PGConf.Russia 2022

My talk covers PostgreSQL infrastructure in Ozon: - how do we tune virtual machines (KVM); - which version control system we have created; - which drivers we make and how it relates to fault tolerance and load balancing; - how do we make "hot" upgrades for the parameters of our virtual machines.

PostgreSQL has a number of peculiarities that you need to take into account not only while maintaining your database but also when designing your database schema. Experienced PostgreSQL are well aware of vacuuming process. On the web one can find tons of materials covering its internals, configurations and monitoring. Many valuable talks about vacuum were given at numerous conferences. However, we still face the common wraparound problem when the maximum possible number of transactions (xid) is reached. It happens even on databases that are relatively small in size. In my presentation, I will share a customer case that looks interesting to me. A chain of mistakes made at different stages of the database's life cycle once caused a disaster. The database fully stopped for one week, we detected a wraparound and spotted corrupted blocks. Maintenance was problematical, and we spent sleepless nights in search of a solution. We managed to achieve a local win as we finally restored the database, but it's not the end of this story, which makes it even more interesting.

The more tables are involved with the query, the more difficult it is for the scheduler to choose a suitable execution plan (both time and memory usage increase). How can we "tell" the planner that it is better to connect this pair of tables first, and the rest can be connected later? What if we see that a part of our query can be improved, but the optimizer does not do this. In my presentation, I will talk about managing the order of connections. I will explain how we can influence the formation of a query plan using the standard "vanilla" PostgreSQL methods.

As opposed to in-memory DBMSs, traditional disc database management systems have an inherent physical limitation. It's their storage system which is by several orders of magnitude slower than CPU operation memory, even in the case of All-Flash NVM. It becomes especially critical for the development of business analytics apps and OLAP scenarios on disc DBMS. We will consider a number how-tos for improving the performance of traditional DBMSs, and then offer a new storage-centric approach to hardware-based speed-up for disc DBMSs.