PgConf.Russia 2020

Odd things in query plan analysis - wasted time and "unnecessary" buffers.
Structural hints in a plan. How to help a developer with optimization without writing a single line of code. How to match plan nodes with query text and take advantage of this information.

From the very beginning, PostgreSQL in Avito has been solving very important tasks. All the main architecture components are built around DBMS. For over 10 years, the project has been actively evolving, and the infrastructure and architecture have changed a lot.

The talk will start with an overview of how PostgreSQL infrastructure and architecture have advanced in Avito over time and which challenges have been successfully resolved.

Then we'll discuss PostgreSQL usage scenarios in Avito in 2020: microservice architecture, sharding, hosting multiple databases on a single server instance, DBaaS (Database discovery, access control, failover, backup, archive, resource sharing, etc.), integration, and team evolution.
And finally, I'll share our backlog and wishlist.

Any DBA needs some kind of tool for historical workload analyse. Assume once at morning your monitoring team will report of sudden performance degradation at 2-3 a.m., and now you need to investigate this issue. What activities was most resource consuming within that hour? There are several tools for solving this problem, and I'll talk about one very easy and convenient tool - pg_profile. It need only a postgres database and a cron-like tool to run, and it will generate a workload profile report for your database as you need it. Ths report will be a good start point for further investigation.

Cloud storage has some unique characteristics compared to traditional storage mainly because it is virtualized and controlled by software. One example is that AWS EBS shows higher throughput with larger I/O size up to 256 KiB without hurting latency. Hence, a user can get only about 4 MiB/sec with 1,000 IOPS EBS volume if the I/O request size is 4 KiB, whereas a user can get about 250 MiB/sec if the I/O request size is 256 KiB. This is because EBS consumes one I/O in a given IOPS budget for every I/O request regardless of the I/O size (up to 256 KiB). Unfortunately, PostgreSQL cannot exploit the full potential of cloud storage because PostgreSQL has designed without considering the unique characteristics of cloud storage.

In this talk, I will introduce the AppOS extension that improves the throughput of a write-intensive workload by 10x by transparently making PostgreSQL cloud storage-native. AppOS works like a storage driver that efficiently exploits the characteristics of cloud storage, such as I/O size dependency to storage throughput and latency, atomic write support in cloud block storage, and fast, but non-durable local SSDs. To do this, AppOS comprises a Linux-compatible file I/O stack including virtual file system, page cache, block I/O layer, cloud storage driver. On top of the file I/O stack, syscall module supports registering pre- and post-handler for file I/O-related system calls in order to transparently work without modifying PostgreSQL codes.

I will focus on presenting key use cases and performance results of the AppOS extension after explaining the internals. Specifically, I will show the performance results of OLTP and some batch workloads using standard benchmarking tools like pgbench and sysbench. I will also present performance results and implications on multiple clouds including AWS, GCP, and Azure.