4 ms·
While row-oriented and column-oriented storage models as laid out in the article are classic database organization patterns, modern databases engines often use
by jandrewrogers 11y ago
While row-oriented and column-oriented storage models as laid out in the article are classic database organization patterns, modern databases engines often use neither as their primary storage model. They still have their use cases, pure OLTP for row-oriented and cold storage for column-oriented, but many workloads these days have a high insert rate concurrent with light analytic queries. This workload shows poor efficiency on either storage model.
Designing a general-purpose database engine today, I would probably use some variant of the vectorized row storage model. Rows are stored within a single page but are decomposed within that page into vectors, the organization of which can be quite complex and may not require any significant row reconstruction. Page sizes are usually larger than OLTP (64-256k versus 4-16k) which makes it suited to modern disk storage.
Vectorized row storage often approaches the insert performance of row-oriented and the light analytics performance of column-oriented if designed and implemented competently. Column-oriented is often more space efficient for cold data storage but depending on the use case may not be significantly faster for queries.
Classic row-oriented and column-oriented designs are at opposite extremes of a set of tradeoffs. Many newer database engines sit at neither extreme while enjoying most of the benefits of both.