PGConf.Russia 2025

Diagnostic Methods: Analyzing the Issue of Long CREATE DATABASE Execution

This talk covers two cases of long CREATE DATABASE execution in PostgreSQL 11, which, in a Corosync/Pacemaker cluster configuration, caused node failover due to a monitoring timeout from the clustering software. The first case is related to a large shared buffers size, while the second involves the use of the WITH OWNER option during database creation and a large pg_shdepend system table (a feature of using the jBPM software).

Poolers play a crucial role in PostgreSQL database operations. In this talk, we’ll discuss the challenges we faced when using and implementing existing poolers and the benefits we gained by developing our own pooler.

I’ll talk about the challenges you’ll face if you decide to implement a new TableAM. What to choose: Generic XLog or Custom RMGR? Why use a Custom SMGR? How to integrate PostgreSQL allocators into third-party libraries, even if they don't officially support it? What’s missing for a columnar engine (including vectorization and late materialization), and how can we work around those limitations?

In the second part, I’ll dive into the internals of pgpro_tam — a new native table engine for OLAP that supports standard data formats, various SMGRs, and, if needed, third-party schedulers and execution engines, all while adhering to ACID principles. This is designed to achieve the fastest analytics on PostgreSQL (not just plugging in DuckDB).

In PostgreSQL performance testing, benchmarks measure query execution time (latency). To get more reliable results, queries are executed repeatedly, generating a dataset of latency values. Performance is often assessed using standard metrics like the median or mean, but we propose a more advanced approach.

In practice, latency distributions are often multimodal, consisting of multiple underlying distributions with distinct characteristics. In such cases, traditional statistical methods are insufficient, requiring a more detailed analysis of the dataset’s structure.

Our work presents a tool that automatically performs statistical analysis of benchmark results, accounting for dataset-specific features. It detects multimodality, identifies the number and boundaries of dominant modes, and determines key distribution parameters—providing deeper insights into PostgreSQL performance variations.