Add collective note to read_parquet docstring

Author: wence-

cpp/include/rapidsmpf/streaming/cudf/parquet.hpp

@@ -20,12 +20,18 @@ namespace rapidsmpf::streaming::node {
 /**
  * @brief Asynchronously read parquet files into an output channel.
  *
+ * @note This is a collective operation, all ranks named by the execution context's
+ * communicator will participate. All ranks must specify the same set of options.
+ * Behaviour is undefined if a `read_parquet` node appears only on a subset of the ranks
+ * named by the communicator, or the options differ between ranks.
+ *
  * @param ctx The execution context to use.
  * @param ch_out Channel to which `TableChunk`s are sent.
  * @param max_tickets Maximum number of tickets to throttle production of chunks. Up to
  * this many tasks can start producing data simultaneously.
  * @param options Template reader options. The files within will be picked apart and used
- * to reconstruct new options for each read chunk.
+ * to reconstruct new options for each read chunk. The options should therefore specify
+ * the read options "as-if" one were reading the whole input in one go.
  * @param num_rows_per_chunk Target (maximum) number of rows any sent `TableChunk` should
  * have.
  *

python/rapidsmpf/rapidsmpf/streaming/cudf/parquet.pyx

@@ -44,6 +44,11 @@ def read_parquet(
         Maximum number of tasks that may be suspended having read a chunk.
     options
         Reader options
+
+    Notes
+    -----
+    This is a collective operation, all ranks participating via the
+    execution context's communicator must call it with the same options.
     """
     cdef cpp_Node _ret
     with nogil: