Skip to content

Vector types


Value vector is Kùzu’s column-oriented in-memory data structure to store a chunk of data of the same data type. The size of ValueVector is defined by DEFAULT_VECTOR_CAPACITY, which is 2048. This is an empirically selected value with the presumption that the data stored in the value vector will fit into CPU cache.

A value vector has the following core fields:

  • data: Stores the actual data which is a trivial byte array managed by a unique pointer.
  • nullMask: Aligned with data and indicate if each entry is NULL or not.
  • auxilaryBuffer: Keeps track of additional data that does NOT fit in data.

Primitive Type ValueVector

For primitive data type, we can represent data in the data field, for example, to represent 0 to 99 elements with INT64 data type, data is simply a int64_t[] with size 100.

String Type ValueVector

Earlier versions of the storage layer required each element to have a fixed size. This however, does NOT work with the STRING type, whose size may vary. We implement STRING as a 16 byte data structure with first 8 bytes as prefix and last 8 bytes as a pointer that points to a memory location storing the rest of string. The fixed length part of STRING is stored in data and the overflow part is stored in auxiliaryBuffer.

Nested Type ValueVector

Although nested type value vectors can be organized as overflow in the same way as STRING, we want to utilize the fact that the child type of a nested type is known and can still be stored in column-oriented value vector.

For STRUCT type, we store children vectors in auxilaryBuffer.


Data to represent:
[{11,12}, {13,14}, {15,16}]
STRUCT value vector
data: [0, 1, 2] // offset
child vector:
data: [11, 13, 15]
child vector
data: [12, 14, 16]

For VAR_LIST type, we store size and offset of each entry in data and elements in auxilaryBuffer.


Data to represent:
[[10], [11,12], [13,14,15]]
List value vector
data: [(0,1), (1,2), (3,3)] // offset,size
child vector:
data: [10,11,12,13,14,15]


Selection state keeps track of data that are valid. This is mainly used to select a subset of data (e.g. data that satisfy a filter) without performing a copy.

A selection state has the following core fields:

  • selectedPositions: a vector of uint16_t that maps to a position of data vector.
  • selectedSize: size of selected positions


Value vector
data: [10, 11, 12, 13, 14]
Selection state:
selectedPositions: [0, 2]
selectedSize: 2
Data being represented:
[10, 12]


A data chunk is a collection of value vector with the same state.


Data chunk
Value vector: [1, 2, 3]
Value vector: [a, b, c]
Data being represented:
[(1,a), (2,b), (3,c)]


A result set is a collection of data chunks that forms a cartesian product.


Result set
Data chunk
Value vector: [1, 2, 3]
Value vector: [a, b, c]
Data Chunk
Value vector: [10, 11]
Data being represented:
[(1,a,10), (2,b,10), (3,c,10), (1,a,11), (2,b,11), (3,c,11)]