Exchange checked tables with Python

Use Parquet as the table boundary. Python does not need to import an R runtime; R does not need reticulate. Exchange an immutable, explicitly named delivery rather than having both runtimes replace a live file.

Python writes a candidate using pyarrow:

import pyarrow as pa
import pyarrow.parquet as pq

candidate = pa.table({"policy_id": pa.array([1, 2], type=pa.int32()),
                      "premium": pa.array([100.0, 200.0], type=pa.float64())})
pq.write_table(candidate, "candidate-001.parquet")

R checks and publishes a separate output. This example requires Arrow and local files. A single Parquet target is a checked file delivery, not a managed lake release with registry history.

library(dataraft.core)
library(dataraft.adapters)
contract <- dr_contract("policies",
  columns = c(policy_id = "integer", premium = "numeric"), key = "policy_id")
result <- dr_product("policies") |>
  dr_add_source(dr_source_parquet("candidate-001.parquet", lazy = FALSE)) |>
  dr_add_contract(contract) |>
  dr_set_target(dr_target_parquet("checked-001.parquet")) |>
  dr_run()
stopifnot(result$status == "completed")

Only after a successful result should the consumer be given the output path:

checked = pq.read_table("checked-001.parquet")
assert checked.schema.field("policy_id").type == pa.int32()

Keep integer widths, nulls and timezone-aware timestamps explicit in Arrow schemas. Represent money as integer minor units with a documented currency rather than assuming floating-point equality. Do not exchange RDS or serialized R functions with Python. This boundary exchanges values; it does not reproduce R validation semantics or automatically share approval and release governance. For a governed workflow, export a specifically approved lake release and include its release ID and file checksum in the delivery manifest.