biodata-cache

biodata-cache is a set of one-line functions that handle the entire process of caching and retrieving data (and metadata) from AIND data assets.

In the background, the cache repackages data/metadata into dataframes and stores them on S3 in versioned folders (data-asset-cache/bdc-v{version}/), or in memory for testing. Each release writes to its own versioned folder, so older versions of the website remain accessible while new versions are deployed. A top-level data-asset-cache/cache_versions.json index lists all available version folders.

Important: this package is not at 1.0. It is changing fast and breaking changes are still occurring, although rarely. To reduce the chance of impact on your code the cache tables are versioned. This does mean that if you want the latest version of the tables you need to keep biodata-cache up-to-date, but it also means your code won't immediately break when I change the way the tables work.

Installation

Note that you must set the backend to S3 or biodata-cache will automatically re-cache the tables locally in memory. This can take a LONG time.

pip install biodata-cache
export BIODATA_CACHE_BACKEND='S3'

Usage

Set backend

export BIODATA_CACHE_BACKEND='S3'

Options are 'S3', 'MEMORY'.

Fetch data

from biodata_cache import unique_project_names

project_names = unique_project_names()

Cache tables

Use get_cache_registry() to see all available cache tables and their metadata (descriptions, S3 paths, columns, etc.) for the installed version:

from biodata_cache import get_cache_registry

registry = get_cache_registry()

Use get_cache_versions() to list all available version folders across all deployed releases:

from biodata_cache import get_cache_versions

versions = get_cache_versions()

Each version stores one registry fragment per table at s3://allen-data-views/data-asset-cache/bdc-v{version}/cache_registry/<table>.json. get_cache_registry() merges these fragments and still reads older monolithic registries. The top-level index s3://allen-data-views/data-asset-cache/cache_versions.json lists all available version folders.

Hive-partitioned tables use key=value directory segments, enabling DuckDB queries like:

import duckdb

duckdb.query(
    """
    SELECT * FROM read_parquet(
        's3://allen-data-views/data-asset-cache/bdc-v0.41/qc/subject_id=123/data.pqt',
        hive_partitioning=true,
        union_by_name=true
    )
"""
)

The raw_to_derived function is not a table stored in S3, instead it is used by passing an asset_name (or list of asset names) and a modality. The function returns the latest derived asset matching the requested pattern.

Cells across every asset

The cell tables use cell_key to join identity, measurements, and transcriptomic data:

Table	Rows	Partition
cell_index	One row per cell	None
cell_properties	Measurements for each cell	asset_name
cell_genes	Genotyping results	subject_id

The cell-by-everything sync job runs after ecephys_units, pophys, and visual_learning. It also reads Visual Coding Neuropixels NWB-Zarr data directly so it does not depend on manual SWDB tables.

Custom cache table

The custom function allows you to store and retrieve your own user-defined DataFrames in the cache by name. This requires write authentication to the active backend.

from biodata_cache import custom
import pandas as pd

df = pd.DataFrame({"col": [1, 2, 3]})
custom("my_data", df)

retrieved_df = custom("my_data")

Update all cache tables

The cache is rebuilt on Code Ocean by a set of per-table sync jobs wired into a Nextflow pipeline (asset_basics first, then the rest in parallel, followed by cell-by-everything). The BIODATA_CACHE_SYNC_JOB environment variable selects the job. The job writes its own registry fragment. See PIPELINE.md for the job list, pipeline layout, and version procedure.

To run a single job (as a capsule does):

from biodata_cache.sync import run_sync_job

run_sync_job()  # reads BIODATA_CACHE_SYNC_JOB, or pass e.g. run_sync_job("qc")

From the repository, use python scripts/run_sync.py <job>.

To rebuild everything in one local process (not used by the pipeline):

from biodata_cache.sync import update_all_tables

update_all_tables()