The BMS LakeAPI

A FastAPI plugin that allows you to expose your data lake as an API, allowing several output formats such as Parquet, Csv, Json, Excel, ...

The LakeAPI also includes a minimal security layer for convenience (Basic Auth), but you can also bring your own.

Unlike roapi, we intentionally do not expose most SQL by default, but limit the possible queries with a config. This makes it easy for you to control what happens to your data. If you want the SQL endpoint, you can enable it. And because we built LakeAPI on the shoulders of giants like FastAPI. A lot of things like documentation and authentication are built in and we don't have to reinvent the wheel.

To run the application with the default config, just do it:

app = FastAPI()
bmsdna.lakeapi.init_lakeapi(app)

To adjust the config, you can do like this:

import dataclasses
import bmsdna.lakeapi

def_cfg = bmsdna.lakeapi.get_default_config() # Get default startup config
cfg = dataclasses.replace(def_cfg, enable_sql_endpoint=True, data_path="tests/data") # Use dataclasses.replace to set the properties you want
sti = bmsdna.lakeapi.init_lakeapi(app, cfg, "config_test.yml") # Enable it. The first parameter is the FastAPI instance, the 2nd one is the basic config and the third one the config of the tables

Installation

Pypi Package bmsdna-lakeapi can be installed like any python package : pip install bmsdna-lakeapi

Basic Idea

Based on a YAML configuration and the data source, LakeAPI will automatically generate GET and/or POST endpoints. Calling the endpoint turns the query into an SQL statement that can be executed with the engine of your choice (duckdb or polars). The result is then seralised into the requested format (Json, CSV, Arrow etc).

This makes it super easy to distribute your data lake data to other systems. We use it internally to feed data to MS SQL Server, SQLite, Streamlit and Web Apps. You can host it wherever you want, we use Azure Websites which works fine even for very large data amounts.

OpenAPI

Of course everything works with OpenAPI and FastAPI. This means you can add other FastAPI routes, you can use the /docs and /redoc endpoints.

So everything will be fully documented automatically, which is really cool. 🔥🔥

Engine

DuckDB is the default query engine. Polars is also supported, but lack some features. The query engine can be specified at the route level and at the query level with the hidden parameter $engine="duckdb|polars". If you want polars, add the required extra.

At the moment, DuckDB seems to have an advantage and performs the best. Also features like full text search are only available with DuckDB.

A note on Deltalake Support

Deltalake is supported WITH Column Mapping, thanks to deltalake2db in Polars and DuckDB. Deletion Vectors are not supported yet.

Default Security

By Default, Basic Authentication is enabled. To add a user, simply run add_lakeapi_user YOURUSERNAME --yaml-file config.yml. This will add the user to your config yaml (argon2 encrypted). The generated Password is printed. If you do not want this logic, you can overwrite the username_retriver of the Default Config

Standalone Mode

If you just want to run this thing, you can run it with a webserver:

Uvicorn: uvicorn bmsdna.lakeapi.standalone:app --host 0.0.0.0 --port 8080

Gunicorn: gunicorn bmsdna.lakeapi.standalone:app --workers 4 --worker-class uvicorn.workers.UvicornWorker --bind 0.0.0.0:80

Of course you need to adjust your http options as needed. Also, you need to pip install uvicorn/gunicorn

You can still use environment variables for configuration

Environment Variables

• CONFIG_PATH: The path of the config file, defaults to config.yml. If you want to split the config, you can specify a folder, too
• DATA_PATH: The path of the data files, defaults to data. Paths in config.yml are relative to DATA_PATH
• ENABLE_SQL_ENDPOINT: Set this to 1 to enable the SQL Endpoint

Config File

The application by default relies on a Config file beeing present at the root of your project that's call config.yml.

The config file looks something like this, see also our test yaml:

tables:
  - name: fruits
    tag: test
    version: 1
    api_method:
      - get
      - post
    params:
      - name: cars
        operators:
          - "="
          - in
      - name: fruits
        operators:
          - "="
          - in
    datasource:
      uri: delta/fruits
      file_type: delta

  - name: fruits_partition
    tag: test
    version: 1
    api_method:
      - get
      - post
    params:
      - name: cars
        operators:
          - "="
          - in
      - name: fruits
        operators:
          - "="
          - in
      - name: pk
        combi:
          - fruits
          - cars
      - name: combi
        combi:
          - fruits
          - cars
    datasource:
      uri: delta/fruits_partition
      file_type: delta
      select:
        - name: A
        - name: fruits
        - name: B
        - name: cars

  - name: fake_delta
    tag: test
    version: 1
    allow_get_all_pages: true
    api_method:
      - get
      - post
    params:
      - name: name
        operators:
          - "="
      - name: name1
        operators:
          - "="
    datasource:
      uri: delta/fake
      file_type: delta

  # use duckdb database as file
  - name: fruits_duck
    tag: test
    version: 1
    api_method:
      - get
      - post
    datasource:
      uri: duckdb/fruits.db
      file_type: duckdb
      table_name: fruits
    params:
      - name: fruits
        operators:
          - "="
          - in
      - name: cars
        operators:
          - "="
          - "in"

  - name: fake_delta_partition
    tag: test
    version: 1
    allow_get_all_pages: true
    api_method:
      - get
      - post
    params:
      - name: name
        operators:
          - "="
      - name: name1
        operators:
          - "="
    datasource:
      uri: delta/fake
      file_type: delta

  - name: "*" # We're lazy and want to expose all in that folder. Name MUST be * and nothing else
    tag: startest
    version: 1
    api_method:
      - post
    datasource:
      uri: startest/* # Uri MUST end with /*
      file_type: delta

  - name: fruits # But we want to overwrite this one
    tag: startest
    version: 1
    api_method:
      - get
    datasource:
      uri: startest/fruits/${IN_URI_YOU_CAN_HAVE_ENVIRONMENT_VARIABLES}
      file_type: delta

  - name: mssql_department
    tag: mssql
    api_method: get
    engine: odbc #requires the odbc extra
    datasource:
      uri: DRIVER={ODBC Driver 17 for SQL Server};SERVER=127.0.0.1,1439;ENCRYPT=yes;TrustServerCertificate=Yes;UID=sa;PWD=${MY_SQL_PWD};Database=AdventureWorks
      table_name: "HumanResources.Department"
    params:
      - GroupName

Partitioning for awesome performance

To use partitions, you can either

• Partition by a column that you filter on. Obviously
• partition on a special column called columnname_md5_prefix_2, which means that you're partitioning on the first two characters of your hex-coded md5 hash. of the hexadecimal md5 hash. If you now filter by columnname, this will greatly reduce the number of files you search for. The number of characters used is up to you, we found two to be useful
• partition on a special column called columnname_md5_mod_NRPARTITIONS, where your partition value is str(int(hashlib.md5(COLUMNNAME).hexdigest(), 16) % NRPARTITIONS). This might look a bit complicated, but it's not that hard :) You're just doing a modulo on your md5 hash, which allows you to which allows you to set the exact number of partitions. Filtering is still done correctly on columnname.

Why partition by MD5 hash? Imagine you have a product id where most id's start with a 1 and some newer ones start with a 2. Most of the data will be in the first partition. If you use an MD5 hash, the data will be spread evenly across the partitions.

With this hack you can get sub-second results on very large data. 🚀🚀

You need to use deltalake to use partitions, and you only need str partition columns for now.

Z-ordering can also help a lot :).

Even more features

• Built-in paging, you can use limit/offset to control what you get
• Full-text search using DuckDB's full-text search feature
• jsonify_complex parameter to convert structs/lists to json, the client cannot handle structs/lists
• Metadata endpoints to retrieve data types, string lengths and more
• Easily expose entire folders by using a "*" wildcard in both the name and the datasource.uri config, see example in the config above
• Config in delta table as json, by using the lakeapi.config Table property. Example in the tests. This allows to have the config for the lakeapi along with your data
• Good test coverage

ODBC Support / Sqlite Support

If you want to use ODBC , install the required extra (named odbc).

ODBC Tables cannot be queried through the sql endpoint for now because of Security Implications. We might allow this later if needed via explicit flag.

ODBC is tested against MS SQL Server using ODBC Driver 17 for MS SQL Server

To use Sqlite, you will need to install adbc_driver_sqlite yourself (pip install adbc_driver_sqlite).

Azure Support

Azure Support is tested and supported. S3/GCP could work in theory but is not tested. Supported configs are taken from object_store create and fsspec:

• https://docs.rs/object_store/latest/object_store/azure/enum.AzureConfigKey.html#variants
• https://github.com/fsspec/adlfs#setting-credentials

Supported settings are only those which are supported by BOTH projects. In practice you can currently use:

• account_name
• account_key
• use_emulator (we add that one for fsspec)
• anon (fsspec/adlfs for anonymous). Adlfs default this to true, we default to false. Set to true if you really want to.

If you want to use Azure Default Credential, just specify account_name without any further option.

Further projects

• lakeapi2sql Allows you to read from lake api and write to MS SQL Server

Work in progress

Please note that this is a work in progress, changes will be made and things may break. Especially at this early stage. We recommend only using the export from bmsdna.lakeapi for now and not to use the submodules.

Limitations

• You can currently only serve local files. We might add support for fsspec later on. You can still mount your data if you use Linux, which is what we do in production

Contribution

Feel free to contribute, report bugs or request enhancements.