Skip to main content


Flow stores data in collections: real-time data lakes of JSON documents. Collections may be added to by captures, or be derived as a transformation of other source collections.

Every collection has a key and an associated schema that its documents must validate against.


Collections are expressed within a Flow catalog specification:

# A set of collections to include in the catalog.
# Optional, type: object
# The unique name of the collection.

# The schema of the collection, against which collection documents
# are validated. This may be an inline definition or a relative URL
# reference.
# Required, type: string (relative URL form) or object (inline form)
schema: anvils.schema.yaml

# The key of the collection, specified as JSON pointers of one or more
# locations within collection documents. If multiple fields are given,
# they act as a composite key, equivalent to a SQL table PRIMARY KEY
# with multiple table columns.
# Required, type: array
key: [/product/id]

# Projections and logical partitions for this collection.
# See the "Projections" concept page to learn more.
# Optional, type: object

# Derivation that builds this collection from others through transformations.
# See the "Derivations" concept page to learn more.
# Optional, type: object


Every Flow collection must declare a schema, and will never accept documents that do not validate against the schema. This helps ensure the quality of your data products and the reliability of your derivations and materializations. Schema specifications are flexible: yours could be exactingly strict, extremely permissive, or somewhere in between.

Schemas may either be declared inline, or provided as a reference to a file. References can also include JSON pointers as a URL fragment to name a specific schema of a larger schema document:

type: object
required: [id]
id: string
key: [/id]

Learn more about schemas


Every Flow collection must declare a key which is used to group its documents. Keys are specified as an array of JSON pointers to document locations. For example:

schema: schema.yaml
key: [/userId]

Suppose the following JSON documents are captured into acmeCo/users:

{"userId": 1, "name": "Will"}
{"userId": 1, "name": "William"}
{"userId": 1, "name": "Will"}

As its key is [/userId], a materialization of the collection into a database table will reduce to a single row:

userId | name
1 | Will

If its key were instead [/name], there would be two rows in the table:

userId | name
1 | Will
1 | William

Schema restrictions

Keyed document locations may be of a limited set of allowed types:

  • boolean
  • integer
  • string

Excluded types are:

  • array
  • null
  • object
  • Fractional number

Keyed fields also must always exist in collection documents. Flow performs static inference of the collection schema to verify the existence and types of all keyed document locations, and will report an error if the location could not exist, or could exist with the wrong type.

Flow itself doesn't mind if a keyed location could have multiple types, so long as they're each of the allowed types: an integer or string for example. Some materialization connectors, however, may impose further type restrictions as required by the endpoint. For example, SQL databases do not support multiple types for a primary key.

Composite Keys

A collection may have multiple locations which collectively form a composite key. This can include locations within nested objects and arrays:

schema: schema.yaml
key: [/foo/a, /foo/b, /foo/c/0, /foo/c/1]

Key behaviors

A collection key instructs Flow how documents of a collection are to be reduced, such as while being materialized to an endpoint. Flow also performs opportunistic local reductions over windows of documents (also called "combines") to improve its performance and reduce the volumes of data at each processing stage.

An important subtlety is that the underlying storage of a collection will potentially retain many documents of a given key.

In the acmeCo/users example, each of the "Will" or "William" variants is likely represented in the collection's storage — so long as they didn't arrive so closely together that they were locally combined by Flow. If desired, a derivation could re-key the collection on [/userId, /name] to materialize the various /names seen for a /userId.

This property makes keys less lossy than they might otherwise appear, and it is generally good practice to chose a key that reflects how you wish to query a collection, rather than an exhaustive key that's certain to be unique for every document.


Projections are named locations within a collection document that may be used for logical partitioning or directly exposed to databases into which collections are materialized.

Many projections are automatically inferred from the collection schema. The projections stanza can be used to provide additional projections, and to declare logical partitions:

schema: anvils.schema.yaml
key: [/product/id]

# Projections and logical partitions for this collection.
# Keys name the unique projection field, and values are its JSON Pointer
# location within the document and configure logical partitioning.
# Optional, type: object
# Short form: define a field "product_id" with document pointer /product/id.
product_id: "/product/id"

# Long form: define a field "metal" with document pointer /metal_type
# which is a logical partition of the collection.
location: "/metal_type"
partition: true

Learn more about projections.


Collections are real-time data lakes. Historical documents of the collection are stored as an organized layout of regular JSON files in your cloud storage bucket. Reads of that history are served by directly reading files from your bucket.

Your storage mappings determine how Flow collections are mapped into your cloud storage buckets.

Unlike a traditional data lake, however, it's very efficient to read collection documents as they are written. Derivations and materializations that source from a collection are notified of its new documents within milliseconds of their being published.

Learn more about journals, which provide storage for collections