4 ms·
I thought DNS (particularly public) was basically immune to DDoS? If one DNS server is down, use the cached result or another server. DNS is some of the most
by paulddraper 10y ago
I thought DNS (particularly public) was basically immune to DDoS?
If one DNS server is down, use the cached result or another server.
DNS is some of the most distributable, cachable data I can imagine.
- facorreia 10y ago> DNS is some of the most distributable, cachable data I can imagine. Indeed. How can it be so broken? And, it seem that it would make sense to use the last known value if upstream is unavailable.
- nzp 10y agoThat would be a horrible idea from a security standpoint. Although not easy by any means for an ordinary attacker, at least one exploit scenario is obvious.
- pepon 10y agoSeriously, please someone explain this. I thought DNS info is propagated among DNS servers, and it is cached on all of them... How can an attack to a single DNS company cause this outage??
- atom_enger 10y agoExplained above for ya
- 274below 10y agoTrue, but DNS records have TTLs that expire, which forces you to reach back out to the authoritative server for the zone. A very decent number of records will have a TTL of 1hr, which means that a total DNS server outage lasting longer than an hour = 100% failure rate. You could always increase the TTL to a day (or whatever other arbitrary value), but that means that you'll be waiting upwards of a day for clients to see any updates to your DNS records. It's a tradeoff between different types of resiliency. This is a simple case of a DNS provider evidently not having their crap together when it comes to a DDoS of any decent size. There are plenty of other DNS providers out there who are working fine currently, this DDoS is limited to just the one. It's all of that company's clients that are impacted. I'm sure that DynETC will post something afterwards about how this was the largest DDoS they've ever encountered by many orders of magnitude and that there was no way for anyone to ever be prepared to take so much traffic... but at this point I think they're inept. edit: just came back up. The TTLs on a few common records explain why this was so obviously a problem... - www.reddit.com: 300 seconds (pointer to a fastly.net address, which is hosted via dynect) - prod.reddit.map.fastlylb.net: 30 seconds - api.twitter.com: 300 seconds - herokussl.com: 3600 seconds So, an outage of five minutes is enough to take out access to all of reddit and twitter. An outage of an hour is enough to take out heroku. Of course, these are best case scenarios: in reality, 50% of your users would lose access to reddit after 15 seconds (the fastly record), and so on.
- atom_enger 10y agoThe problem here lies with the authoritative nameservers. You have to retrieve the record from somewhere initially. The trouble here is that since these companies are using Dyn and only Dyn to be their Nameservers. When a dns request is made for github.com they're attempting to contact ns3.p20.dynect.net or another nameserver provided by Dyn. So yes caching will work for a little while, but only for the length of the TTL which your client is designed to respect. Once that TTL expires, your client and upstream DNS provider will attempt to contact that nameserver for a fresh record. Since you can't contact the nameservers, the website is effectively offline for all name resolution. Now if you know the record you can always forge the record locally in /etc/hosts or in a local dns resolver which allows for overrides. Hope this helps.
- indigomm 10y agoIt was true that one DNS nameserver record == one physical server even at large providers. But we're beyond that now. Each nameserver at a physical location can be a cluster of hosts. Beyond that, with the use of anycast that single nameserver record may map to different clusters positioned around the world. This is how the root servers work and why they are more difficult to attack. Of course small DNS providers will find it hard to run a system this way, but the larger providers follow the same architecture - anycast and multiple servers at each location. Google and OpenDNS for a start use this pattern - the famous 8.8.8.8 and 8.8.4.4 are in fact multiple server clusters all around the world.
- atom_enger 10y agoSure - I assumed that's what the large organizations were doing but for the sake of explanation to those who didn't understand why the problem existed in the first place I didn't want to complicate my answer with something even more complex :) Appreciate your help doing so