Per-Entity Metadata Working Group

["Cantor, Scott" <>]

On 9/14/16, 3:58 PM, 
"
 on behalf of Nick Roy" 
<
 on behalf of 
>
 wrote:

>    I'll also just leave this here for you to peruse/think about: 
>    
> https://loadbalancer.org/blog/gslb-why-do-global-server-load-balancers-suck

There are lot of those articles around, many of which I've read very 
recently, and they all share some characteristics:

- they're years old, sometimes 10+
- they don't actually offer a solution to the problems

When state isn't a consideration, I would agree that certainly one has more 
options, and it's probably easier to manipulate DNS or just make clients a 
little smarter, and you'll end up most of the way there anyway. And if that's 
how people serve content now, that's fine, though of course they're doing so 
on the backs of browsers having been forced to do an absolute ton of changes 
to handle DNS and failover differently than they did in 1995. In other words 
"GSLB sucks if you can make Firefox do it all for you". Well, we're not 
Firefox unfortunately.

Speaking more generally, if you add sticky state (and this doesn't apply to 
MDQ, I'm talking in general), suddenly all these tricks fail horribly, and 
you either do GSLB, hack your own GSLB (could be as simple as hot/cold of 
course), or you just hope you never have a network outage I guess.

The modern GSLB devices really do the crazy things you thought you'd have to 
make them do to pull it off, because they had to do something to justify the 
prices. They proxy each other in real time behind the scenes to prevent 
client addresss switches from breaking apps. Jim Fox described doing that 
himself with cookies + Apache many years back. It is basically the only way 
to avoid a real-time SPOF from your network layer.

Do they work? I'm probably going to find out soon and will certainly let you 
know. If they don't, then there really isn't much I can do about it, so 
hot/cold it is.

-- Scott