What Happened At The Posette Event for Postgres 2026
Written by Nikos Vaggalis
Friday, 31 July 2026
Posette is a recurring event run by the Postgres team at Microsoft. that brings together Postgres experts, developers, users and contributors from around the globe to present all things Postgres. 2026 marks its 5th year.
This year the event's topics highlighted modern strategies for scaling, securing and managing databases in AI-driven and cloud-native environments, and utilizing the Model Context Protocol to connect AI agents with data while enforcing safety layers.
Security was framed as a proactive design habit, emphasizing the least-privilege principle, schema isolation and the hardening of function execution paths. Of course there was coverage of AI integrated tools and the terminal that simplify database administration through natural language and visual execution plans.
Following is a brief overview of the main themes discussed:
AI And modern data workloads A major focus was on integrating Postgres with AI, including building Model Context Protocol servers to allow AI agents to safely and reliably query databases. Speakers highlighted using Azure HorizonDB for Production RAG pipelines and leveraging Postgres as an "Everything Database" that handles vectors, JSON and graphs to eliminate fragile ETL pipelines. Extensions like pg_duckdb and pg_lake were also showcased for bringing fast, OLAP analytics and data lakehouse capabilities directly into Postgres.
Performance, infrastructure and scaling Several talks emphasized the critical role of hardware and tuning. Sessions highlighted that moving from network attached storage to local NVMe storage dramatically reduces disk I/O bottlenecks for vacuums and queries. Additional operational strategies included speeding up Citus cluster scaling via snapshot based node additions and how to effectively maintain large tables through partitioning and tuning autovacuum.
New Postgres features (Versions 17-19) Contributors detailed massive improvements in recent and upcoming Postgres versions. Postgres 18 introduces asynchronous I/O for vacuuming, dynamic worker scaling and eager freezing to handle table bloat efficiently. It also brings new constraint features like temporal primary keys (to prevent overlapping time periods) and NOT ENFORCED constraints, alongside major query planner optimizations like B-tree skip scans and automatic self-join eliminations. For PostgreSQL 19, hackers are actively laying the groundwork for a multithreaded architecture and built in support for querying property graphs. Security, replication and internals Security discussions covered the evolution of Postgres authentication from basic trust parameters to modern token-based identity systems. Experts recommended securing production environments through least privilege roles, schema isolation and search path hardening, as well as finding edge case vulnerabilities via fuzzing. Deep technical dives explored the internals of the Write-Ahead Log (WAL) , the lifecycle of an INSERT statement through the reorder buffer and building event-driven architectures using logical replication and CDC tools like Debezium, wal2json and Drasi.
Developer tooling and application patterns Finaly for application developers, talks covered practical design patterns like using range types, exclusion constraints and PostGIS to solve complex business logic. Speakers warned about the risks of accidentally serializing concurrent transactions when misusing LISTEN/NOTIFY and offered better alternatives for robust in-database queues. Ecosystem tooling updates included measuring SQL test coverage with a new tool called pgcov and managing Postgres natively across AI-powered IDEs like Cursor and VS Code.
Fortunately for those who couldn't attend, the talks have been uploaded in total on YouTube. Since there are many interesting talks, we singled out a few:
"An MCP for your Postgres DB", explores the cutting-edge integration of AI agents with databases using the newly adopted Model Context Protocol. This talk is exceptionally interesting because it tackles the very real danger of giving AI models unrestricted access to your data. The speaker demonstrates how an AI model without proper guardrails might carelessly execute a command to delete 20,000 rows just because a user casually suggested it.
The session provides highly practical solutions for safely taming these agents, such as using progressive schema discovery to avoid overwhelming the LLM's context window and enforcing strict read only database roles to guarantee AI cannot mutate your production data.
"Fuzzing PostgreSQL" takes an adversarial approach to database testing. Rather than testing a database with expected inputs, fuzzing throws continuously mutated garbage data at the system to see where its assumptions break. It is particularly interesting because it frames Postgres not just as a data store, but as a massive, overlapping attack surface similar to an operating system or a web browser. The talk highlights how this automated, continuous fuzzing helps uncover deep, non obvious vulnerabilities in complex components like the SQL parser and network protocols to ensure bulletproof reliability.
And "LISTEN carefully: How notify can trip up your database", serves as an excellent real world horror story about the unintended consequences of using seemingly elegant built-in features at scale. The talk details how the lightweight LISTEN/NOTIFY asynchronous messaging mechanism caused a massive production bottleneck as customer traffic grew. It breaks down the root cause, which is putting NOTIFY inside a trigger inadvertently forced concurrent transactions to queue up sequentially, creating accidental serialization as they all competed for an access exclusive lock on a global catalog table. The creative workaround, decoupling the notification process using an unlogged queue table combined with transaction level advisory locks, offers a brilliant lesson in advanced PostgreSQL performance architecture.
That's not all, however. These are just three of the noteworthy talks. There's many more on the playlist so you'd better start watching now!
Researchers at the University of California San Diego are looking at ways to give old mobile phones a second life by using them for “phone cluster computing.”