Data orientation refers to how tabular data is represented in a linear memory model such as in-disk or in-memory.The two most common representations are column-oriented (columnar format) and row-oriented (row format).[1] [2]
The choice of data orientation is a trade-off and a architectural decision in databases, query engines, and numerical simulations.[1] As a result of these tradeoffs, row-oriented formats are more commonly used in Online transaction processing (OLTP)and column-oriented formats are more commonly used in Online analytical processing (OLAP).[2]
Examples of column-oriented formats include Apache ORC,[3] Apache Parquet,[4] Apache Arrow,[5] formats used by BigQuery, Amazon Redshift and Snowflake.Predominant examples of row-oriented formats include CSV, formats used in most relational databases, in-memory format of Apache Spark, and Apache Avro.[6]
Tabular data is two dimensional in nature - data is represented in rows and columns.However, modern operating systems logically represent data in a linear memory model, both in-disk and in-memory.[7] [8] [9] Therefore, a table in a linear memory model requires projecting its two-dimensional items in a one-dimensional space.Data orientation refers to the decision taken in this projection.There are two prominent choices of orientation: row-oriented and column-oriented.[1] [2]
In row-oriented, the elements of the table
| column 1 | column 2 | column 3 | |
|---|---|---|---|
| item 11 | item 12 | item 13 | |
| item 21 | item 22 | item 23 |
are stored linearly as
| item 11 | item 12 | item 13 | item 21 | item 22 | item 23 |
I.e. each row of the table is located one after the other.In this orientation, values on the same row are close in space (e.g. similar address in an addressable space).
In column-oriented, the elements of the table
| column 1 | column 2 | column 3 | |
|---|---|---|---|
| item 11 | item 12 | item 13 | |
| item 21 | item 22 | item 23 |
are stored linearly as
| item 11 | item 21 | item 12 | item 22 | item 13 | item 23 |
I.e. each column of the table is located one after the other.In this orientation, values on the same column are close in space (e.g. similar address in an addressable space).
See list of column-oriented DBMSes for more examples.
The data orientation is an important architectural decision of systems handling data because it results in important tradeoffs in performance and storage.[8] Below are selected dimensions of this tradeoff.
Row-oriented benefits from fast random access of rows.Column-oriented benefits from fast random access of columns.In both cases, this is the result of less page or cache misses when accessing the data.[8]
Row-oriented benefits from fast insertion of a new row.Column-oriented benefits from fast insertion of a new column.
This dimension is an important reason why row-oriented formats are more commonly used in Online transaction processing (OLTP), as it results in faster transactions in comparison to column-oriented.[2]
Row-oriented benefits from fast access under a filter.Column-oriented benefits from fast access under a projection.[4] [3]
Column-oriented benefits from fast analytics operations.This is the result of being able to leverage SIMD instructions.[5]
Column-oriented benefits from smaller uncompressed size.This is the result of the possibility that this orientation offers to represent certain data types with dedicated encodings.[4] [3]
For example, a table of 128 rows with a boolean column requires 128 bytes a row-oriented format (one byte per boolean)but 128 bits (16 bytes) in a column-oriented format (via a bitmap).Another example is the use of run-length encoding to encode a column.
Column-oriented benefits from smaller compressed size.This is the result of a higher homogeneity within a column than within multiple rows.[4] [3]
Because both orientations represent the same data, it is possible to convert a row-oriented dataset to a column-oriented dataset and vice-versa at the expense of compute.In particular, advanced query engines often leverage each orientation's advantages, and convert from one orientation to the other as part of their execution.As an example, an Apache Spark query may