PgConf.Russia 2019

As any ordinary software developers, we just pursued a goal to develop a system robust for high loads, and even succeeded. The system architecture was fine, but the data volume was keeping increased and revealed the painful issues and errors that nobody had expected. We faced very strange queries seemed to be unbelievable. In my short talk I would like to share sad experience of arised-from-nothing high loads in DBMS and solving the challenge.

This talk will represent a new platform of Distributed Control System for Nuclear Power Plant operation. Participants will learn about control system for very complicated automation objects. In a hard real time node more than 150 special subsystems are operating in order to control various technological processes of nuclear power plant (NPP), such as reactor control system for more than 1000 MW power unit with a turbine weighing more than 2000 tons. More than 100K of data gained from sensors are resulting in up to 500K of parameters representing 5 branches of physical processes: neutron kinetics, hydrodynamics, chemistry and radiochemistry, and physics of strength. Deviations may cause the whole system to become a huge DDoS source made of useful diagnostic information which is always much larger than the network and hardware are capable to manage. This may lead to normal operation failure. The talk will reveal the approaches to solve the issue.

You will learn about hardware and software architecture of such systems, about backup and replication, data redundancy and technological diversity. How to manage high loads, what is QoS, and what will happen in case of normal operation system failure, as for example was at Fukushima. But, hey, there should be a talk about coding! So, no SSD and HDDs, only InMemory, data structures from tens of millions of elements, and forget about processor cache as it does not work. Imagine your newest 4-generation Xeon has lost all the advantages and turned into a "pumpkin", so let's roll up your sleeves and examine timings, synchronicity, and try to make the most of your hardware, discovering the weakest link from processor, operating system and a network.

One of the most soft after features of Postgres v10 is logical replication. In this presentation we will cover what Logical Replication is, how it compares to Binary (Streaming Replication), how Logical Replication works, configuring Logical Replication, Logical Replication limitations, gotchas, security and management. We will also discuss potential deployed architectures with Logical and Binary Replication and some neat features of the underlying technology.

At the end of this presentation an attendee with a reasonable understanding of how to manage Postgres will be able to configure Logical replication for use.

As is often seen in OLAP and batch processing workloads, the more complex a query (containing many joins, filters, aggregates), the more there is a possibility of row count estimation errors, which leads to planner choosing an inefficient execution plan.

To address that problem, I developed a tool called pg_plan_advsr as a PostgreSQL extension, which corrects the estimation errors by repeatedly feeding back the information collected during query execution to the planner.

The tool has three features:

1. Automatic plan tuning by repeatedly feeding execution information to planner
2. Preserve all plans generated during plan tuning in a history table
3. Create and store optimizer hints to be able to reproduce plans generated during tuning process

I verified the effectiveness of pg_plan_advsr by enabling it when running the join order benchmark (JOB) against PG 10.4 and observed its execution time shortening to 50% of the original. Therefore, it is useful for user who would like to do plan tuning for OLAP and batch processing.

I will talk about the following things in this presentation:

• Principles behind pg_plan_advsr and its architecture
• Detailed information about the measurements done with JOB
• Possible future enhancements
• Using aqo and pg_plan_advsr together (experimental)