OasisDataManager

A Python library providing unified data access patterns across different storage backends and DataFrame engines, as part of the OasisLMF catastrophe modelling platform.

It abstracts:

• File I/O — local filesystem, AWS S3, Azure Blob Storage
• DataFrame reading — pandas, Dask, PyArrow
• Data pipelines — fetch → filter → SQL → transform, composable via a fluent API

---

Installation

# Core install (pandas + local storage only)
pip install oasis-data-manager

# Development install
pip install -e .
pip install -r requirements.txt

# With optional cloud and distributed features (Dask, S3, Azure, PyArrow)
pip install -e ".[extra]"

---

Quick start

from oasis_data_manager.df_reader.backends.pandas import OasisPandasReader
from oasis_data_manager.filestore.backends.local import LocalStorage

storage = LocalStorage("/data")

# Read a CSV and get a pandas DataFrame
df = OasisPandasReader("accounts.csv", storage).as_pandas()

# Chain filters
df = (
    OasisPandasReader("accounts.csv", storage)
    .filter([lambda x: x[x["PortNumber"] == "1"]])
    .as_pandas()
)

---

Storage backends

Three backends are provided. All share the same interface.

Local

from oasis_data_manager.filestore.backends.local import LocalStorage

storage = LocalStorage(root_dir="/data")

AWS S3

from oasis_data_manager.filestore.backends.aws import AwsS3Storage

storage = AwsS3Storage(
    bucket_name="my-bucket",
    access_key="AKIA...",
    secret_key="...",
    root_dir="models/",          # optional sub-path within the bucket
)

Azure Blob Storage

from oasis_data_manager.filestore.backends.azure import AzureABFSStorage

storage = AzureABFSStorage(
    account_name="myaccount",
    account_key="...",
    azure_container="my-container",
    root_dir="models/",          # optional sub-path within the container
)

Configuration dict pattern

Used throughout the OasisLMF platform to configure storage from serialisable dicts:

from oasis_data_manager.filestore.config import get_storage_from_config

config = {
    "storage_class": "AwsS3Storage",
    "options": {
        "bucket_name": "my-bucket",
        "access_key": "AKIA...",
        "secret_key": "...",
    }
}
storage = get_storage_from_config(config)

Common storage operations

# Open a file (context manager, like built-in open)
with storage.open("path/to/file.csv") as f:
    data = f.read()

# Copy a file to a local temp directory
local_path = storage.get("remote/file.parquet", "/tmp/")

# Upload a file
storage.put("/tmp/output.csv", "remote/output.csv")

# Delete
storage.delete_file("remote/old.csv")

---

DataFrame readers

Reader backends

Class	Engine	Formats	Filter behaviour
OasisPandasReader	pandas	CSV, Parquet	In-memory (post-load)
OasisDaskReader	Dask	CSV, Parquet	In-memory via dask-sql
OasisPyarrowReader	PyArrow	Parquet only	Predicate pushdown (pre-load)

Format-specific subclasses (OasisPandasReaderCSV, OasisDaskReaderParquet, etc.) are available for Pandas and Dask, but only base OasisPyarrowReader is available and uses parquet.

Fluent API

All readers share the same chainable interface. The actual file read is lazy — it happens on the first access to .df or when .as_pandas() is called.

from oasis_data_manager.df_reader.backends.pandas import OasisPandasReader
from oasis_data_manager.df_reader.backends.dask import OasisDaskReader
from oasis_data_manager.filestore.backends.local import LocalStorage

storage = LocalStorage("/data")

# Pandas — CSV
df = OasisPandasReader("losses.csv", storage).as_pandas()

# Pandas — Parquet (detected automatically from extension)
df = OasisPandasReader("losses.parquet", storage).as_pandas()

# Dask
df = OasisDaskReader("losses.csv", storage).as_pandas()

Filtering

Pass a list of callables; each receives the DataFrame and must return a (filtered) DataFrame.

df = (
    OasisPandasReader("locations.csv", storage)
    .filter([
        lambda x: x[x["CountryCode"] == "US"],
        lambda x: x[x["LocNumber"].notna()],
    ])
    .as_pandas()
)

OasisPandasReader and OasisDaskReader apply filters after loading the full file into memory. OasisPyarrowReader accepts a filters kwarg (list of tuples or list of lists) that is pushed down into the Parquet engine before any data is read into memory — use this for large Parquet files where row-group skipping matters.

from oasis_data_manager.df_reader.backends.pyarrow import OasisPyarrowReader

# AND of conditions — list of tuples
df = (
    OasisPyarrowReader("losses.parquet", storage)
    .read(filters=[("CountryCode", "==", "US"), ("TIV", ">=", 1_000_000)])
    .as_pandas()
)

# OR of AND-groups — list of lists
df = (
    OasisPyarrowReader("losses.parquet", storage)
    .read(filters=[[("CountryCode", "==", "US")], [("CountryCode", "==", "GB")]])
    .as_pandas()
)

Supported operators: ==, !=, <, <=, >, >=, in, not in.

SQL (Dask only)

Requires dask-sql. The reserved table name is table.

from oasis_data_manager.df_reader.backends.dask import OasisDaskReader

df = (
    OasisDaskReader("locations.csv", storage)
    .sql("SELECT LocNumber, Latitude, Longitude FROM table WHERE CountryCode = 'US'")
    .as_pandas()
)

Arbitrary queries

.query(fn) passes the raw DataFrame to any callable and returns the result directly (not a reader).

count = OasisPandasReader("losses.csv", storage).query(lambda df: len(df))

Configuration dict pattern

from oasis_data_manager.df_reader.config import get_df_reader

config = {
    "path": "accounts.csv",
    "storage": storage,
    "options": {"dtype": {"LocNumber": str}},
    "engine": "OasisPandasReaderCSV",
}
reader = get_df_reader(config)
df = reader.as_pandas()

---

Complex data pipelines

ComplexData composes storage fetch, SQL filtering, and post-read transformations into a single reusable class.

FileStoreComplexData — files not handled by the df_reader (e.g. HDF5)

from oasis_data_manager.complex import FileStoreComplexData, Adjustment
import h5py
import pandas as pd

class NormaliseAdjustment(Adjustment):
    @classmethod
    def apply(cls, df):
        df["loss"] = df["loss"] / df["loss"].max()
        return df

class EventLossData(FileStoreComplexData):
    filename = "event_losses.hdf5"
    sql = "SELECT * FROM table WHERE event_id > 1000"
    adjustments = [NormaliseAdjustment]

    def to_dataframe(self, result) -> pd.DataFrame:
        f = h5py.File(result)
        return pd.DataFrame({"event_id": list(f["event_id"]), "loss": list(f["loss"])})

# Run the pipeline
df = EventLossData(storage=storage).run().as_pandas()

RestComplexData — HTTP endpoints

from oasis_data_manager.complex import RestComplexData

class ExposureAPI(RestComplexData):
    url = "https://api.example.com/exposures"
    timeout = 30

    def get_headers(self):
        return {"Authorization": "Bearer my-token"}

    def handle_response(self, response):
        return response.json()["data"]

df = ExposureAPI().run().as_pandas()

---

Exceptions

Exception	Description
OasisDataManagerException	Base exception for this library
OasisException	Backward-compatible alias for the above
MissingInputsException	Raised when a required input file is not found

from oasis_data_manager.errors import OasisDataManagerException, MissingInputsException

---

Import paths

# Storage backends
from oasis_data_manager.filestore.backends.local import LocalStorage
from oasis_data_manager.filestore.backends.aws import AwsS3Storage
from oasis_data_manager.filestore.backends.azure import AzureABFSStorage
from oasis_data_manager.filestore.config import get_storage_from_config

# DataFrame readers
from oasis_data_manager.df_reader.backends.pandas import OasisPandasReader, OasisPandasReaderCSV, OasisPandasReaderParquet
from oasis_data_manager.df_reader.backends.dask import OasisDaskReader, OasisDaskReaderCSV, OasisDaskReaderParquet
from oasis_data_manager.df_reader.backends.pyarrow import OasisPyarrowReader
from oasis_data_manager.df_reader.config import get_df_reader

# Exceptions
from oasis_data_manager.errors import OasisDataManagerException, OasisException

Deprecated module paths (filestore/backends/aws_s3.py, filestore/backends/azure_abfs.py) still work but emit a DeprecationWarning.

---

Development

# Install dev dependencies
pip install -e .
pip install -r requirements.txt

# Run tests
pytest

# Skip type checking and import sorting for faster iteration
pytest --no-header -p no:mypy -p no:isort tests/df_reader/

# Cloud integration tests (requires Docker)
docker compose up -d
pytest tests/filestorage/test_aws.py tests/filestorage/test_azure.py
docker compose down

# Linting
flake8 --select F401,F522,F524,F541 --show-source ./
autopep8 --diff --exit-code --recursive --max-line-length 150 --ignore E402 .

# Build
python setup.py sdist && python setup.py bdist_wheel

---

License

Part of the OasisLMF platform. See repository for licence details.