Skip to main content



Projections are an advanced concept of Flow. You can use Flow without knowing the details of projections, but this section may help you better understand how Flow works.

Flow documents are arbitrary JSON, and may contain multiple levels of hierarchy and nesting. However, systems that Flow integrates with often model flat tables of rows and columns, without hierarchy. Others are somewhere in between.

Projections are the means by which Flow translates between the documents of a collection and a table representation. A projection defines a mapping between a structured document location, given as a JSON-Pointer, and a corresponding field name used as, for example, a CSV file header or SQL table column.

Many projections are inferred automatically from a collection’s JSON schema, using a field that is simply the JSON Pointer with its leading slash removed. For example, a schema scalar with pointer /myScalar will generate a projection with field myScalar.

You can supplement by providing additional collection projections, and a document location can have more than one projection field that references it:

schema: session.schema.yaml
key: [/user/id, /timestamp]
# A "user/id" projection field is automatically inferred.
# Add an additional field that doesn't have a slash.
user_id: /user/id
# Partly decompose a nested array of requests into a handful of named projections.
"first request": /requests/0
"second request": /requests/1
"third request": /requests/2

Logical partitions

Projections can also be used to logically partition a collection, specified as a longer-form variant of a projection definition:

schema: session.schema.yaml
key: [/user/id, /timestamp]
location: /country
partition: true
location: /agent/type
partition: true
location: /agent/network
partition: true

Logical partitions isolate the storage of documents by their differing values for partitioned fields. Flow extracts partitioned fields from each document, and every unique combination of partitioned fields is a separate logical partition.

Every logical partition has one or more physical partitions into which their documents are written, which in turn controls how files are arranged within cloud storage.

For example, a document of "acmeCo/user-sessions" like:

{"country": "CA", "agent": {"type": "iPhone", "network": "LTE"}, ...}

Might produce files in cloud storage like:


country, device, and network together identify a logical partition, while pivot identifies a physical partition. utc_date and utc_hour is the time at which the journal fragment was created.

Learn more about physical partitions.

Partition selectors

When reading from a collection, Flow catalog entities like derivations, materializations, and tests can provide a partition selector, which identifies the subset of partitions that should be read from a source collection:

# Partition selectors are included as part of a larger entity,
# such as a derivation or materialization.
# `include` selects partitioned fields and corresponding values that
# must be matched in order for a partition to be processed.
# All of the included fields must be matched.
# Default: All partitions are included. type: object
# Include partitions from North America.
country: [US, CA]
# AND where the device is a mobile phone.
device: [iPhone, Android]

# `exclude` selects partitioned fields and corresponding values which,
# if matched, exclude the partition from being processed.
# A match of any of the excluded fields will exclude the partition.
# Default: No partitions are excluded. type: object
# Skip sessions which were over a 3G network.
network: ["3G"]

Partition selectors are efficient as they allow Flow to altogether avoid reading documents that aren’t needed.