Revving Up For Edge Computing | Hacker News Reader

Revving Up For Edge Computing

semiengineering.com

47 pointsSemiTom18 comments

Does anyone else feel like "Edge" is made up term just for deluding investors?

The client/server split has been around forever. Edge means the same thing as a client. With IOT you have more clients. Modern smartphones and ras-pi type hardware are already great at neural net inference... The future is already here.

I work at Red Hat and am involved with edge technologies, while if you look at it in a generalized way, sure, it seems like "huh, it's client server with lots of clients" but it does really refer to novel new architectures being deployed by customers where there are more computing resources out in the field than would normally be present.

more simply, Chick-Fil-A's post about their edge efforts really made it clear for me when the term first came out - a miniature datacenter in every store. this is definitely novel: https://medium.com/@cfatechblog/edge-computing-at-chick-fil-...

I always felt like edge computing was more along the lines of clients that do all the computation themselves, then federate their data among peers, instead of each client connecting to a server.

There do seem to be many different definitions of "edge". The parent and grandparent above, big CDNs with "edge functions", etc.

Has anyone compiled a list of concrete example use cases?

Not to my knowledge, but I've found one that calls itself an 'edge' service, and some I might classify as 'edge'.

- snips.ai calls itself an 'edge' ai service because all the computation is done on your device.

- ledger cli is one small project that I consider to fall under the 'edge' umbrella, because all it's 'computation' is done on your own device, even though that's a tiny amount of computation.

- taskwarrior is another that I might call edge, because it also handles all computation, reporting, etc, on your device. There can be a server involved, but it only handles synchronization of data, i.e. federation

I would not include ledger in that category. Ledger is just software.

I can see your point. If ledger is edge, then so is firefox, or notepad. The definition would be too broad to be useful.

That's one of the most interesting architectural / conceptual blog posts I've seen in a while, thanks for sharing!

(it makes me think there must be a ton of interesting scenarios around applications where images/containers migrate between the core/cloud and the edge on-demand (or simply have edge replicas), and situations in which mapreduce style computation could be farmed out to edge clusters before sending aggregated results home)

So every edge device run it's own Docker container and the fog or the nodes act as a swarm manager before depositing the relevant information into the cloud? I've always thought that we would see drone-based air traffic controllers to manage and handle air traffic, especially as we can gain flying vehicles in the future.

I'm not sure I completely grasped the air-traffic control example you mentioned, but I think you're on the same track, yep. Low-latency, localized computation within a well-connected physical area.

To try to make it more concrete, here's a scenario:

Imagine that smartphones are able to opt-in to become part of an edge cluster which communicates over some kind of local network fabric (be that wi-fi / 5G / similar).

Now imagine that you have 2,000+ fans at a sporting event or live performance, all with their phones and many taking live recordings.

Shifting high-fidelity video/audio data from that many devices at an event back to the cloud in realtime might not be particularly feasible and/or useful for various network contention, bandwidth, and latency reasons (both client and server-side).

But if those devices were already running a containerized application to perform -- say, 3D image stitching[0], for example -- you could collect, compute and redistribute results via the cluster in near-real-time, potentially providing some pretty immersive audience experiences (3D highlights and replays and all kinds of interesting augmented reality interpolation of audience & performers).

It'd also raise questions around who owns/authorizes the on-device computation, who has permission and copyright over the data captured, and various other issues.

All very hypothetical ideas, but technically feasible. Predicting fast food order demand is much more practical challenge to begin with, no doubt :)

[0] - https://www.geekwire.com/2013/microsoft-updates-photosynth-w...

It sounds a bit like pre-cloud business IT architecture. Put the servers somewhere on-site.

It is really hard to see much new here. 20 years ago you might put a Windows 2000 server in each restaurant... What's changed is we have better server management tools: ansible/puppet/docker/kube...

The reason we moved to the cloud was because the Internet gained massive capacity and there were economies of scale in huge datacenters.

What chick-fil-a is doing is'nt wrong, but feels more like the pendulum swinging than actually new tech.

Somewhat.

I like the Samsung SmartThings/AWS Lambda model for internet of things applications in most respects.

What I don't appreciate is that a light switch that is powered by AWS will sometimes fail to toggle the lights or may be a bit slow depending on the internet.

So if "edge" means having the ability to move serverless or container applications away from the cloud when that makes sense, I am all for it. That is, be able to write a lambda function and run it in a box in the office or run it in the AWS cloud.

However, if "edge" means that you pay American Tower prices to rent an "edge" server right at your cell phone tower, that seems silly in so many ways -- not least that you defeat the point of mobility just by moving to another cell tower!

Pretty sure that you're not renting right at the tower. The provider will have an aggregation point in the middle of large metros and round-trip latency from your phone should still be under 5 ms, which is way better than going all the way to the nearest cloud datacenter. (If it was right at the tower then you'd see an order or two lower even, but I don't think that's the model being promoted). At least that's my understanding.

This is a whole new layer of compute and storage between the cloud and the endpoint devices. It's still a client/server model, but it provides more control over the data closer to the source. This has a number of benefits. One is that you can control who sees that data. Second, response time is faster. And third, you can get rid of more useless data much more quickly and cheaply closer to the source of that data than if you have to send it.

Where Is The Edge? https://semiengineering.com/where-is-the-edge/

This particular article basically uses Edge Computing as a synonym for client-side computing. I've more commonly seen the phrase Edge Computing used in a different and more interesting way. Basically, a typical web app architecture would have servers for dynamic content in 2 or 3 places, mainly for redundancy, and use a CDN to distribute static content. The CDN in this case is the Edge, but there's no compute: just caching of static assets. That's not a great fit for some applications (those with high bandwidth or low latency requirements on dynamic data especially), but there's not really been provider support for that, so if you wanted to do it you had to deal with your own hardware in a large number of locations. If a cloud provider adds Compute to the Edge (ie Lambda@Edge from AWS, Cloudflare Workers), this architecture becomes doable for small shops.

Not investors, middle and upper management. Gotta justify those consulting fees with new and hip stuff.

I have always thought of edge computing as a CDN that serves more than just static assets. The power of the "edge" becomes apparent when you realize commodity hardware that's using RISC/RISC-V can now do on-device machine learning or host basic REST services. It's also key for lowering latencies and increasing availability in a globally distributed service.

Reply on news.ycombinator.com