PGConf.SPB 2023

There are a lot of different databases. We need some formal criteria to compare databases to each other.
The very first idea is to divide SQL and NoSQL.
NoSQL is a popular class of platforms developed in 2000s. Indeed, the rejection of SQL is not a fresh idea because there were predecessors of the relational database model, such as network and hierarchical models.
The fresh «NoSQL» stream consists of the graph, object, and key-value models.
Time-series, wide column, and «document-oriented» models are just extensions of the key-value model. Their advantage is the possibility to parse either key or value on a database server.
The facilities of SQL are much more extensive than the key-value interface. So, the simplified interface is just a charge for the ability to build a distributed database.
So, the data model is the first axis, and the distribution is the second one.
It’s not trivial to release a distributed relational database. The reason is that distributed transaction is one of the most complex problems in IT, and one SQL operator can involve all the nodes in a single transaction.
There are attractive efforts to create a distributed relational database. You should pay attention to Cockroach or Yugabyte. But these platforms haven’t got widespread.
One day a man invented the in-memory cache. As random access memory got cheaper, in-memory technologies came to databases. Every considered class of platforms contains at least one in-memory member. TimesTen and SolidDB are relational and monolithic; Tarantool, Ignite, etc. are key-value and distributed; VoltDB is relational and distributed.
Now the storage environment becomes the third axis.
You can remember Teradata, Greenplum, MS PDW, and a few more distributed relational platforms. They are very successful commercial software. It’s true, but these platforms are not intended to process transactions.
So the fourth axis is the load type: OLTP vs. OLAP.
I would like to draw a 4-dimension cube on the blackboard, but I can’t :)
There are no clear borders between the described classes. Relational databases get some non-relational facilities, while non-relational platforms implement SQL. Disk-based systems become in-memory features, while in-memory databases learn to store data on disk. Monolithic platforms become distributed versions.
The main idea of this presentation is the following: you have first to define the class of platforms for your solution and then choose a platform inside a class.
Not all the classes are equal. Monolithic platforms are much more robust than distributed ones. Relational model is universal in contrast to NoSQL. On-disk storage is cheaper than in-memory.
That’s why a relational monolithic on-disk platform is almoast always the right choice. So, choose PostgreSQL! This platform really covers more than 90% of problems.

Traditionally, fault tolerance in Postgres is implemented using built-in replication mechanisms and external utilities that monitor the state of running Postgres instances and react accordingly when various failures occur. In this presentation, I will tell you what we like and what we don’t like about this approach, which alternative we see, what we have been able to achieve to date and what we want to get done by the time of release, which is planned for December.

Facing the restrictions on the use of Western software, our bank has developed and put into commercial operation a reliable mechanism for data replication between two DBMSs. This mechanism allowed the bank's teams without any special restrictions to replace the product from Oracle (Oracle Golden Gate was previously used in the bank for a long time). In the presentation, we will talk about the main advantages of our software and the nuances of its use.

One of the requirements for the operation of the Corosync/Pacemaker failover cluster is the fencing of the failed node. In virtual environments, fencing is implemented by disabling the virtual machine through a hypervisor, in a cluster on physical servers - through IPMI/ILO. But what if it is impossible to organize fencing, for example, in the cloud environment? During this presentation, I will list and explain alternative methods of fencing a failed node.