PgConf.Russia 2016

Postgresql's buffer manager has parts where it's showing its age. We'll discuss how it currently works, what problems there are, and what attempts are in progress to rectify its weaknesses.

• Lookups in the buffer cache are expensive
• The buffer mapping table is organized as a hash table, which makes efficient implementations of prefetching, write coalescing, dropping of cache contents hard
• Relation extension scales badly
• Cache replacement is inefficient
• Cache replacement replaces the wrong buffers

Whenever multiple users, processes, or threads are concurrently modifying data which is shared among them, problems can occur if race conditions are not handled somehow. These problems are particularly acute in a database which provide ACID semantics. A set of changes grouped into a database transaction must appear atomically, both to concurrent transactions and in terms of crash recovery. Each transaction must move the database from one consistent state (with regard to business rules) to another. For programming efficiency, each transaction must be able to be coded independently of what other transactions may happen to be running at the same time. In the event of a crash, all modifications made by transactions for which the application was notified of successful completion, and all modifications which had become visible to other transactions, must still be completed upon crash recovery. Over the years, various strategies have been employed to provide these guarantees, and sometimes the guarantees have been compromised in one way or another. This talk will cover the approaches taken to provide these guarantees or compromised variations of them, with an emphasis on the Serializable Snapshot Isolation (SSI) technique available in PostgreSQL (and so far not in any other production product). While SSI already performs faster and with higher concurrency than any other technique for managing race conditions with most common workloads, there are many opportunities for further enhancing performance, some of which would require the assistance of people expert in the various index access methods; these issues will be discussed. The talk will also present some rough ideas about how SSI techniques might be used with XTM in a distributed system.

Time will be reserved at the end of the talk for group discussion of optimizations and possible application in distributed environments.

Heroku Postgres is a cloud database service and the largest provider of PostgreSQL as a service anywhere. We operate more than 1,000,000 PostgreSQL databases with a team of about 10 people. We may be the most efficient DBAs in history, with approximately 100,000 databases per person on our team! This talk will introduce the opportunity and challenges of building and operating a cloud database service, as well as discussing the strategies we use to build, operate, and scale this product and team for the last six years now. We will include details about * a brief introduction to the service to provide context * strategies to design and build such a data service * operational war stories like how to recover from losing thousands of servers at once, * common challenges users have with Postgres * and a basic overview of the technical architecture

This is a complementary talk to Will Leinweber's talk, which will go into much more depth on the architecture of the software we have written.

In the speech we consider the current PostgreSQL planner model, then the possibilities of applying machine learning methods for planner improvement and the obtained results.