PGConf.Russia 2022

Disaster-tolerant cluster implies minimal data loss in the event of the main data center disaster and switching to the backup data center. The data loss takes place due to the asynchronous replication between the main data center and the backup one. However, there's a solution that can resolve the existing problem and ensure zero data loss in the case of the main data center disaster. My talk covers this solution for zero data loss.

Logical decoding and replication slots introduced in PostgreSQL 9.4 (released in 2014) created a solid foundation for implementing built-in core logical replication in version 10 (released in 2017). Unfortunately, there are a few limitations that make logical replication not very useful in real-world scenarios. Logical decoding currently isn’t supported on the standby server, and PostgreSQL allows creating logical replication slots only on the primary server. Or in other words, logical slots are lost on failover/switchover.

Postgres hackers made many attempts to address the problem, and most of them resulted in not too much success. Although, there is one little function introduced in PostgreSQL 11 that made it possible to implement failover of logical replication slots externally.

In my talk I will tell a story of how Patroni solves the problem of logical replication slots failover without using invasive third-party extensions, dig down into some of the Postgres internals in order to prove why this approach is safe, and finally, we will discuss limitations and potential downsides of this solution.

We'll talk about the ways to change the database records without updates. To do this, we'll explore what's going on at the lowest level and see what can happen if we make changes in data files in different operation modes.

Previously, we have explained the internals of Avito, discussed, where and how we store your classifieds, and how they appear on the search results. In the recent 3-4 a lot has changed in Avito. We got rid of logical replication, stopped using standby servers for reads, removed nearly all stored procedures and our custom failover solution, migrated all our classifieds into a sharded DBMS, switched from the monolith app to microservices. I'll explain why we made such decisions, list some of the problems we encountered and describe the current state of our development process.