Input and Output

Where are my data?

New users often ask this question thinking of passive data repositories – collections of files and folders and a separate collection of metadata – information about how the files were collected and what they contain. Let’s address metadata first, since the answer there is easy: Everything goes in the database! Any information about the experiment that would normally be stored in a lab notebook, in an Excel spreadsheet, or in a Word document is entered into tables in the database. These tables can accommodate numbers, strings, dates, or numerical arrays. The entry of metadata can be manual, or it can be an automated part of data acquisition (in this case the acquisition software itself is modified to enter information directly into the database).

Depending on their size and contents, raw data files can be stored in a number of ways. In the simplest and most common scenario, raw data continue to be stored in either a local filesystem or in the cloud as collections of files and folders. The paths to these files are entered in the database (again, either manually or by automated processes). This is the point at which the notion of a data pipeline begins. Below these “manual tables” that contain metadata and file paths are a series of tables that load raw data from these files, process it in some way, and insert derived or summarized data directly into the database. For example, in an imaging application, the very large raw .TIFF stacks would reside on the filesystem, but the extracted fluorescent trace timeseries for each cell in the image would be stored as a numerical array directly in the database. Or the raw video used for animal tracking might be stored in a standard video format on the filesystem, but the computed X/Y positions of the animal would be stored in the database. Storing these intermediate computations in the database makes them easily available for downstream analyses and queries.

Do I have to manually enter all my data into the database?

No! While some of the data will be manually entered (the same way that it would be manually recorded in a lab notebook), the advantage of DataJoint is that standard downstream processing steps can be run automatically on all new data with a single command. This is where the notion of a data pipeline comes into play. When the workflow of cleaning and processing the data, extracting important features, and performing basic analyses is all implemented in a DataJoint pipeline, minimal effort is required to analyze newly-collected data. Depending on the size of the raw files and the complexity of analysis, useful results may be available in a matter of minutes or hours. Because these results are stored in the database, they can be made available to anyone who is given access credentials for additional downstream analyses.

Won’t the database get too big if all my data are there?

Typically, this is not a problem. If you find that your database is getting larger than a few dozen TB, DataJoint provides transparent solutions for storing very large chunks of data (larger than the 4 GB that can be natively stored as a LONGBLOB in MySQL). However, in many scenarios even long time series or images can be stored directly in the database with little effect on performance.

Why not just process the data and save them back to a file?

There are two main advantages to storing results in the database. The first is data integrity. Because the relationships between data are enforced by the structure of the database, DataJoint ensures that the metadata in the upstream nodes always correctly describes the computed results downstream in the pipeline. If a specific experimental session is deleted, for example, all the data extracted from that session are automatically removed as well, so there is no chance of “orphaned” data. Likewise, the database ensures that computations are atomic. This means that any computation performed on a dataset is performed in an all-or-none fashion. Either all of the data are processed and inserted, or none at all. This ensures that there are no incomplete data. Neither of these important features of data integrity can be guaranteed by a file system.

The second advantage of storing intermediate results in a data pipeline is flexible access. Accessing arbitrarily complex subsets of the data can be achieved with DataJoint’s flexible query language. When data are stored in files, collecting the desired data requires trawling through the file hierarchy, finding and loading the files of interest, and selecting the interesting parts of the data.

This brings us to the final important question:

How do I get my data out?

This is the fun part. See Queries for details of the DataJoint query language directly from MATLAB and Python.

Interfaces

Multiple interfaces may be used to get the data into and out of the pipeline.

Some labs use third-party GUI applications such as HeidiSQL and Navicat, for example. These applications allow entering and editing data in tables similarly to spreadsheets.

The Helium Application (https://mattbdean.github.io/Helium/ and https://github.com/mattbdean/Helium) is web application for browsing DataJoint pipelines and entering new data. Matt Dean develops and maintains Helium under the direction of members of Karel Svoboda’s lab at Janelia Research Campus and Vathes LLC.

Data may also be imported or synchronized into a DataJoint pipeline from exising LIMS (laboratory information management systems). For example, the International Brain Lab synchronizes data from an Alyx database. For implementation details, see https://github.com/int-brain-lab/IBL-pipeline.

Other labs (e.g. Sinz Lab) have developed GUI interfaces using the Flask web framework in Python.

GUIs, Google Docs integration, LIMS integration, Slack integration, etc.