S3 Bucket Organization
You can think of your data buckets as a grid. The horizontal dimension reflects the data lifecycle phase (raw, refined, curated). The vertical dimension reflects topic or domain.
There's no one folder structure that works for everyone. For instance, if you organize experiments by
department/datethen it's tedious for users to look at everything that happened on a given date. And vice versa. Since there's no one folder structure that works for everyone in the business, files are copied so that they can be found through multiple folder paths. But copies reduce uniqueness. As a result users don't know which copy of the file to trust. These copies then diverge, further reducing trust in data.
As users try to keep files organized they put metadata in the file names and paths. This is fragile. First, files and folders can be moved, thus severing their connection to metadata stored in the path. Second, file names with metadata can only get so long before they become unusable. Finally, metadata and folder-naming conventions vary across users, making metadata in the path of low consistency, low quality, and low trust.
We call file paths physical views because they are fixed addresses for data on disk or in blob storage.
For the above reasons reason, packages offer both metadata tags and infinite logical views atop physical locations in S3. Metadata tags prevent file names from getting longer and longer to hackily include metadata.
With Quilt packages, you can include a given S3 object in as many packages without ever copying that object. Or you can organize multiple S3 buckets into single packages under any logical folder structure that you wish (by using the Quilt catalog or the
Since packages separate metadata from logical view paths, but include metadata and data in the package manifest, data and metadata can never be accidentally separated. File paths also become simpler and easier to trust.
Below is an example of how you might organize three data domains according to a three-phase data lifecycle.
In addition to the three lifecycle phases above it is useful to have on or more "sandbox" buckets where users can create experimental packages without fear of overwriting or disrupting business data.
Generally, you'll need three S3 buckets for each data domain.
Below are a few questions you can use to determine how many data domains and lifecycle phases you'll need:
- 1.Are these data related in a meaningful way?If not, favor separate buckets.
- 2.Are these data all a the same level of "doneness"; are they at the same stage in the data quality lifecycle (raw, refined, curated)?
- 4.Do humans or pipelines have delete permissions against these buckets?If so, data should be duplicated to another bucket on push to ensure durability.
- 5.Does this data need to be sealed and linked to an IND filing or ELN?If so, favor curated or "production" buckets where all data are self-contained and delete access is exceptional or non-existent.
- 6.Are we mixing validated and non-validated data?If so, favor segregation at the bucket level.
- 7.Is this data part of a staging or dev environment?If so, consider how this data will be copied to more stable buckets, or why it is OK to lose this data at any time.
- 8.Do you want to be able to trace the data provenance back to its source, for example from instrument, to scientist, to ELN entry?If so, consider isolated, only-instruments-can-write buckets for the early data lifecycle, and read-only or write-only buckets for the curated (ELN) phase.