Struct data type
This document defines struct, a data type for arrays whose elements are fixed-size records composed of named, typed fields — commonly referred to as "structured arrays" or "record arrays".
The struct data type closely models NumPy's structured arrays, where each element consists of multiple named fields, each with its own data type.
Background
Structured arrays allow a single array to represent tabular or record-like data without using separate arrays per column. For example, a geospatial dataset might store (latitude: float64, longitude: float64, elevation: float32) as a single structured array rather than three separate arrays.
Each element of a structured array is a scalar of a fixed size in bytes. The size is the sum of the sizes of all fields.
Data type representation
Name
The name of this data type is the string "struct".
Note: Implementations SHOULD also support reading arrays that use the legacy
"structured"data type name, including legacy field definitions in tuple form (see the"structured"specification for details).
Configuration
This data type requires a configuration object. The configuration object MUST contain a "fields" key whose value is an ordered JSON array of field objects.
Each field MUST be a JSON object with the following keys:
"name": A non-empty string that identifies the field. Field names MUST be unique within the samestructdata type; nestedstructtypes have independent namespaces."data_type": A valid Zarr v3 data type representation whose size in bytes is fixed and known at the time the array is opened:- For core data types, this MUST be a string (e.g.
"float32","int32","uint8"). - For extension data types that require configuration (e.g.
numpy.datetime64), this MUST be an object with a"name"key and a"configuration"key. - Variable-length data types (e.g.
"string") MUST NOT be used as field types, as they do not have a fixed encoded size.
- For core data types, this MUST be a string (e.g.
The "fields" array MUST contain at least one field. The order of fields in the array is significant: it defines the encoding order for the bytes codec and MUST be preserved by implementations.
The struct data type MAY be used recursively: a field's data type MAY itself be "struct", enabling nested record types.
Examples
The following is an example of array metadata for an array of 2D point records, each with an x and a y coordinate stored as 32-bit floats:
{
"zarr_format": 3,
"node_type": "array",
"shape": [100],
"data_type": {
"name": "struct",
"configuration": {
"fields": [
{"name": "x", "data_type": "float32"},
{"name": "y", "data_type": "float32"}
]
}
},
"chunk_grid": {
"name": "regular",
"configuration": {"chunk_shape": [100]}
},
"chunk_key_encoding": {"name": "default"},
"fill_value": {"x": 0.0, "y": 0.0},
"codecs": [{"name": "bytes", "configuration": {"endian": "little"}}]
}The remaining examples show only the configuration object for brevity.
The following shows a field using a parametrized data type. The timestamp field uses numpy.datetime64, which requires a configuration object specifying unit and scale_factor:
{
"fields": [
{
"name": "timestamp",
"data_type": {
"name": "numpy.datetime64",
"configuration": {
"unit": "s",
"scale_factor": 1
}
}
},
{"name": "value", "data_type": "float32"}
]
}The following shows a nested struct field. The outer record has a point field that is itself a struct type with x and y sub-fields, plus a scalar value field:
{
"fields": [
{
"name": "point",
"data_type": {
"name": "struct",
"configuration": {
"fields": [
{"name": "x", "data_type": "float32"},
{"name": "y", "data_type": "float32"}
]
}
}
},
{"name": "value", "data_type": "float64"}
]
}Bytes codec encoding
When the bytes codec is used, each struct scalar is encoded as the packed concatenation of the encoded bytes of each field's value, in field declaration order. No padding bytes are inserted between fields, regardless of alignment considerations.
For nested struct types, encoding proceeds depth-first: each field is encoded completely before the next sibling field. A nested struct field is encoded as the packed concatenation of its own sub-fields, recursively.
The total encoded size of a struct scalar in bytes is the sum of the encoded sizes of all fields.
As a concrete example, consider the following struct type:
{
"fields": [
{"name": "id", "data_type": "int32"},
{"name": "flags", "data_type": "uint8"},
{"name": "value", "data_type": "float64"}
]
}The encoded byte layout is:
byte: 0 1 2 3 4 5 6 7 8 9 10 11 12
├───────────────┼───┼───────────────────────────────┤
field:│ id │ f │ value │
│ (int32) │ │ (float64) │
└───────────────┴───┴───────────────────────────────┘The total encoded size is 4 + 1 + 8 = 13 bytes, with field byte offsets of 0, 4, and 5 respectively.
For nested struct types, the same principle applies recursively. Consider:
{
"fields": [
{
"name": "point",
"data_type": {
"name": "struct",
"configuration": {
"fields": [
{"name": "x", "data_type": "float32"},
{"name": "y", "data_type": "float32"}
]
}
}
},
{"name": "value", "data_type": "float64"}
]
}The nested point field is encoded completely (all sub-fields) before value:
byte: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
├───────────────┼───────────────┼───────────────────────────────┤
field:│ point.x │ point.y │ value │
│ (float32) │ (float32) │ (float64) │
└───────────────┴───────────────┴───────────────────────────────┘The total encoded size is (4 + 4) + 8 = 16 bytes, with point at offset 0 (containing x at 0, y at 4) and value at offset 8.
Endianness
When a struct type contains multi-byte numeric fields, the bytes codec MUST be configured with an explicit endian setting (e.g. {"name": "bytes", "configuration": {"endian": "little"}}). All multi-byte fields MUST use the byte order specified by the endian parameter.
Struct types composed entirely of single-byte fields (e.g. uint8, int8) have no byte-order ambiguity and MAY omit the endian configuration.
Fill value representation
The fill_value for arrays with the struct data type MUST be a JSON object mapping each field name to its fill value. Every field defined in the struct type MUST have a corresponding entry in the fill value object. Each field's value MUST be a valid fill value for that field's data type.
Note: Explicit fill values are required for all fields because implicit defaults (such as "zero") are ambiguous for extension data types. For example, a NumPy
datetime64initialized with zero represents the Unix epoch, whereas a C++ datetime implementation might default to a null or invalid state. Requiring explicit fill values ensures consistent behavior across implementations and future extensions.
"fill_value": {"x": 1.23, "y": 4.56}For nested struct fields, the value must itself be an object mapping the nested field names to their fill values:
"fill_value": {"point": {"x": 1.0, "y": 2.0}, "value": 3.14}Codec compatibility
The struct data type works with any array-to-bytes codec that encodes each array element as a contiguous, fixed-size binary blob equal to the total encoded size of the struct. The bytes codec is the standard choice for this role. Other array-to-bytes codecs (for example, a future per-field encoding codec) MAY be used provided they satisfy this constraint. Byte-manipulation codecs (e.g. gzip, zstd, blosc) MAY be applied on top for compression.
Variable-length codecs (e.g. vlen-utf8) are not compatible with the struct data type.
Notes
Note: Fields are packed contiguously with no inter-field padding. This matches the default NumPy structured dtype layout. NumPy's "aligned" structured dtype layout (created with
align=True), which inserts padding for memory alignment, is NOT supported by this specification.
Note: Field names MUST be unique within a
structdata type. Implementations MUST rejectstructtypes with duplicate field names.
Note: The order of fields in the
"fields"array is significant and MUST be preserved. Fields are encoded in declaration order.
Change log
- Relaxed the JSON schema to permit field data types declared as a JSON object without a
configurationkey, e.g.{"name": "float64"}(#63).