PGConf.Russia 2021

In this presentation, we'll reveal the process and peculiarities of our Oracle to PostgreSQL migration project. We have moved our ELDOKA corporate ECM platform to PostgreSQL, and during this talk we will explain how we ensured object-based and role-based access to data, which features were missing in the community-developed PostgreSQL, how we organized our work with spatial data and changed approach to storing our file content. We will also share our experience in time and resource savings, as well as data replication between our nodes. Performance test results will also be shared with the audience.

Finding the nearest neighbor can be required for various tasks. For example, when you need to find the closest object to a given point on the map. This task looks trivial to non-programmer (a person can easily cope with it if they have a map). In a software developer's reality, this task doesn't have a common solution available to everyone. To get rid of this headache, programmers often create ad hoc solutions also known as "crutches". These workarounds don't look nice and often ruin the mood of a creative programmer who needs to go to a beer pub to cope with the cognitive dissonance :)

Indeed, while a person has a typical field of view and a map with a certain scale, the programmer has only one given point and a huge number of other points (i.e. billions of stars). This multitude of points gets a lot of incoming requests, including the write requests, not just read ones. You can write a perfect query in SQL, however, the real-world query execution plan will be depressingly long. To find the closest neighbor, you will have to read the entire table, compute all the distances from the given point and return the given number of good enough results. Indexing doesn't help in this case, as you will have to fully scan the search tree and read the entire table in random order. This will take much longer than simple table reading. In reality, tasks, where you need to efficiently find nearest neighbors, aren't limited to spatial search. It can also be used for classification tasks, finding typos, data clustering, and deduplication. All such tasks will benefit from efficient nearest neighbor search in DBMSs that are now a de facto standard for storing the data. What do we mean by "efficient search"? It means that our search is fast, concurrent, scalable, and supports various data types (most likely, non-standard ones). We implemented such KNN search in PostgreSQL 11 years ago. I will cover its implementation, today's state and share some use cases for KNN.

Throughout the release cycle of version 14, I had the opportunity to describe the accepted patches in articles dedicated to each of the five release commitfests. The first two articles on Postgres 15 also have some remarks on PostgreSQL 14. If you have read them, you can skip my presentation :-)). On September 30, 2021, PostgreSQL 14 was officially released. Nothing will change in this major release, so it's the right time for a final overview. However, presenting each and every new feature in a 45-min. talk isn't possible. That's why I will be sharing with you my personal top of version 14 innovations.

The built-in partitioning mechanism in PostgreSQL has been developing for several years. However, there is no magic button that would transform an ordinary table into a partitioned one. I will explain how to implement partitioning in a production system, how to prepare your table and apps for partitioning, and what pitfalls can make a DBA's life harder. We will also take a deep dive into various techniques of transferring data between partitions, also considering their pros, cons and limitation.