Mental model
Quilt represents datasets as packages. A package is an immutable collection of related files with a handle of the form AUTHOR/DESCRIPTION
, a cryptographic top-hash (or hash of hashes) that uniquely identifies package contents, and a backing manifest.
The manifest is serialized as file that contains entries. Manifest entries are tuples of the following form:
(LOGICAL_KEY, PHYSICAL_KEYS, HASH, METADATA)
Logical keys are user-facing friendly names, like "README.md"
. Physical keys are fully qualified paths to bytes on disk, or bytes in S3. A hash is a digest of the physical key's contents, usually SHA-256. Metadata are a dictionary that may contain user-defined keys for metadata like bounding boxes, labels, or provenance information (e.g. {"algorithm_version": "4.4.1"} to indicate how a given file was created).
Package manifests are stored in registries. Quilt supports both local disk and Amazon S3 buckets as registry. A registry may store manifests as well as the primary data. S3 was chosen for its widespread adoption, first-class versioning support, and cost/performance profile. The Quilt roadmap includes plans to support more storage formats in the future (e.g. GCP, Azure, NAS, etc.).
By way of illustration first entry of a package manifest for the COCO machine learning dataset are shown below.
Buckets are branches
In Quilt, S3 buckets are analogous to branches in git. Each bucket is a self-contained registry for one or more packages. As package data and schemas are refined, you can promote a package to a new bucket to signify its increased data quality.
We generally recommend a minimum of three buckets for the data lifecycle:
Raw
Stage
Production
See Quilt workflows for more on how you can control data quality with schemas.
Last updated