Supply stream and mr to column synthesizers - #23209

cpp/src/io/parquet/reader_impl_preprocess.cu (1)

1226-1240: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Stream mismatch: host_row_offsets still allocated on _stream instead of the new stream parameter.

synthesize_source_index_column now takes an explicit stream/mr, and every other allocation/op in this block correctly uses stream (lines 1235-1239), but host_row_offsets on line 1230 is still allocated against the member _stream. The subsequent async H2D copy (make_device_uvector_async(host_row_offsets, stream, ...)) and label_segments run on stream, while the pinned staging buffer's lifetime/deallocation is tracked against _stream. If a future caller passes a stream different from _stream, the pinned buffer could be reclaimed/reused once _stream completes, before the copy on stream finishes — a stream-ordering race on the staging buffer. Today it happens to be safe only because the sole caller (finalize_output) passes _stream for both parameters, but that coincidence defeats the purpose of this refactor.

🐛 Proposed fix

     auto host_row_offsets =
-      cudf::detail::make_empty_pinned_vector<cudf::size_type>(num_sources + 1, _stream);
+      cudf::detail::make_empty_pinned_vector<cudf::size_type>(num_sources + 1, stream);

As per coding guidelines, "Propagate stream and memory-resource parameters through internal APIs consistently" and "Ensure async memcpy staging buffers outlive the copy they are staging."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cpp/src/io/parquet/reader_impl_preprocess.cu` around lines 1226 - 1240, In
synthesize_source_index_column, the pinned staging buffer allocation for
host_row_offsets is still tied to _stream while the rest of the work uses the
explicit stream parameter. Update that allocation to use the same stream passed
into the function so the pinned buffer lifetime and async H2D copy in
make_device_uvector_async remain ordered on the correct stream. Keep the
existing label_segments and synchronization flow, but make the stream usage
consistent throughout this block.

Source: Coding guidelines