Hacker News Reader

Comment by daviddumenil | Hacker News Reader

This has been offered by their competitor Incapsula for some time under the name 'Global Server Load Balancing'

https://www.incapsula.com/load-balancing-failover.html

And AWS Route 53: https://aws.amazon.com/route53/ Or Microsoft Traffic Manager: https://azure.microsoft.com/en-us/services/traffic-manager/

And possibly many, many other providers.

It would be interesting to know if there's some innovation here.

Because Cloudflare uses Anycast there is no change in the public IP address for a request that hits us. It will be routed to the nearest data center (of which we now have 100 cities covered globally). The Traffic Manager product controls how that traffic is the routed to the origin server (if the request could not be served from cache).

As the active monitoring is performed from each location separately the Traffic Manager applies its policies in each location and can automatically route around connectivity problems in the Internet as well as origin server problems. All without changing public DNS.

This allows us to do very fast failover and route changes because this is independent of the propagation time for DNS on the public Internet. Solutions based on changing DNS records are suboptimal.

Somewhat off topic, but since CloudFlare keeps mentioning "routing the nearest data center" as one of its benefits...

I have a number of clients near ICN (Seoul) who have been happy with CloudFlare for some time, but now the vast majority of their users are being routed through LAX. Not Tokyo, nor Osaka, nor any of your other locations throughout East Asia, but LAX all the way across the Pacific Ocean! This detour adds about 300ms to every uncached request, and 150ms to every cached request. All the time I spent helping my clients shave a few milliseconds off their response times, now completely obliterated!

When the Pro plan ($20/mo) users contacted CloudFlare to find out what was going on, they were told to upgrade to the Business plan ($200/mo). Some of them did, but their users were still being routed halfway around the world. When they contacted support again, they were told to consider upgrading to an Enterprise plan. What is this, a cheap webhosting shop that tells people to upgrade whenever they're having a problem?

I don't know what happened around late summer that suddenly made Korean and Japanese locations off limits to all but the highest paying customers (local ISPs demanding more money?) but whatever it is, it has made CloudFlare barely usable where I live.

That experience with support doesn't sound right at all. Please email jgc @ cloudflare so I can follow up with you on this.

Okay, I'll gather some information from people affected by this and get back to you.

This is definitely correct. Every Cloudflare pro plan site I know of now routes to LAX for South Korean users. For business sites, going over a certain amount of traffic, they also route to LAX for South Korean users.

I also have confirmed with Cloudflare support, that sites going over a certain amount of traffic have to upgrade to Business -> Enterprise plan to keep South Korean users to route to ICN (Seoul). This is causing huge continous performance problems for nearly all Cloudflare sites in South Korea.

http://www.todayhumor.co.kr/cdn-cgi/trace http://www.ilbe.com/cdn-cgi/trace http://www.issuein.com/cdn-cgi/trace

All show loc=KR, while using a slow LAX colocation. The first two from what I know, are business plan sites.

It doesn't have anything to do with how much they're paying (if Cloudflare uses the same network for all customers). Most likely, an Asian IP transit provider found a peer in an IX in LA and decided it was better to "cold potato" the data over there than deliver it to a nearby connection. Or it decided to de-peer for some stupid reason, or a route became "AS best path" to LA despite being actually worse in a way BGP can't figure out, et cetera.

Sometimes you can fix this with BGP communities, sometimes you have no control over it and are basically screwed unless you find another transit option that works better for everyone.

I've been building an Anycast network all week. Right now my problem is Australia. I make Sydney work better by peering and half the world starts best path routing to it. Debugging these problems is really tricky, particularly in Asia/Pacific, which is full of expensive submarine cables and, frankly, crappy communication monopolies that love to fight with eachother over peering.

Basically you need to get a contact to their NOC team and get them to check the BGP routes for those customers. I have no idea how they diagnose, but I myself would ask for some traceroute -A yoursite.com outputs from the people with the weird routing and go from there. Don't expect overnight results, the fix may require another transit hookup or peering agreement.

Yep, I understand the difficulties. It's bad enough trying to get decent international transit in these parts as a local customer.

But if the matter really is beyond CloudFlare's control, their support should not be encouraging people to upgrade to a plan that costs 10x as much.

I saw the same thing with China. We have something hosted on IBM Softlayer in HKG that takes a detour through LA. Apparently all China traffic for Cloudflare must go through LA and that's just now China works, we were told.

That's awesome and definitely a different product.

Thank you very much for sharing!

Neither of those services are http(s) load balancers and are merely (specialised) DNS managers.

It was my understanding based on the website's copy that this was also a DNS GLB:

Expanding on Cloudflare’s highly available DNS infrastructure and global Anycast network, the Cloudflare Traffic Manager is a cloud-based load-balancing solution that enhances the performance and availability of your Internet facing infrastructure. Designed for scale, flexibility and robustness, CTM enables your applications to deliver the ultimate end user experience through the highest level of performance and reliability.

It has since been clarified that this product has the ability to re-route traffic based on availability and it does not base its balancing in returning different results for the same DNS query.

P.S.: I'm not saying that the copy "says" this is DNS GLB, I'm just saying it was vague enough for me to assume it was a DNS GLB.

I hope you'll apply for Early Access and put our new service through its paces.

A colleague just published more details here https://blog.cloudflare.com/cloudflare-traffic-manager-the-d...

Reply on news.ycombinator.com