Geo-distributed web hosting

[sudo]

Hi,

Does anyone here have a use for, or know others with a use for, geo-distributed web hosting? Something close to what companies like Tronkle (http://tronkle.com/index.php/gdh/description) are offering...

If there is a use-case, on what scale would it be helpful (static/database-backed site)?

Interested to hear your thoughts. Tx.

 

[koffiejunkie]

At work I deal with a few clients who have this. There are generally two use cases:

1. Redundancy, i.e. if Elon Musk miscalculates and lands on our DC, your site can switch over

2. Improving response time by cutting out trans continental or trans oceanic latency. You might, say, have clients in the US and in ZA, and have hosting in a DC in JHB and in the US somewhere. US based clients will hit your US based servers, while ZA clients will hit your local servers. The complication with this is keeping data in sync between locations - it's particularly challenging with database-based sites.

For static content this is already very commonly used - think Akamai...

 

[sudo]

For static content this is already very commonly used - think Akamai...

Thanks.

Essentially Akamai - and like services - are just CDNs though, correct me if I'm wrong?

What I was asking about was more like a traditional web hosting service for smaller businesses that will take care of all the hassle of distributing the site and dns to global locations for visitors... Does a CDN address that kind of problem?

Tx

 

[koffiejunkie]

Essentially Akamai - and like services - are just CDNs though, correct me if I'm wrong?

What I was asking about was more like a traditional web hosting service for smaller businesses that will take care of all the hassle of distributing the site and dns to global locations for visitors... Does a CDN address that kind of problem?

That's exactly what a CDN does, but as I said, only for a limited application. i.e. static content.

I don't know of any hosting service that does geo-distributed hosting by default. It's not a trivial setup, and if you're hosting anything other than a completely static site, your site has to be designed with this in mind, i.e. each instance has to know about the other one and they need to talk to one another. If it's a database backed site, for example, let's say you run a site that accepts some sort of user content - a blog that accepts comments, a dating site, etc). Now you need bi-directional database replication between two data centers. This is bad enough between two hosts in the same network, but if both are very busy and they have to deal with the increased latency, things can go wrong in a hurry, and you can end up with a split brain situation. And trust me, that's a very very bad place to be.

 

[koeksGHT]

Set up multiple servers around and there must be a way to choose a nameserver based on IP location?

 

[tnero]

Take a look at https://www.cloudflare.com/

 

[ambo]

a way to choose a nameserver based on IP location?

Thats not quite how they do it. Rather they look at the address that the DNS query comes from and send different A (or AAAA) records back depending on where they guess that you are located.

Its quite an art. Thats why guys like Akamai are earning so much money doing it well.

 

[koffiejunkie]

Thats not quite how they do it. Rather they look at the address that the DNS query comes from and send different A (or AAAA) records back depending on where they guess that you are located.

I'm not sure about that. The DNS servers will return the same IP regardless of where you're querying from. Your nodes will have the same IP address. We then use anycast to route your packets to the nearest node with that address.

 

[ambo]

I'm not sure about that. The DNS servers will return the same IP regardless of where you're querying from. Your nodes will have the same IP address. We then use anycast to route your packets to the nearest node with that address.

Anycast is generally only user for stateless protocols like UDP. There are a few nasty caveats with doing TCP over anycast. For example: using anycast you are at the mercy of the ISPs who will choose where they send their traffic whereas geodns allows you to load balance quite effectively.

That said, there are a few CDNs now using anycast to serve their content but its far from the norm.