update docstring

python/pyarrow/_dataset_parquet.pyx

@@ -669,7 +669,10 @@ cdef class ParquetFragmentScanOptions(FragmentScanOptions):
     pre_buffer : bool, default True
         If enabled, pre-buffer the raw Parquet data instead of issuing one
         read per column chunk. This can improve performance on high-latency
-        filesystems.
+        filesystems (e.g. S3, GCS) by coalesing and issuing file reads in
+        parallel using a background I/O thread pool.
+        Set to False if you want to prioritize minimal memory usage
+        over maximum speed.
     thrift_string_size_limit : int, default None
         If not None, override the maximum total string size allocated
         when decoding Thrift structures. The default limit should be

python/pyarrow/parquet/core.py

@@ -1742,11 +1742,12 @@ class ParquetDataset:
     different partitioning schemes, etc.
 pre_buffer : bool, default True
     Coalesce and issue file reads in parallel to improve performance on
-    high-latency filesystems (e.g. S3). If True, Arrow will use a
+    high-latency filesystems (e.g. S3, GCS). If True, Arrow will use a
     background I/O thread pool. This option is only supported for
     use_legacy_dataset=False. If using a filesystem layer that itself
     performs readahead (e.g. fsspec's S3FS), disable readahead for best
-    results.
+    results. Set to False if you want to prioritize minimal memory usage
+    over maximum speed.
 coerce_int96_timestamp_unit : str, default None
     Cast timestamps that are stored in INT96 format to a particular resolution
     (e.g. 'ms'). Setting to None is equivalent to 'ns' and therefore INT96