Use ParquetWriter rather than Dataset.writedataset (#216)

The pyarrow write dataset function has shown to consume a lot of memory
while also taking longer to execute. It has features we use around
partitioning and file visitor functions, but those features can be
replicated without too much difficulty given the way we use them.

python_src/src/lamp_py/aws/s3.py

@@ -20,7 +20,7 @@
 import boto3
 import pandas
 import pyarrow
-import pyarrow.dataset as ds
+import pyarrow.compute as pc
 import pyarrow.parquet as pq
 from pyarrow import Table, fs
 from pyarrow.util import guid
@@ -336,7 +336,8 @@ def write_parquet_file(
     @s3_dir - the s3 bucket plus prefix "subdirectory" path where the
         parquet files should be written
     @partition_cols - column names in the table to partition out into the
-        filepath.
+        filepath. NOTE: the assumption is that the values in the partition
+        columns of the incoming table are uniform.
     @visitor_func - if set, this function will be called with a WrittenFile
         instance for each file created during the call. a WrittenFile has
         path and metadata attributes.
@@ -349,26 +350,34 @@ def write_parquet_file(
     )
     process_logger.log_start()
 
-    # generate partitioning for this table write based on what columns
-    # we expect to be able to partition out for this input type
-    partitioning = ds.partitioning(
-        table.select(partition_cols).schema, flavor="hive"
-    )
+    # pull out the partition information into a list of strings.
+    partition_strings = []
+    for col in partition_cols:
+        unique_list = pc.unique(table.column(col)).to_pylist()
+        assert len(unique_list) == 1
+        partition_strings.append(f"{col}={unique_list[0]}")
+
+    table = table.drop(partition_cols)
 
+    # generate an s3 path to write this file to
     if basename_template is None:
         basename_template = guid() + "-{i}.parquet"
-
-    ds.write_dataset(
-        data=table,
-        base_dir=s3_dir,
-        filesystem=fs.S3FileSystem(),
-        format=ds.ParquetFileFormat(),
-        partitioning=partitioning,
-        file_visitor=visitor_func,
-        basename_template=basename_template,
-        existing_data_behavior="overwrite_or_ignore",
+    write_path = os.path.join(
+        s3_dir, *partition_strings, basename_template.format(i=0)
     )
 
+    process_logger.add_metadata(write_path=write_path)
+
+    # write teh parquet file to the partitioned path
+    with pq.ParquetWriter(
+        where=write_path, schema=table.schema, filesystem=fs.S3FileSystem()
+    ) as pq_writer:
+        pq_writer.write(table)
+
+    # call the visitor function if it exists
+    if visitor_func is not None:
+        visitor_func(write_path)
+
     process_logger.log_complete()