Application architecture of large live broadcast platform

It's been a while since we entered the live broadcast field .

Like most of you , New to the field of live broadcasting, people are very curious about the macro structure of this field .

Here I study according to my time , To sum up, share it with interested partners to sweep literacy .

At present, there are many live broadcast platforms , Shaking, / fast 、B standing 、 Bucket / Tiger, etc .

The efficient interaction mode of live broadcasting is the ability that all industries need to have , With the development of network technology, it is also an inevitable trend .

Put aside the different content of each live broadcast , A complete live broadcast platform needs to include at least the following core functions .

Watch the live broadcast 、 Send the barrage 、 Reward and give gifts 、 Anchor interaction .

We take an ordinary C End user perspective , Take a look at the general application architecture behind these core functions .

Watch the live broadcast

When we enter the live broadcast room, we first watch the live broadcast content .

【 Pull flow 】

The live broadcast content is through Streaming media player Play out , And streaming media player through Stream Address To pull Stream data , Reprocess related decode Work , Finally, the picture and sound are presented .

In order to ensure that users all over the country and even overseas can watch the live broadcast smoothly , You need strong physical network coverage .

Streaming media data passes upstream CDN Origin push （ Or push-pull combination ） To all edge nodes .

Users in different regions through overall situation CDN-DNS After scheduling , Visit nearby CDN node .

In order to cover all areas as much as possible , It is impossible for every company to build a huge network infrastructure .

At this time, we need to rely on a powerful public cloud platform （BAT、 Huawei 、 jinshan 、 Bytes, etc ）.

Public cloud manufacturers spent a huge amount of money on building physical links , As an application-oriented enterprise, you only need to use cloud product capabilities , Everything is service .

When we watched the live broadcast happily , You will be curious about the source of these live content , How are these pictures collected , How to transmit data .

【 Push flow 】

Usually there is a broadcasting tool for broadcasting , The launch tool is specially used to collect streaming media （ Audio and video equipment ） Information and media stream synthesis , It can also be used. OBS And other powerful streaming tools .

When the broadcasting tool collects media information, it will conduct local compression , And then through RTMP（Real Time Message Protocol） The protocol pushes data to RTMP The server .

RTMP After the server receives the streaming media data , The stream information will be decoded and HLS（HTTP Live Streaming） File slice .

take RTMP The server is deployed on the public virtual machine , It is to use the internal services of the public cloud （LAAS、PAAS） The affinity of , It can maximize the convenient coverage to CDN Edge node .

You may be curious , Why is streaming used RTMP Agreement and pull flow HLS agreement .

The broadcast end uses RTMP, Because RTMP be based on TCP Real time protocol , It can ensure the reliability and real-time of streaming .

See end use HLS Because HLS It is based on static small files , This is more convenient in large-scale millions PCU（Peak concurrent users ） when , Use static CDN Easy to distribute advantage to cover .

Use HLS At the same time, it is also convenient for dynamic switching between different code rates , So as to select the best viewing experience according to the user's network situation .

such as , Users are under weak network , We'll switch to low bit rate to watch .

When the user network is good , We'll switch to high bit rate to watch , At the same time, you can also use some advanced viewing experience and interaction .

Of course ,HLS There are many advantages , In compression 、 Encryption and decryption are very mature .

stay 2C Large scale live broadcast ,flv It's the past tense ,HLS Is the future .

【P2P】

In order to save as much as possible CDN The huge cost , Will use P2P（Peer to Peer） Technology to reduce public cloud bandwidth usage .

Again , In order to make STUN The server 、 Signaling server Closer to the user , Both servers are deployed on public cloud edge hosts .

Further crushing costs , You can also use PCDN（P2PCDN） service .

Due to the strong terminal network coverage of the public cloud （ Router 、 Terminal equipment, etc ） Can be CDN Distribution content is further marginalized .

Use tens of thousands of edge computing capabilities to extend content and computing to the end of users .

Actually P2P The technology is already very mature and has a better history than WEB The Internet has a long history .

Send the barrage

When watching the live broadcast, a very important appeal is to interact with the anchor , And the barrage is the interactive way with the largest amount of information .

bullet chat It is now an indispensable media element in the video field .

The barrage was originally in the field of on-demand , Japan's Niconico It can be regarded as the ancestor of barrage interaction .

A standing （Acfun） It is basically the cultural birthplace of domestic bullet screen video websites .

But at present, those who play the bullet screen into a new dimension in China are B standing （bilibili） 了 .

Barrage from send to receive , rely on A long connection middleware .

Considering the platform availability , Long connection services as a whole need to support disaster recovery , The whole architecture needs to support the mixed deployment of multiple machine rooms .

At the bullet screen message delivery end, it is necessary to do computer room line detection , Select the best computer room according to the relevant data after exploration .

There is another advantage of using public cloud , You can enjoy a certain amount “ Peer to peer connection ” Value added service .

The overall architecture is increasingly moving towards hybrid clouds , It is to balance the core data closed loop of private cloud as much as possible , At the same time, with the powerful network coverage of the public cloud .

Public cloud high-performance hosts can generally carry millions of connections at low frequencies , The bottleneck basically lies in memory and bandwidth .

The cost of the machine is actually good , Because machines can theoretically expand infinitely , But the bandwidth has a physical upper limit , So it's more expensive .

【 Public opinion control 】

Any Internet company with a certain volume , According to the industry and content , We need to control the public opinion of users' participation behavior .

The barrage area is a major position that needs to be controlled , It is also one of the core battlefields in charge of each security center .

The characteristics of the barrage are precisely difficult to control , Usually the idea is based on AI Algorithm （ Such as NLP） Filter the content of the barrage with pre sensitive words 、 Emotional analysis, negative filtering, etc .

Actually ,AI What the algorithm can deal with is relatively simple and easy to recognize semantics .

Chinese characters and culture are broad and profound , Plus complicated words 、 Words stem together , Even the most powerful algorithm cannot absolutely recognize .

Manual auxiliary approval is required , Labor also needs a certain “ Culture ” Profound and quite extensive “ width of knowledge ”.

So it must be a very large audit team .

General application enterprises , The technical content of making barrage function is not high , And now cloud vendors 、 Open source sdk, With a little combination, the architecture can basically be built .

Instead, it's , Security 、 Public opinion control is the most critical and key investment .

Now, more and more companies are doing third-party public opinion discovery and control , There are still many social public opinion control companies , Subdivided vertical classes are relatively few .

Give gifts and rewards

After watching a hard-working performance by the anchor , You need to express your sincerity by rewarding .

A seemingly simple gift , In fact, there are many systems involved , These systems are located at different levels of responsibility .

But whether in the field of live broadcasting or e-commerce 、 education , It's basically similar , Just selling different goods .

On the gift chain , Order 、 Account 、 Clearing and settlement is actually a relatively mature transaction 、 Financial model .

These two pieces are basically set according to the pattern , The purpose is to complete the company's revenue recognition and capital settlement .

【 The whole link 】

We divide the whole link into two layers , The first level is the sales business level , This floor consists of a series of sales units , Like the mall 、 Promotion Center, etc .

In the field of live broadcasting , The gift giving system is located on this floor , Gift giving can be considered as a virtual commodity transaction .

Gift giving scenes can be very diverse , But they are all in disguise to promote the conclusion of the transaction .

Once the goods are sold successfully , According to the nature of business , The amount and time point of revenue recognition are also different .

In live gift giving , The gift given is basically to confirm the income in near real time , At this point, you need to trade orders （ Or business order ） To carry this part of the transaction responsibility .

At the lower level of the sales business layer is the transaction settlement layer , It is the division of income 、 Settlement 、 Payment part .

The core of this part is the transaction supporting the business sale , Settle the income of each role . It is the most complex part of the transaction settlement system , It is also prone to asset losses .

No matter what we sell , To run completely in the trading system , First of all, we need to commercialize the concept of virtual sales .

After having goods, you can sell them on the shelves through the sales layer .

The business of gift giving is relatively simple , Deduct the virtual assets of personal account , Add it to the anchor account , It completes the basic business process .

When this is done , Then you need to separate the transaction order . Separate computing platforms are required 、 External merchants / How much do participants in various roles such as institutions get .

After a successful gift giving, at least a platform is needed 、 The host （ Or labor union ） Make a certain share .

This part is the responsibility of the clearing part of the clearing and settlement system , This part is charged 、 Allocate and calculate the settlement details of each role separately .

The settlement system will retrieve the settlement cycle of the contract on a regular basis （T+N/ Monthly settlement, etc ） Settle accounts and make payments to merchants , Most of this part has an accounting period .

in short , Gifts and rewards are biased towards the field of transaction settlement , The core of the whole system is accounting 、 transaction 、 Financial and other business knowledge .

At the same time, in the system design , Data consistency 、 Reconciliation process and scenario are the core of the whole system architecture design .

Anchor interaction

In the studio , Whether for the anchor or the platform , Are the ultimate goal . But this ultimate goal cannot be achieved in one step , It needs to be constantly transformed through other scenes .

Most of the needs of live broadcast users are hormone needs , How to use various interactive tools 、 Form of interaction , Let users and anchors interact with each other more , To promote gift giving .

From the perspective of users, the forms of interaction are mainly divided into three categories , bullet chat 、 gesture 、 Gift giving .

Such as shaking / Double click the screen to like , Long press of tiger teeth to recall the shortcut panel ,B Station sending special barrage trigger “ Thermal storm ” Special effects, etc .

At the same time, almost every family has special gift interaction , Change the theme of the room by giving special gifts 、 The change of the anchor's dress .

All forms of interaction need unified abstract management , It can be summed up in one sentence ： “ In the specified room for the specified period of time , Enable one or more interactive activities / How to play ”.

As for a specific interactive play , What materials are needed for this game , What trigger media are needed , In addition to the channel part, you can subscribe , Others need customized development .

such as , Energy storage in the interaction between bullets and curtains , In addition to the general barrage channel , Which words , How many times in how many time windows , Then trigger what business logic .

Or combo gift storage , In addition to the gift giving channel is a general part , What gifts , How many time windows to send , How much money is satisfied is where the business logic needs to capture and process .

The whole interaction needs to get through the viewing end and the broadcasting end .

There is a clear difference between watching end effects and broadcasting end effects , Most of the viewing terminals are special effects elements that can interact in the live studio , And the special effects on the broadcasting end are finally reflected in the stream .

such as , An interactive special effect for gift giving , user “ continuity ” When the gift reaches a certain threshold, it triggers the interactive logic of the broadcaster , Finally, input the message into the stream , At the same time, see whether the requirements are added SEI（Supplemental Enhancement Information） Information .

Room theme

If we compare the live studio to a real room , We hope it's a thousand people and a thousand faces , And it has certain customization .

meanwhile , We hope the room can automatically sense the holiday according to the calendar for entertainment 、 Seriousness .

If we compare the live studio to an online community unit , This place is not beyond the law , Need to be affected by different degrees of public opinion 、 Social sensitivity 、 Legal safety supervision .

The live broadcasting room is divided according to the business form , There are many dimensions to divide , For example, commercial rooms （ Take the goods / Online retailers / distribution ）、 Game room 、K Song room 、 Projection hall, etc .

Final , Different rooms need to be opened 、 Control some capabilities , Interactive entertainment , Gift giving , Anchor behavior, etc .

Calendar and theme management are the basic genes of the whole room generation .

It is necessary to carry out engineering design for the whole structure of the room , Be interface oriented 、 Open standards oriented .

Only in this way can it be customized and extensible .

The whole room probably has several core areas , ___ Gift area 、 Player area 、 The barrage area 、 Interactive area .

The implementation of these areas needs to be divided into two layers , The first layer realizes the core basic functions of the kernel .

This part of the function can directly exclude topic relevance .

The second layer is other decorative functions , The implementation of this part requires theme awareness .

On top of that , We assume to realize a special barrage effect in the barrage area .

The implementation plug-in passes the engineering interface and standard , Get the calendar & The theme can perceive the attributes of the festival . In terms of implementation, basic functions and effect functions can be separated .

This design is also convenient for automatic testing and business inspection of effect functions on special festivals throughout the year .

The room recommendation algorithm needs to add a calendar recommendation factor , This part is easy to understand . What is more interesting is that the center of the room structure is the outlet of the finished room .

Activity guarantee

Finally, share the content of technical support .

This part is a concise summary , Do not describe the process .

in short , Guarantee mainly focuses on two key points .

The first point is that most of our familiar application systems are generally resistant to high concurrency , This includes a series of middleware 、DB、 cache 、 Microservice suite, etc .

The second point is to follow PCU Related linear resources CDN bandwidth 、P2P bandwidth 、 Long connection bandwidth 、 Dot 、 Real time report calculation .

According to the view 、 Look at the two ends of the broadcast separately .

Broadcast end , We need to focus on ensuring the availability of streaming , So basically, there are at least two streams on the push side backup.

Look at the end ,QPS、TPS The relevant interface is basically related to slope or sudden increase of flow .

（ notes ： The peak value of an interface can be carried 10wQPS, Reaching this peak is minutes per second plus 1w Slowly increase , still 1s arrive 10w It's different .）

The essence of the slope is to test the capacity expansion speed of related resources at the back end .

In a large concurrency scenario , Sometimes we need to expand capacity in advance to prevent sudden increase of traffic .

This is also the difficulty of evaluating expansion resources according to the slope .

On the other hand, it is with PCU Linearly related .

PCU The direct result is related public cloud bandwidth 、 The proportion of connection pool increases , This part needs to do a good job of resource pre purchase .

When the number of people online is very high , Data reporting 、 Real time data processing also increases linearly .

This part of storage resources needs to be prepared in advance , Computing resources can be flexibly scheduled .

All the above scenarios can be carried by the system within a reasonable capacity , But once unexpected traffic comes in, we also need to have a self-protection mechanism and switching capacity .

This part is related , Current limiting 、 Degradation capability is also essential .