diff --git a/toshi_hazard_store/model/revision_4/pyarrow_aggr_dataset.py b/toshi_hazard_store/model/revision_4/pyarrow_aggr_dataset.py
new file mode 100644
index 0000000..11316b3
--- /dev/null
+++ b/toshi_hazard_store/model/revision_4/pyarrow_aggr_dataset.py
@@ -0,0 +1,84 @@
+"""pyarrow helper function"""
+
+import csv
+import logging
+import pathlib
+import uuid
+from functools import partial
+from typing import TYPE_CHECKING, Iterable, Union
+
+import pandas as pd
+import pyarrow as pa
+import pyarrow.dataset
+import pyarrow.dataset as ds
+
+log = logging.getLogger(__name__)
+
+if TYPE_CHECKING:
+    from .hazard_aggregation import HazardAggregation
+    from .hazard_realization_curve import HazardRealizationCurve
+
+
+def write_metadata(output_folder: pathlib.Path, visited_file: pyarrow.dataset.WrittenFile) -> None:
+    meta = [
+        pathlib.Path(visited_file.path).relative_to(output_folder),
+        visited_file.size,
+    ]
+    header_row = ["path", "size"]
+
+    # NB metadata property does not exist for arrow format
+    if visited_file.metadata:
+        meta += [
+            visited_file.metadata.format_version,
+            visited_file.metadata.num_columns,
+            visited_file.metadata.num_row_groups,
+            visited_file.metadata.num_rows,
+        ]
+        header_row += ["format_version", "num_columns", "num_row_groups", "num_rows"]
+
+    meta_path = pathlib.Path(visited_file.path).parent / "_metadata.csv"  # note prefix, otherwise parquet read fails
+    write_header = False
+    if not meta_path.exists():
+        write_header = True
+    with open(meta_path, 'a') as outfile:
+        writer = csv.writer(outfile)
+        if write_header:
+            writer.writerow(header_row)
+        writer.writerow(meta)
+    log.debug(f"saved metadata to {meta_path}")
+
+
+def append_models_to_dataset(
+    models: Iterable[Union['HazardRealizationCurve', 'HazardAggregation']],
+    output_folder: pathlib.Path,
+    dataset_format: str = 'parquet',
+) -> int:
+    """
+    append realisation models to dataset using the pyarrow library
+
+    TODO: option to BAIL if realisation exists, assume this is a duplicated operation
+    TODO: schema checks
+    """
+
+    def groomed_models(models):
+        for model in models:
+            yield model.as_pandas_model()
+
+    df = pd.DataFrame(groomed_models(models))
+
+    table = pa.Table.from_pandas(df)
+
+    write_metadata_fn = partial(write_metadata, output_folder)
+
+    ds.write_dataset(
+        table,
+        base_dir=str(output_folder),
+        basename_template="%s-part-{i}.%s" % (uuid.uuid4(), dataset_format),
+        partitioning=['nloc_0'],
+        partitioning_flavor="hive",
+        existing_data_behavior="overwrite_or_ignore",
+        format=dataset_format,
+        file_visitor=write_metadata_fn,
+    )
+
+    return df.shape[0]