PGConf.Russia 2018

Avito is the biggest classified site of Russia, and the third largest classified site in the world (after Craigslist of USA and 58.com of China). In Avito, ads are stored in PostgreSQL databases. At the same time, for many years already the logical replication is actively used. With its help, the following issues are successfully solved: the growth of data volume and growth of number of requests to it, the scaling and distribution of the load, the delivery of data to the DWH and the search subsystems, inter-base and internetwork data synchronization etc. But nothing happens "for free" - at the output we have a complex distributed system. Hardware failures can happen - it is natural - you need to be always ready for it. There is plenty of samples of logical replication configuration and lots of success stories about using it. But with all this documentation there is nothing about samples of the recovery after crashes and data corruptions, moreover there are no ready-made tools for it. Over the years of constantly using PgQ replication, we have gained extensive experience, rethought a lot, implemented our own add-ins and extensions to restore and synchronize data after crashes in distributed data processing systems. In this report, we would like to show how our experience can be shifted to a new logical replication subsystem in 10th version of PostgreSQL. In the current implementation, these are only non-trivial solutions - there is a number of issues for the community, that come down to implementing simple recovery mechanisms - as simple as configuring the replication in 10th version.

Last year we announced Skala-SR / Postgres Pro database machine with a hardware and software support for remote direct memory access (RDMA) as a key feature. The first appliances have been already installed at customer sites, and even with the first version customers got some constructions, that were impossible without RDMA and CPU offload (got with Mellanox networking). However, capabilities of this equipment are much wider, and this talk is dedicated to current works and prospective developments for this topic.

It's not a secret for anyone that statistics can not be transferred with a major upgrade. For small and not heavily loaded databases this is not a problem, you can quickly collect new statistics. But we have databases with a volume of about 5TB and a load of about 100k rps, for which it became a big problem: taking off without statistics, the replicas could not even replay WAL. In my report I'll tell you what tricks we went to upgrade these databases with requirements of 100% read only availability, about what mistakes were made, and about how these errors were painfully corrected. The result of these errors was the extension called "pg_dirty_hands", in which we will collect various hacks, which can be last resort to repair data corruption.

In the last six months I have been working with a massive OLAP environment with 20TB shards, spanning around 400TB of data. Come to listen to how we make it all work, the challenges, and the skills involved. This talk has very little in common with the 10TB and Beyond talk because the data environments are very different.

We will cover analytics performance, data alignment, reasons for building extensions in C, and moving data around between servers in multiple data centers.