Throughout the year, Russians regularly go on holidays - New Year holidays, May and other shorter weekends. And this is the traditional time for serial marathons, spontaneous purchases and sales on Steam. In the pre-holiday period, retail and logistics companies experience an increased load: people order gifts in online stores, pay for their delivery, buy tickets for trips, communicate. Calendar peaks in demand are also a good stress test for online cinemas, gaming portals, video hosting and streaming music services - all of them work βfor wearβ during the holidays.
We describe how the uninterrupted availability of content is organized using the example of the Okko online cinema, which relies in its work on the power of the Linxdatacenter data center.
Previously, additional equipment for local placement was purchased for seasonal consumption surges, and βwith a marginβ. However, when βTime Hβ came, it often turned out that companies either could not or did not have time to cope with the correct configuration of servers and storage systems on their own. It was simply not possible to solve these problems in the course of the development of emergency situations. After a while, the understanding came: peaks in demand for content and online services are perfectly worked out with the help of third-party resources that can be purchased using the pay-as-you-go model - payment for the volume actually consumed.
Today, almost all companies that anticipate a surge in demand for their resources during the holidays (the so-called burst) order in advance the expansion of communication channel bandwidth. Those companies that place applications and databases on the resources of the data center, increase computing power in the clouds for the holiday peaks, ordering the necessary virtual machines, storage volume, etc. in data centers.
How not to miss the calculations
To prepare for peak loads, coordinated work between the provider and the client is important. The main points in this work can be called an accurate forecast of a load jump in terms of time and volume, careful planning and the quality of interaction with colleagues inside the data center, as well as with a team of IT specialists on the side of the content provider.
A number of solutions help to organize the quick provisioning of resources necessary to ensure that the new episode of your favorite series does not freeze on the screen of your tablet.
- Firstly, these are workload balancers: these are software solutions that carefully monitor the load level of servers, storage systems and networks, allowing you to optimize the operation of each system for the task at hand. Balancers evaluate the level of availability of both hardware and virtual machines, not allowing system performance to sag, on the one hand, and preventing the infrastructure from overheating and slowing down, on the other. Thus, a certain level of resources is maintained, which can be quickly transferred to solve urgent problems (a sharp jump in calls to the portal with video content, an increase in orders for a certain product, etc.).
- Secondly, CDN. This technology allows users to receive content from the portal without buffering delays by accessing it from the geographical point closest to the user. In addition, CDN neutralizes the detrimental effect on traffic transmission processes caused by channel congestion, communication breaks, packet loss at channel junctions, etc.
All-seeing Okko
Consider an example of preparation for the holidays of the Okko online cinema, which uses our sites in Moscow and St. Petersburg.
According to Aleksey Golubev, Technical Director of Okko, in addition to calendar holidays (high season), the company has periods when the main movie releases from majors are released:
βEvery year during the holiday season, Okko is growing in terms of traffic approximately twice as compared to the previous year. So, if last New Year's season the maximum peak load was 80 Gb / s, then in 2018/19 we expected 160 - the traditional doubled increase. However, we got over 200 Gbps!β
Okko always prepares for peak load slowly, throughout the year, as part of a project codenamed "New Year". Previously, Okko used its own infrastructure, the company has its own distributing content delivery cluster, on its own hardware and with its own software. During the year, Okko's technical specialists gradually bought new servers, increased the throughput of their cluster, assuming a doubling of annual growth. Additionally, new uplinks and operators were connected - in addition to large players like Rostelecom, Megafon and MTS, they also connected traffic exchange points and the smallest operators. This approach made it possible to deliver the service to the maximum number of customers using the shortest route.
Last year, after analyzing the cost of equipment, labor costs for expansion, and compared with the cost of using third-party CDNs, Okko realized that it was time to try a hybrid distribution model. The double growth during the New Year holidays is followed by a decline in traffic, and February is the lowest season. And it turns out that your equipment is idle at this time. By the summer, the decline is leveled, and by the autumn season a new upsurge begins. Therefore, in preparation for the new 2019, Okko took a different path: they finalized their software in order to be able to distribute the load not only on themselves, but also on external CDNs (Content Delivery Network). Two such CDNs were connected, into which excess traffic βmergedβ. The internal bandwidth of Okko's IT infrastructure was ready to withstand that same double growth, but partner CDNs were prepared in case resources were overwhelmed.
βThe decision not to increase its CDN saved Okko about 20% of the distribution budget in CAPEX. Plus, the company saved several man-months by shifting the work of setting up the equipment to the shoulders of a partner.β Alexey Golubev comments.
The distribution cluster (internal CDN) in Okko is implemented at two Linxdatacenter sites in Moscow and St. Petersburg. Complete mirroring of both content and its caching (distributing nodes) is provided. Accordingly, the Moscow data center processes Moscow and several regions of Russia, and the St. Petersburg data center handles the North-West and the rest of the country. Balancing occurs not only on a regional basis, but also depending on the workload of the nodes in a particular data center, the presence of a movie in the cache and a number of other factors are also taken into account.
The enlarged service architecture looks like this in the diagram:
Physically, the service and product development support consists of about ten racks in St. Petersburg and several racks in Moscow. There are a couple of dozen servers for virtualization and almost two hundred "iron" servers for everything else - distribution, service support and the office's own infrastructure. The interaction of the content provider with the data center during the period of peak loads does not differ in any way from the current work. All communication is limited to an application to the support service, and in case of an emergency - by call.
Today, more than ever, we are closer to the present, 100% uninterrupted scenario of content consumption online, since all the technologies necessary for this are already available. The development of online streaming is happening very quickly. The popularity of legal models of content consumption is growing: Russian users are gradually getting used to the fact that you have to pay for content. Moreover, not only for cinema, but also for music, books, educational materials on the net. And in this regard, the delivery of the most diverse content and with the least network delays is the most important criterion in the operation of online services. And our task, as a service provider, is to close the need for resources on time and with a margin.
Source: habr.com