{"id":96515,"date":"2020-10-12T19:43:02","date_gmt":"2020-10-12T17:43:02","guid":{"rendered":"https:\/\/prohoster.info\/blog\/administrirovanie\/odyssey-roadmap-chto-eshhyo-my-hotim-ot-pulera-soedinenij-andrej-borodin-2019"},"modified":"2020-10-12T19:43:02","modified_gmt":"2020-10-12T17:43:02","slug":"odyssey-roadmap-chto-eshhyo-my-hotim-ot-pulera-soedinenij-andrej-borodin-2019","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/administrirovanie\/odyssey-roadmap-chto-eshhyo-my-hotim-ot-pulera-soedinenij-andrej-borodin-2019","title":{"rendered":"\ud83e\udd47Windows 10 + Linux. Setting up the KDE Plasma GUI for Ubuntu 20.04 in WSL2. Step-by-step guide | ProHoster","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

In the report, Andrey Borodin will discuss how they considered the experience of scaling PgBouncer when designing the connection pooler. Odyssey<\/a><\/noindex>, how they deployed it in production. Additionally, we will discuss what features we would like to see in new versions of the pooler: it is important for us not only to meet our needs but also to develop the community of users. Odyssey<\/a><\/noindex>.<\/p>\n

<\/a><\/noindex><\/p>\n

Video:<\/p>\n

\n

\r\n \"Play\r\n
<\/div>\r\n <\/div><\/center><\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

Hello everyone! My name is Andrey. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

At Yandex, I work on the development of open-source databases. Today, our topic is about connection poolers. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

If you know how to name a connection pooler in Russian, please let me know. I am eager to find a good technical term that should establish itself in technical literature. <\/p>\n

<\/p>\n

The topic is quite complex because many databases have a built-in connection pooler, and you don't even need to know about it. Some settings, of course, exist everywhere, but it's not the same with Postgres. Meanwhile, there is a talk by Nikolay Samokhvalov at HighLoad++ 2019 about configuring queries in Postgres. I understand that the audience here consists of people who have already perfected their query configurations and are now encountering rarer systemic problems related to networking and resource utilization. At times, this can be quite difficult, as the problems are not obvious.<\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

Yandex has Postgres. Many Yandex services reside in Yandex.Cloud. We handle several petabytes of data, generating no less than a million queries per second in Postgres. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

We provide a fairly standard cluster for all services \u2013 this includes a primary node, two regular replicas (synchronous and asynchronous), backups, and the scaling of read requests on the replica. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

Each cluster node runs Postgres, which besides Postgres and monitoring systems also has a connection pooler installed. The connection pooler is used for fencing and serves its primary purpose. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

What is the primary purpose of a connection pooler? <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

Postgres adopts a process model for database operations. This means that one connection corresponds to one process, one Postgres backend. In this backend, there are many different caches, which are quite expensive to maintain separately for different connections. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

Additionally, in the Postgres code, there is an array called procArray. It contains primary data about network connections. Almost all algorithms that process procArray have linear complexity, traversing the entire array of network connections. This is quite a fast loop, but with a large number of incoming network connections, everything becomes a bit more costly. When things get a little more expensive, you can ultimately pay a very high price for a large number of network connections. <\/p>\n

<\/p>\n

\"\ud83e\udd47Windows<\/p>\n

<\/p>\n

There are three possible approaches:<\/p>\n

<\/p>\n