5 ms·

Maximum theoretical downtime for a website: 30 minutes

3 points by paraschopra 16y ago

niyazpk 16y ago
AFAIK not only does the browser cache the IP address, but the many DNS servers (user ISP or any others) also cache the IP addresses. I wouldn't expect all the DNS servers out there to check for a new IP every couple of minutes. Many of them manage their cache according to their policies, not strictly according to the arbitrary constraints set by the individual domains. BTW shouldn't some default variant load in the site if your service is not working?
- paraschopra 16y ago
  Today, almost all DNS servers who cache entries respect TTL values. We did a number of benchmarks and observed that TTL was respected in all cases. Ultimately it boils down to the browser caching DNS. Yes, default variant loads if service isn't working. The slowdown part occurs when browser tries to connect to our service for loading JS but our servers don't respond. The timeout for this kind of behavior is 60 seconds. That means for only after 60 seconds, browser will abandon its attempt for connecting.
kilburn 16y ago
<blockquote> Our research led us to conclude that even for the best configurations in the world (yes, that includes Google, Amazon and all biggies), maximum theoretical downtime (for a few users) will be 30 minutes. You cannot possibly escape this limitation. </blockquote> You've overestimated your own abilities. "The biggies" don't need to use DNS failovers, they use anycast routing and the like. Therefore, their "maximum theoretical downtime" is probably much lower than 30 minutes...
- paraschopra 16y ago
  Even with Anycast, a portion of their users will still be routed to old, down server until DNS cache expires. No? I'd love to get clarified as this is a problem we desperately want to solve.
  - skorgu 16y ago
    IANANetwork Engineer but basically with anycasting you can announce a single IP from two different places and different providers will route to whichever is 'closest'. If your datacenter goes down one of those routes stops being advertised and providers will re-route to the next best advertised route. BGP cycle time is variable but way less than 30 minutes in many cases. This presumes you can essentially have two 'active' datacenters, not all application designs allow for this. You can still use BGP as a failover mechanism, just announce the route to your production IP from DC1 and if that dies, announce it from DC2 once you fail over. Again, limited by response time and BGP cycle time but that's faster than DNS.
    - paraschopra 16y ago
      Thanks for your info. I researched a bit and it sounds like Anycast is a very expensive option. But, anyway, I agree it could be faster than relying on DNS routing.
paraschopra 16y ago
Just FYI, I found this excellent guide http://www.tenereillo.com/GSLBPageOfShame.htm http://www.tenereillo.com/GSLBPageOfShame.htm

3 points by paraschopra 16y ago

AFAIK not only does the browser cache the IP address, but the many DNS servers (user ISP or any others) also cache the IP addresses. I wouldn't expect all the DNS servers out there to check for a new IP every couple of minutes. Many of them manage their cache according to their policies, not strictly according to the arbitrary constraints set by the individual domains. BTW shouldn't some default variant load in the site if your service is not working?

paraschopra 16y ago
Today, almost all DNS servers who cache entries respect TTL values. We did a number of benchmarks and observed that TTL was respected in all cases. Ultimately it boils down to the browser caching DNS. Yes, default variant loads if service isn't working. The slowdown part occurs when browser tries to connect to our service for loading JS but our servers don't respond. The timeout for this kind of behavior is 60 seconds. That means for only after 60 seconds, browser will abandon its attempt for connecting.

kilburn 16y ago

<blockquote> Our research led us to conclude that even for the best configurations in the world (yes, that includes Google, Amazon and all biggies), maximum theoretical downtime (for a few users) will be 30 minutes. You cannot possibly escape this limitation. </blockquote> You've overestimated your own abilities. "The biggies" don't need to use DNS failovers, they use anycast routing and the like. Therefore, their "maximum theoretical downtime" is probably much lower than 30 minutes...

paraschopra 16y ago
Even with Anycast, a portion of their users will still be routed to old, down server until DNS cache expires. No? I'd love to get clarified as this is a problem we desperately want to solve.
- skorgu 16y ago
  IANANetwork Engineer but basically with anycasting you can announce a single IP from two different places and different providers will route to whichever is 'closest'. If your datacenter goes down one of those routes stops being advertised and providers will re-route to the next best advertised route. BGP cycle time is variable but way less than 30 minutes in many cases. This presumes you can essentially have two 'active' datacenters, not all application designs allow for this. You can still use BGP as a failover mechanism, just announce the route to your production IP from DC1 and if that dies, announce it from DC2 once you fail over. Again, limited by response time and BGP cycle time but that's faster than DNS.
  - paraschopra 16y ago
    Thanks for your info. I researched a bit and it sounds like Anycast is a very expensive option. But, anyway, I agree it could be faster than relying on DNS routing.

paraschopra 16y ago

Just FYI, I found this excellent guide http://www.tenereillo.com/GSLBPageOfShame.htm http://www.tenereillo.com/GSLBPageOfShame.htm