When I use the term "DNS Round Robin" I generally mean in in the sense of the "cheap load balancing technique" as OP describes it.
But that's not the only way DNS can be used for global high availability. Most of the time, it's just hard for people with different (technology) backgrounds to communicate well.
The best load balancing technique (if money is not a problem) is generally considered to be:
- A Anycast'ed global network of 'intelligent' DNS servers,
- and a set of globally spread out datacenters,
- where each DNS node implements Split Horizon DNS,
- and monitoring of availability and traffic flows are available to the 'intelligent' DNS nodes in some fashion,
- so that the user DNS request flows to the nearest DNS server via IP Anycast,
- and this DNS server hands out a low-TTL A Record / set of A Records for the nearest / best datacenter for this end user via 'intelligent' split horizon DNS.
Using anycast for DNS is generally fine, because DNS responses are stateless and almost extremely short. So if the BGP routes change it's highly unlikely to interrupt a DNS query.
Anycast is less suited for the longer and stateful HTTP conversations, thus this system uses split horizon DNS. A HTTP session between a client and server is kept to one datacenter; it generally cannot fail over to another datacenter without breaking the session.
As I indicated with "set of A Records" what I would call 'DNS Round Robin' can be used together with the setup above. It is typically used to spread the traffic load over multiple highly available load balancers in each datacenter (so that you can get better redundancy, use smaller/cheaper load balancers, not overwhelm the Unix network buffers of a single host server, etc).
So, is it true that, with multiple data centers
and HTTP traffic, the use of DNS RR is the ONLY
way to assure high availability?
No it's not true, not if by 'DNS Round Robin' we simply mean handing out multiple A records for a domain. But it's true that clever use of DNS is a critical component in any global high availability system. The above illustrates one common (often best) way to go.
Edit: The Google paper "Moving Beyond End-to-End Path Information to Optimize CDN Performance" seems to me to be state-of-the-art in global load distribution for best end-user performance.
Edit 2: I read the article "Why DNS Based .. GSLB .. Doesn't Work" that OP linked to, and it is a good overview -- I recommend looking at it. Read it from the top.
In the section "The solution to the browser caching issue" it advocates DNS responses with multiple A Records pointing to multiple datacenters as the only possible solution for instantaneous fail over.
In the section "Watering it down" near the bottom, it expands on the obvious, that sending multiple A Records is uncool if they point to datacenters on multiple continents, because the client will connect at random and thus quite often get a 'slow' DC on another continent. Thus for this to work really well, multiple datacenters on each continent are needed.
This is a different solution than my steps 1 - 6. I can't provide a perfect answer on this, I think a DNS specialist from the likes of Akamai or Google is needed, because much of this boils down to practical know-how on the limitations of deployed DNS caches and browsers today. AFAIK, my steps 1-6 are what Akamai does with their DNS (can anyone confirm this?).
My feeling -- coming from having worked as a PM on mobile browser portals (cell phones) -- is that the diversity and level of total brokeness of the browsers out there is incredible. I personally would not trust a HA solution that requires the end user terminal to 'do the right thing'; thus I believe that global instantaneous fail over without breaking a session isn't feasible today.
I think my steps 1-6 above are the best that are available with commodity technology. This solution does not have instantaneous fail over.
I'd love for one of those DNS specialists from Akamai, Google etc to come around and prove me wrong. :-)
You can't rely on UDP to deliver packets in order because the specification doesn't provide those guarantees. Even assuming the most ideal situation, a single piece of ethernet cable between two hosts, there is still the matter of the OS, the network stack, the NIC driver, and the libc implementation that your writing against.
At every step in that chain, the writers of that code will have chosen NOT to prioritise ordering UDP packets even if they arrive in order for the simple reason that they don't have to.
One contrived example could be the data structure that incoming packets are read into, which might be a ring buffer. Packets arriving in order, will be placed, in order into the ring buffer, but it may be simpler for the driver writer to dump them to the upper layers of the networking code in memory order, hence randomising their ordering.
Taking your situation, a virtual machine run on a shared infrastructure that will be run for volume, not performance, then the probability of predicting the order UDP packets will be received will be low.
In short, if the spec says you can't rely on UDP packet ordering. You can't rely on it, and you can't try to tweak the environment to give a stronger guarantee than the spec ever promised.
Best Answer
First off, there is no exact formal definition of what a "Point of Presence" (POP) or a "Datacenter" (DC) is.
Next, I would say you're somewhat looking at this from the wrong perspective. What's called POP and DC is often not defined by the tech group of the ISP. They're often chosen by, or chosen with significant input from, the sales/marketing group. So don't rely too much on these words.
My personal rough take on these words is that:
I don't quite agree. A CDN also has a commercial need pushing their design in the opposite direction -- the more POPs the CDN has, the more duplication they have of content (i.e. more storage is required, and objects are served fewer times from cache). CDN's mainly want to add POPs when: