Known for its reliability and ability to manage complex data, the open source database management system has established itself as a benchmark for artificial intelligence applications.
As PostgreSQL celebrates its fortieth anniversary in 2026, the object-oriented relational database management system (SGBDRO) has never been so good. More than 55% of developers use it according to the latest annual Stack Overflow survey, ahead of MySQL, SQLite and Microsoft SQL Server. The adoption rate is even around 60% among professionals working on artificial intelligence projects. In terms of popularity, PostgreSQL comes fourth behind Oracle Database, MySQL and Microsoft SQL Server according to the famous DB-Engines ranking.
How to explain such enthusiasm? The advantages of Postgre, as its aficionados call it, are not lacking. First of all, it is a truly open source and community database without the shadow of a private company behind it, unlike MySQL (Oracle), MariaDB (K1 investment fund) and MongoDB (developed by the company of the same name). PostgreSQL is published under the PostgreSQL in-house license. A license, similar to BSD or MIT, that provides more protection than Server Side Public License (SSPL). It requires the source code to be made accessible when modifications or new functionalities are offered to third parties as a service.
PostgreSQL also relies on a particularly rich and organized community. It has several thousand contributors around the world, including employees of tech giants such as Amazon and Google. In France, the PosgreSQLFr association has set up an inter-company working group which brings together Air France, Carrefour, EDF, Maif, Orange and SNCF. The core team is made up of seven authoritative members including Peter Eisentraut and Bruce Momjian from EDB and Magnus Hagander from Redpill Linpro. The global development group approves the release of a new major version once a year, incorporating new features, and minor versions, for fixes, at least once a quarter. Released on September 25, PostgreSQL 18 notably improves the performance of workloads thanks to a new input-output subsystem.
This mode of governance is, for Adrien Nayrat, freelance PostgreSQL expert, a guarantee of trust. “The priority has always been the reliability and robustness of the engine. If certain features are inconsistent with the established roadmap, they are quickly discarded. This way, the code does not suffer from rushed or poorly integrated additions.” This preference given to reliability could harm performance, PostgreSQL having at one time been singled out for its slowness. MySQL, for example, met the needs for high reading performance of an application or website. PostgreSQL took over when the criteria of robustness, stability and compliance with SQL rules became essential.
This is no longer the case today, according to Lætitia Avrot, PostgreSQL & security practice leader at EDB. “PostgreSQL has no reason to be ashamed of its performance. In migration projects, it exceeds or equals Oracle in 90% of cases.” Over the years, the gap has indeed narrowed between Oracle Database and its open source alternative. One reason among others which explains why the database is found in almost all organizations, whatever their size. Hyperscalers (AWS, Google Cloud or Microsoft Azure) offer a version of PostgreSQL in their cloud as a managed service. PostgreSQL adoption is also made easier by the fact that it is also widely used for personal use.
An architecture designed to be extensible
PostgreSQL also makes it possible to manipulate large volumes of complex and heterogeneous data, including for critical use cases. The DBMS supports JSON and JSONB storage formats. Another key advantage: the PostgreSQL architecture was designed from the outset to be extensible. “This facilitates the development of advanced functionalities and allows it to adapt to new uses, starting with those of artificial intelligence,” says Lætitia Avrot. Among the extensions offered, one really stands out. As the name suggests, pgvector allows PostgreSQL to store, index, and query vector data.
“The principle consists of representing information, a text or an image, in the form of large digital vectors,” explains Adrien Nayrat. This vectorization and the creation of specific indexes allow calculations of approximate distances between the data. This approach paves the way for similarity search. Classically, an LLM will search for images similar to the one presented to it.
To illustrate the capabilities of pgvector, Lætitia Avrot cites the famous “Chihuahua Challenge”, the objective of which is to distinguish the photo of a chihuahua head from that of a muffin, the resemblance between the two being confusing. Similarity search finds use cases in the professional world. Among the most used algorithms in e-commerce, the nearest neighbor search method (k-NN) allows a merchant site to suggest products similar to those purchased by a customer. Pgvector opens the way to other use cases such as the detection of cases of fraud in the banking sector or the detection of anomalies during the analysis of medical images.
The centralization of complex and vector data made possible is a key element in the eyes of Adrien Nayrat. “Storing this data in the same place minimizes the risks of inconsistencies linked to synchronization between different sources. Any modification is automatically reflected. Updates, version upgrades or maintenance and backup operations are also facilitated.” There is no longer any need to use a vector database like Pinecone. PostgreSQL also offers full-text search functions, so the question of making it coexist with a specialized engine like Elasticsearch no longer arises.
A platform of choice for RAG
Launched in 2021, the pgvector extension has gained a solid reputation among the developer population, as evidenced by its more than 18,000 stars on GitHub. Despite this popularity, it is unlikely to be natively integrated into PostgreSQL. This very demanding integration process is rarely granted. pgvector is not the only AI-oriented extension in the PostgreSQL ecosystem. As its name suggests, pgvectorscale improves the performance and scalability of searches on vector data. Developed by Timescale, the pgai extension allows you to integrate AI capabilities directly into the database, such as authorizing calls to LLMs simply using SQL commands.
This on-board management of functionalities can impose PostgreSQL as the platform of choice for recovery-augmented generation or RAG (Retrieval-Augmented Generation) systems. Another extension, pgrag, aims to create pipelines dedicated to these RAGs without leaving the PostgreSQL environment. These last three extensions remain experimental and are far from reaching the level of maturity and adoption of pgvector.
PostgreSQL has, of course, some downsides. Like any open source solution, its installation, configuration, and maintenance can be complex for small organizations that do not have dedicated skills. Although its community is active, its support cannot match the commercial service of a proprietary solution like Microsoft SQL Server. Although PostgreSQL offers advanced security features, it lacks Transparent Data Encryption (TDE), i.e. data encryption at the storage layer. What companies like EDB or Percona offer to fill.
But the other databases have not said their last word. On October 24, MariaDB announced the availability of MariaDB Enterprise Platform 2026, which it presents as “the reference platform for creating the new generation of intelligent applications”. This major version notably integrates RAG pipelines and AI agents which interact with data in natural language. The battle for databases in the AI era is on.
