Dear habread readers, we are ready to provide you with video recordings of all reports from the HPC High Load Conference, which took place on August 9 in Moscow.
Right today (or at any time convenient for you) in the program: Sergey Averin XEK , Badoo (participated in the creation of such projects as Habr, dirty, leprosorium, autokadabra, dribbler, trendclub, etc.); Yuri Nasretdinov, Badoo; Andrew Sumin AndrewSumin , Mail.ru Group (previously Yandex, HeadHunter); Konstantin Osipov, Mail.Ru Group (formerly Oracle, Sun, MySQL, Ringrows, Spylog, Interpro); Daniil Pavlyuchkov, ITmozg.ru. ')
Badoo’s desktop application project manager, Sergey Averin, made a presentation: “Not all databases are equally useful.” Some theses:
stay on MySQL, do not give in to fashion on NoSQL;
iron is cheaper than people. At the start, scale with hardware and standard open source software (and its settings);
an ordinary startup needs at least six months to get out of the resources of one server;
think about the consistency in advance, at the level of architecture, and then later cut it yourself;
Web services that are not billing or processing do not require high connectivity and synchronicity;
search for MySQL FULLTEXT - doomed. Use Sphinx or other specialized engines;
The main thing in reliability is iron and redundancy on it. Badoo uses Percona, and NoSQL is very popular.
The head of the Mail.Ru Group developers, Konstantin Osipov, spoke on the topic “The diversity of NoSQL solutions and Tarantool's own development”.
Tarantool is as fast as Redis or Memcache, but takes up less memory. Holds 400 thousand simple requests per second;
Tarantool stores 60 million sessions for the entire Mail.ru on two physical machines;
Look at the VoltDB RDBMS, which solves a number of horizontal scaling problems.
The head of ITmozg development, Daniil Pavlyuchkov, made a presentation: “Web site profiling, server side”.
Some theses:
profiling should start with rough testing at low loads, then with no less rough benchmarks;
then load tests, which are based on real data from peak loads. Then a stress test. stress test is needed in order to check why the site will fall in the first place;
use XHGUI to visualize XHProf data. And the strace utility for tracing system calls;
for PHP, use XHProf from Facebook because it can be allowed to be produced (adds only 50 ms) + visualization.
And that's not all good news for today. On October 6, a professional development festival in IT - BitByte was held, during which expert presentations were held in the HPC, Upgrade, and Soft Skills sections. We will also start gradually uploading the processed videos on the site, and then we will add all the video records to Habr. Keep for updates. Here you can read about how BitByte went and what was said there.