OrioleDB

En lisant la release note v3.0.3 de wal-g, j'ai découvert l'extension PostgreSQL nommée OrioleDB.

OrioleDB is a new storage engine for PostgreSQL, bringing a modern approach to database capacity, capabilities and performance to the world's most-loved database platform.

OrioleDB consists of an extension, building on the innovative table access method framework and other standard Postgres extension interfaces. By extending and enhancing the current table access methods, OrioleDB opens the door to a future of more powerful storage models that are optimized for cloud and modern hardware architectures.

Le projet OrioleDB a commencé en février 2022 par un développeur de Supabase : Alexander Korotkov.

Les commentaires de ce thread Hacker News semblent très enthousiastes https://news.ycombinator.com/item?id=30462695.

Dans la page "Introductions" de la documentation, je lis :

Differentiators

The key technical differentiations of OrioleDB are as follows:

No buffer mapping and lock-less page reading

In-memory pages in OrioleDB are connected with direct links to the storage pages. This eliminates the need for in-buffer mapping along with its related bottlenecks. Additionally, in OrioleDB in-memory page reading doesn't involve atomic operations. Together, these design decisions bring vertical scalability for Postgres to the whole new level.

MVCC is based on the UNDO log concept

In OrioleDB, old versions of tuples do not cause bloat in the main storage system, but eviction into the undo log comprising undo chains. Page-level undo records allow the system to easily reclaim space occupied by deleted tuples as soon as possible. Together with page-mergins, these mechanisms eliminate bloat in the majority of cases. Dedicated VACUUMing of tables is not needed as well, removing a significant and common cause of system performance deterioration and database outages.

Copy-on-write checkpoints and row-level WAL

OrioleDB utilizes copy-on-write checkpoints, which provides a structurally consistent snapshot of data every moment of time. This is friendly for modern SSDs and allows row-level WAL logging. In turn, row-level WAL logging is easy to parallelize (done), compact and suitable for active-active multimaster (planned).

J'ai lu le billet "Rethinking PostgreSQL buffer mapping for modern hardware architectures". Je pense avoir compris que l'implémentation actuelle de PostgreSQL utilise un "buffer mapping" autrefois bien adapté aux contraintes matérielles.

J'ai compris qu'OrioleDB propose une nouvelle approche, spécialement conçue pour tirer parti des SSD rapides, ce qui lui permet d’atteindre des performances nettement supérieures à celles de l’implémentation existante.