dr_source_iceberg() reads a metadata location with an optional snapshot ID; dataraft.adapters::dr_target_iceberg() writes to an already attached DuckDB Iceberg REST catalog. Authentication and catalog attachment stay outside definitions. The backend must support transactional candidate validation; no fallback, retention guarantee or branch/merge behavior is implied. See DuckDB's Iceberg catalog documentation.