Introduction

Data storage is of much concern for researchers in every discipline, and a subject of much debate with lots of solutions written.

At its core, a storage format of experiment data has to provide three things:

1. Provide an organizational structure for data, so it can be found again at a later date and be navigated easily by researchers who did not generate the data originally.

2. Permit safe archival of data that will be readable decades in the future and can ideally be processed with many existing tools.

3. Store metadata about the experiment and other descriptive information

In addition to that, a storage layout for raw experiment data also has to fulfill a few more requirements:

4. Support datasets of multiple modalities: Electrophysiology, Tables, Matrices, Video recordings, …

5. Allow massively parallel write and read access on the level of individual datasets

The EDL format is a specification for a storage layout for data from scientific experiments, with primary focus on neuroscientific data. It was originally devised exclusively for the Syntalos DAQ system, which has governed many of its design decisions.

Architecture Overview

Unlike file-based formats for data storage, EDL uses a hierarchy of directories to organize data into logical groups. This permits it to not be limited to one mode of data storage, while also taking full advantage of all performance optimizations done by the operating system at filesystem level when reading or writing data. Additionally, writing to multiple datasets in parallel is possible naturally by means of accessing individual files.

Collections, Groups and Datasets

EDL distinguishes three fundamental organizational units for sorting data: Collections, Groups and Datasets. Each one of these units is represented on the filesystem by a directory containing a manifest.toml file in the TOML markup language which describes the unit type and additional metadata about is contents.

A Collection is the highest organizational unit, similar to a file in single-file based data storage schemes. It defines basic information, like the unique experiment ID and possible authors of the data collection. A collection is never contained in any other EDL unit type.

A Group is a grouping of datasets or other groups. It never contains data on its own, and may contain more groups for further hierarchical nesting, or datasets.

A Dataset contains the actual experiment data in one of the formats permitted in EDL. Dataset directories represent leafs of the directory tree and will never contain more directories. A dataset may have a unit of type group or collection as parent.

Types of organizational units in EDL in an inheritance hierarchy

Metadata

Metadata is stored in the TOML text-based human-readable markup language. Fields that are generic and specified in EDL are stored in manifest.toml files in each directory, while the experimenter can define and add arbitrary custom metadata in a separate attributes.toml file to describe a group, collection or dataset.

Data

Experiment data must be stored in a set of well-defined formats, only one type of data is permitted per dataset. In addition to allowing one primary type of data per dataset, EDL also permits the use of one additional “auxiliary” type of data. This auxiliary data is usually closely tied to the primary data and describes additional information about it, for example mapping of frame numbers to timestamps for videos, or timesync information for data acquired with Syntalos.

Example

The following graph shows the use of all three organizational unit types in an example directory hierarchy. Each node represents a new directory.

Schema of an example EDL directory hierarchy

Comparison with Related Projects

HDF5

The Hierarchical Data Format (HDF) is a file format to store large amounts of data in a single, structured file. Currently, HDF5 is the most-used container format for a variety of other file formats, and is especially useful to store array data.

While HDF5 is an excellent format to store structured data, and has the advantage of being just one file instead of a directory structure like EDL, it also has some disadvantages which may make EDL preferable in some applications:

• Writing in parallel to multiple datasets in a HDF5 file is very slow due to thread synchronization, making HDF5 not a great choice for massively parallel data acquisition.

• In case of an error while writing (e.g. due to a crash of the writing process), the entire HDF5 file may be corrupted, instead of just one file. EDL is a bit more robust.

• HDF5 has a limited set of ways to compress its array data. EDL, by using file-formats designed for the recorded modalities, can compress the stored data a lot more (e.g. by using the FFV1 video codec for video data).

HDF5’s advantages over EDL are it being a single file, its wide industry support and there being a single library to load the entirety of an HDF5 file (while EDL relies on other tools to read the specific data, e.g. FFmpeg or Neo).

Exdir

The Experimental Directory Structure (Exdir) is an open file format specification for experimental pipelines. Like EDL, it is modeled using the same abstractions that HDF5 uses as well.

Exdir and EDL are extremely similar to each other, and Syntalos used Exdir for a short time. However, there are some key differences:

• EDL uses TOML for metadata and Exdir uses YAML. While YAML is great to write for humans and sometimes less verbose than TOML, it has quite a lot of ambiguities and pitfalls for humans and programs to run into, which sometimes makes it more difficult to be used for structured data storage.

• Exdir adheres fairly strictly to the HDF5 abstract data model and the data types that HDF5 supports. While arrays are well-supported as Numpy arrays, there are no standards for “raw” datasets (containing videos or images). For use in a DAQ system like Syntalos, the “raw” dataset type would have to be massively extended.

• Each Exdir dataset can only contain one format, while EDL supports “auxiliary data” that describes the contained data further and is not pure textual TOML metadata. This auxiliary data may for example be timestamps for a video file. In Exdir, this data would have needed to be split into two datasets, which users found very confusing. EDL allows strongly linked data to be kept together in one directory.

• Exdir has no unique ID for the dataset, and also does not mandate the ID to be added as metadata to files contained in its structure. This is a feature that we really wanted for EDL, and that was a bit difficult to retrofit into Exdir.

Overall, Exdir is a fine format that simply has a few less features compared to what we needed for Syntalos. Its advantage over EDL is that is maps a lot more cleanly to HDF5 data structures and probably needs less processing to be converted into HDF5 than EDL would.

NWB

The Neurodata Without Borders format is a HDF5-based file format and structure to store neuroscientific data and make it easy to be interchanged between many groups in the neuroscientific community. It being based on HDF5 makes it not very suitable for massively parallel direct-write data acquisition though, which is why Syntalos and other DAQ systems can’t use it directly.

However, some groups do convert EDL into NWB-HDF5 files after the data acquisition has been completed, which is a good approach to get the best of both worlds.