Building an edge computing platform

This is the last post in the series, now we’re going to design an edge computing platform using Lua, Nginx, and Redis cluster. Previously, we explained how to add code, not dynamically, in nginx.

Platformize the computations

The platform is meant to provide a way to attach Lua code dynamically into the edge servers. The starting point can be to take the authentication code and port it to this solution.

At the bare minimum, we need a Lua phase name, an identifier and the Lua code. Let’s call this abstraction the computing unit (CU).

If we plan to add a computing unit dynamically to the NOTT, we need to persist it somewhere. A fast data store for this could be Redis.

overview_architecture

We also need to find a way to encode the computing unit into some of the Redis data types. What we can do is to use the string to store the computing unit. The key will be the identity and the value will hold the Lua phase and code separated by the string “||.

The platform needs to know all of the computing units, therefore, we need to list them. We could use the keys command but it can be very slow depending on how much data we have.

A somewhat better solution would be to store all the identities in a set data type, providing an O(N) solution, where N is the number of CUs.

KEYS pattern would also be O(N) however with N being the total number of keys in the data store.

Now that we know how we’re going to encode the computing unit, we need to find a method to parse this string separated by || and also a process to evaluate this string as code in Lua.

To find a proper and safe delimiter is difficult, so we need to make sure that no Lua code or string will ever contain ||.

To split a string we’ll use a function taken from Stackoverflow PepeLaugh - Discord Emoji and for the code evaluation, Lua offers the loadstring function.

But now some new questions arise, what happens if the code is syntactically invalid or when the CU raises an error? and how can we deal with these issues?

To deal with syntax errors, we need to validate the returned values from the function loadstring where the first value is the function instance and the last is the error.

And to protect against runtime error, Lua has a builtin function called xpcall (pcall meaning protected call) that receives a function to execute and a second argument which is an error handler function.

With all this in mind, we can develop the core of our platform. Somehow we need to get all the computing units from Redis, parse them to something easier to consume and finally, we can execute the given function.

Before we start to code, we can create a prototype that will replicate the authorization token system we did before but now using Redis to add and fetch the computing unit as well as shielding the platform from broken code.

To test these lines of code, we can go to the terminal and simulate calls to the proper nginx location. Therefore we can understand if the expected behavior is shown.

Screen Shot 2020-04-14 at 12.59.56 PM

Since we’re comfortable with our experiment, we can start to brainstorm thoughts about the code design and performance trade-offs.

Querying Computer Units

The first decision we can take is about when we’re going to fetch all the computing units (CUs). For the sake of simplicity, we can gather all the CUs for every request but then we’re going to pay extra latency for every client’s request.

Screen Shot 2020-04-14 at 11.36.37 PM
the step 2 can generate several round trips in Redis cluster

To overcome these challenges, we’ll rely on two known techniques,  caching and  background processing.

We’ll move the fetch and parse logic to run in the background. With that running periodically, we then store the CUs into a shared memory where the edge computing core can lookup without the need for additional network connections.

Openresty has a function called ngx.timer.every(delay, callback), it runs a callback function every delay seconds in a “light thread” completely detached from the original request. This background job will do the fetch/parser instead of doing so for every request.

Once we got the CUs, we need to find a buffer that our fetcher background function will store them for later execution, openresty offers at least two ways to share data:

  • a declared shared memory (lua_shared_dict) with all the Nginx workers
  • encapsulate the shared data into a Lua module and use the require function to import the module

The nature of the first option requires software locking. To make it scalable, we need to try to avoid this lock contention.

The Lua module sharing model also requires some care:

“to share changeable data among all the concurrent requests of each Nginx worker, there is must be no nonblocking I/O operations (including ngx.sleep) in the middle of the calculations. As long as you do not give the control back to the Nginx event loop and ngx_lua’s light thread scheduler (even implicitly), there can never be any race conditions in between. “

source

Edge Computing Bootstrapping

The usage of this edge computing lua library requires you to start the background process and also to explicitly call an execution function for each location and lua phase you want to add it to.

In the past example, we started, on the first request, at the rewrite phase, this will also initiate the background job to update every X seconds.

On the access phase, we’re going to execute the available CUs. If a second request comes in, it’ll “skip” the start and just execute all the cached CUs.

no_extra_cost

While using these APIs solve our challenge of adding unnecessary latency, it comes at a price: now when we add a new CU to the data store, it might take X seconds to be available for workers.

The rewrite_by_lua_block was used because it’s the first phase where we can start a background job that can access cosockets and works with lua-resty-lock (a library used by resty-redis-cluster).

funny behavior will happen, related to the eventual consistency nature of this solution: if a client issues a request1, in a given time for a Worker1, later the same user does another request2 and a different Worker2 will accept it. The time in which each worker will run the update function will be different.

This means that the effective deployment of your CUs will be different even within a single server. The practical consequence for this is that the server might answer something different for a worker in comparison to another one. It will eventually be consistent given the x seconds delay declared as update interval.

worker_memory_update

Nginx worker load balancing relies on the event-based model and OS-dependent mechanisms to efficiently distribute requests among worker processes.

How to use it

Adding the CU via the Redis cluster will make it work.

Computing Edge Use Cases

Let’s list some of the possible usages for this platform so we can think a little bit ahead of time.

  • access control – tokens, blacklist, origin
  • change response
  • decorate headers
  • generate content
  • traffic redirect
  • advanced caching

The options are endless, but let’s try to summarize the features that we didn’t add to the code yet.

When we implemented the request counter we used redis as our data store so it’s safe to assume that somehow the CUs might use redis to persist data. Another thing we could do is to offer sampling, instead of executing the task for all requests we could run it for 3% of the them.

Another feature we could do is to allow filtering by the host. In this case, we want a given CU to execute in a specific set of machines only, but we also can achieve that in the CU itself if we need to.

The persistence needs to be passed for the CU, we can achieve that by wrapping the provided raw string code with a function that receives an input and pass this argument through the pcall call.

We’ll have access to the edge_computing in our CUs as if it was a global variable.

And finally, we can mimic the sampling technique by using a random function. Let’s say we want a given CU to be executed 50% of the time.

We need to encode the desired state at the datastore level and before we run the CU, we check if the random number, ranging from 1 to 100, is smaller or equal to the desired sampling rate.

Pure random distribution is not the most adequate, maybe in the future, we can use an algorithm similar to the power of two choices.

Future

Any platform is a complex piece of software that needs to be flexible, accommodate many kinds of usage, be easy to use, be safe and it still needs to be open for future innovation/changes such as:

  • a control plane where you can create CUs using UI
  • if the execution order is important then change the current data types
  • add usage metrics per CU
  • wrapper the execution with a timeout/circuit breaker mechanism

I hope you enjoyed these blog posts, they were meant to mostly show some trade-offs and also to deepen the nginx and Lua knowledge.

 

Empowering Nginx with Lua code

This is the second post in the series where we develop an edge computing platform. In this post, we’ll add some code/behavior to the front end servers. Here’s a link to the previous entry.

Add code inside the front end

The OTT service we did before don’t employ any kind of authentication thus the users can watch the streams for free. To solve this authentication issue we can add Lua code embed into nginx.

OpenResty – an Nginx with support for LuaJIT 2.0/2.1 code.

To run Lua code inside nginx you need to understand a little bit of the request phases within the server. The request will travel across different stages where you can intercept it using Nginx directives and add the code.

Screen Shot 2020-04-20 at 2.09.49 PM

Just for the sake of learning, the authentication logic will be a straightforward token system. During the access phase, we’ll deny access for those with no proper authentication. Once a user has the required token it’s going to be persisted in form of a cookie.

Fixed token with no expiration is unsafe for production usage, you should look for something like JWT.

The edge server can run useful behavior/code, now let’s laid out some examples that demonstrate the power we can have while executing functions at the front end.

Suppose a hacker, behing the IP 192.168.0.253, is exploting a known issue, that is going to be fixed soon. We can solve that by blacklisting his/her IP. Adding lua code, to the same phase, can fix this problem.

You can access all the nginx variables using the api ngx.var.VARIABLE.

Nginx has the deny directive to solve this problem although it doesn’t allow a dynamic way to update the IP list. We would need to reload the server every time we want to update the IPs.

It’s wanted to avoid different domains to consume our streams, to prevent that, we’re going to examine the referer header and reject all the requests not originated from our domain.

CORS and CSP will be safer to solve this issue.

To on-the-fly change the response from the backend, we’ll add a custom HLS tag in the playlist.

To decorate the HTTP headers, we’ll attach new ones exposing some metrics from the server and for that matter, it can rely on the ngx.header[‘name’] API.

Finally, we’ll count how many requests a given user (based on her/his IP) does and expose it through a custom HTTP header. The counter was stored in Redis.

All this is working, if you want, you can test it by yourself.

Conclusion

Did you notice a pattern? Every time we want add a new feature, we need to:

  • write a little bit of Lua code
  • attach it to a request phase directive in nginx
  • deploy and reload the edge servers

That’s why we need to build an edge computing platform, to put code faster into production and avoid server reload.

Building an open-source OTT platform

Create software from “scratch” might not be a good idea at first but it’s often a great way to study a specific technology or even to deepen your knowledge in a particular field of computer science.

In this three-post series, we’re going to build a simple video platform using open-source software, will add features to it so it handles the computional power on the front end (edge computing) and we’ll conclude designing a platform that will enable us to add code/features dynamically to the servers.

Screen Shot 2020-04-21 at 9.07.02 AM

 

An over-the-top (OTT) is a streaming media service offered directly to viewers via the Internet. OTT bypasses cable, broadcast, and satellite television platforms. Now you don’t need to spend your money that much.

Edge computing is the ability to put computation and storage closer to the place where it is demanded, in simpler terms is the code running within your front end servers.

We’re going to design two distinct services: a simple video streaming solution and an edge computing platform for this video streaming service.

  1. Building NOTT – an open source OTT video platform
  2. Add edge computation to NOTT – empower nginx with lua code
    • token authentication code
    • IP blacklist
    • forbid other sites
    • add a custom HLS tag on the fly
    • expose metrics in HTTP headers
    • count user request per IP
  3. Platformize the edge computing – using lua + redis
  4. $profit$

The NOTT

The new OTT is a VERY SIMPLE open-source video platform that expects an input signal and produces an output stream. It was made mostly as an excuse to discuss and design an edge computing platform around it.

NOTT
NOTT is built using a simple html5 app
Screen Shot 2020-04-12 at 8.11.15 AM
NOTT architecture

The UI app is a simple static html5 file served from nginx. We’re using Clappr (backed by hls.js and shaka) as the selected player. The front end works as a caching layer for the video streaming and it also hosts the NOTT app.

The live streaming reaches the platform through FFmpeg, the broacasting, which is also used to transcode the input producing multiple renditions. The nginx-rtmp acts as a packager, converting the RTMP input into the adaptive output streaming format known as HLS.

The main selling point of our OTT platform is that it has the popular TV channel color bar (60fps) and the legendary TV show big buck bunny (partner’s licensed content). :slightly_smiling_face:

Compatibility: I didn’t test on all platforms (browsers, ios, android, CTVs), video is hard and NOTT won’t cover 100% of the devices but it should work in most places.

How does it work?

To broadcast the color bar TV show into the platform, we’ll use FFmpeg. It has some filters that are capable to create synthetic color bar frames at a given rate. It also offers an audio source filter known as sine can be used to create artificial sound.

This command creates color bar pictures at 60 frames per second and a sine wave sound at 48000 hertz. It encodes them to the video codec h264 using the libx264 and to the audio codec aac. Finally, we send them to the transcoder/packager using RTMP.

The ingest server runs nginx-rtmp and it acts as input service, receiving the FFmpeg synthetic stream. It also transcodes (spawning FFmpeg processes for that) and creates the HLS format in a given folder.

The front end servers will consume the streaming via HTTP backed by this ingest server.

The front end server we chose was nginx, a scalable web server and reverse proxy. This will be the endpoint where the final users can access the html5 application to watch the stream. It will also work as a caching layer for scalability.

Finally, the app is a simple HTML static file that instantiates the player.

How to use it

The entire platform was conceived with Linux containers in mind so you just need to run make run and this is going to start it all. You also need to start the color bar in a different tab by running make broadcast_tvshow and point your browser to http://localhost:8080/app.

Conclusion

The genuine reason we created this simplistic video platform is to have a software where we can explore the computation at the edge. The next post will be empowering the Nginx front end with Lua code to add features to NOTT, things like authentication and IP blacklist.

How to build a distributed throttling system with Nginx + Lua + Redis

graph

At the last Globo.com’s hackathon, Lucas Costa and I built a simple Lua library to provide a distributed rate measurement system that depends on Redis and run embedded in Nginx but before we explain what we did let’s start by understanding the problem that a throttling system tries to solve and some possible solutions.

Suppose we just built an API but some users are doing too many requests abusing their request quota, how can we deal with them? Nginx has a rate limiting feature that is easy to use:

This nginx configuration creates a zone called mylimit that limits a user, based on its IP, to be able to only do a single request per minute. To test this, save this config file as nginx.conf and run the command:

We can use curl to test its effectiveness:

screen shot 2019-01-25 at 9.51.19 pm

As you can see, our first request was just fine, right at the start of the minute 50, but then our next two requests failed because we were restricted by the nginx limit_req directive that we setup to accept only 1 request per minute. In the next minute we received a successful response.

This approach has a problem though, for instance, a user could use multiple cloud VM’s and then bypass the limit by IP. Let’s instead use the user token argument:

There is another good reason to avoid this limit by IP approach, many of your users can be behind a single IP and by rate limiting them based on their IP, you might be blocking some legit uses.

Now a user can’t bypass by using multiple IPs, its token is used as a key to the limit rate counter.

screen shot 2019-01-25 at 10.22.00 pm

You can even notice that once a new user requests the same API, the user with token=0xCAFEE, the server replies with success.

Since our API is so useful, more and more users are becoming paid members and now we need to scale it out. What we can do is to put a load balancer in front of two instances of our API. To act as LB we can still use nginx, here’s a simple (workable) version of the required config.

Now to simulate our scenario we need to use multiple containers, let’s use docker-compose to this task, the config file just declare three services, two acting as our API and the LB.

Run the command docker-compose up and then in another terminal tab simulate multiple requests.

When we request http://localhost:8080 we’re hitting the lb instance.

screen shot 2019-01-25 at 10.58.25 pm

It’s weird?! Now our limit system is not working, or at least not properly. The first request was a 200, as expected, but the next one was also a 200.

It turns out that the LB needs a way to forward the requests to one of the two APIs instances, the default algorithm that our LB is using is the round-robin which distributes the requests each time for a server going in the list of servers as a clock.

The Nginx limit_req stores its counters on the node’s memory, that’s why our first two requests were successful.

And if we save our counters on a data store? We could use redis, it’s in memory and is pretty fast.

screen shot 2019-01-25 at 11.28.41 pm

But how are we going to build this counting/rating system? This can be solved using a histogram to get the average, a leaky bucket algorithm or a simplified sliding window proposed by Cloudflare.

To implement the sliding window algorithm it’s actually very easy, you will keep two counters, one for the last-minute and one for the current minute and then you can calculate the current rate by factoring the two minutes counters as if they were in a perfectly constant rate.

To make things easier, let’s debug an example of this algorithm in action. Let’s say our throttling system allows 10 requests per minute and that our past minute counter for a token is 6 and the current minute counter is 1 and we are at the second 10.

last_counter * ((60 current_second) / 60) + current_counter
6 * ((60 10) / 60) + 1 = 6 # the current rate is 6 which is under 10 req/m

To store the counters we used three simple (O(1)) redis operations:

  • GET to retrieve the last counter
  • INCR to count the current counter and retrieve its current value.
  • EXPIRE to set an expiration for the current counter, since it won’t be useful after two minutes.

We decided to not use MULTI operation therefore in theory some really small percentage of the users can be wrongly allowed, one of the reasons to dismiss the MULTI operation was because we use a lua driver redis cluster without support but we use pipeline and hash tags to save 2 extra round trips.

Now it’s the time to integrate the lua rate sliding window algorithm into nginx.

You probably want to use the access_by_lua phase instead of the content_by_lua from the nginx cycle.

The nginx configuration is uncomplicated to understand, it uses the argument token as the key and if the rate is above 10 req/m we just reply with 403. Simple solutions are usually elegant and can be scalable and good enough.

The lua library and this complete example is at Github and you can run it locally and test it without great effort.

Behind the scenes of live streaming the FIFA World Cup 2018

Screen Shot 2018-07-13 at 5.01.13 PM

Globo.com, the digital branch of Globo Group, had the rights to do the online live streaming of the FIFA World Cup 2018  for the entire Brazilian national territory.

We already did this in the past and I think that sharing the experience may be useful for the curious minds that want to learn more about the digital live streaming ecosystem as well as for the people interested in how Brazil infrastructure and user’s demand behave in an event with this scale.

Before the event – Road to the world cup

In average, we usually ingest and process about 1TB of video and users fetches around 1PB every single day. Even before the World Cup started, the live stream of a single soccer match had a peak of more than 500K simultaneous users with more than 400k requests per second.

When comparing these numbers to previous events such as the Olympic Games or the FIFA World Cup 2014 we can see an exponential evolution in demand.

Screen Shot 2018-07-13 at 5.00.44 PM

Back in 2014, Globo.com CDN was equipped with 20Gbps network interfaces. Now, the nodes were upgraded with 40Gbs, 50Gbs, and 100Gbps NICs. Processors were also upgraded enabling us to deliver 84Gbps on a single machine as part of the preparation for the World Cup.

I’m glad to say that the Linux/kernel fine-tune required was minimal since the newer kernel versions are very well tuned by default.

Screen Shot 2018-07-13 at 5.00.59 PM

We broke the simultaneous users record set by 2014 FIFA world cup way before the first 2018 World Cup matches. We also noticed an increase in the overall bitrate which likely point that the Internet infrastructure in Brazil improved significally in the past four years.

Plataform overview – The strategy 1:1:1

Let’s not focus on the workflow before the video arrives at our ingest encoders. Just think that it’s coming from Russia’s stadiums and reaching our ingest encoders directly. With this simplification in place, we can assume that there are basically two different users of interacting with the video platform: the ones producing the video and others consuming in the other end.

Screen Shot 2018-07-07 at 4.36.18 PM

Consumers of the video are the visitors of our internet properties and they watch the live content throughout Globo.com video player, which is responsible for requesting video content to Globo.com’s CDN or one of our CDN partners.

Globo.com player is based on Clappr, an open source HTML5 player that uses hls.js and shaka as its core playback engines.

Globo.com CDN nodes are mostly built on top of  OSS projects such as Linux, Nginx (nginx-lua), Lua Programming Language and redis. Our origin is made of multiple ingest points and a mix of solutions such as FFmpeg, Elemental and  OBS. A Cassandra cluster is also deployed with the responsibility of storing and manipulating video segments.

OSS projects play a key role in all the initiatives we have within our technology and engineering teams. We also rely a lot on dozens of open source libraries and we try as much as we can to give stuff back to the community.

If you want to know how this architecture works you can learn from the awesome post: Globo.com’s live video platform for the 2014 FIFA World Cup

Constrained by bandwidth – Control the ball

The truth is: the Internet is physically limited, it doesn’t matter if you got more servers, in the end, if a group of users have a link to us of 10Gb/s that’s all we stream to them.

Or we can explore external CDNs more pops but I hope you got the idea! 🙂

In a big event, such as the World Cup, there will be some congestions on the link between our CDN and the final users, how we tackle this problem (of a limited bandwidth) can be divided into three levels:

  1. OS :: TCP congestion control – the lowest level to control the connection, when it’s saturated, this control is applied to each user.
  2. Player :: ABR algorithm – it watches metrics such as network speed, CPU load, frame drop among others to decide whether it should adapt to a better or the worst bitrate quality.
  3. Server :: group bitrate control – when we identify that a group of users, which uses the same link, are using a link that is about to saturate, we can try to help the player to use to a lower bitrate and accommodate more users.

During the event – Goals

Even before the knockout stage, we were able to beat all of our previous records, serving about 1.2M simultaneous users during this match. Our live CDN delivered, at its peak, about 700K requests/s and our worst response time was half a second for a 4 seconds video segment.

Some of our servers were able to reach (peak) 37Gb/s in bandwidth. We also delivered the 4K live streaming using HEVC with a delay of around 25 seconds.

We are constantly evolving the platform and looking at the bleeding edge technologies such as AV1. With the help of the open source community and the growing amount of talents on our technology teams, we hope to keep beating records and delivering the best experience to our users.

References