Software-as-a-Service is a great thing. With just a swipe of a credit card, what once took months to provision and set up has now become something that can be purchased and configured before lunch. But as great as it is to use a SaaS service, building a reliable one can be equally challenging - especially in the cloud, where servers can disappear at any moment and infrastructure is expected to heal dynamically.
search.usa.gov is a SaaS that serves thousands of other government agencies. During our recent project to migrate the search.usa.gov server infrastructure to Amazon Web Services’ (AWS) Elastic Compute Cloud (EC2), we were presented with a challenging combination of requirements whose solution seems worth sharing:
search.usa.gov, or at “masked” domain name such as
search.usa.govrecords were required to be signed by DNSSEC
As we’ll see, this seemed at first to be a “pick two out of three” situation. Satisfying all three requirements at once required a spiffy open-source software feature and a two-tier DNS server architecture that we’ll describe in this post.
All problems in computer science can be solved by another level of indirection, except of course for the problem of too many indirections. - David Wheeler
The simplest of our three requirements was to allow customers to choose whether to use the
search.usa.gov website directly, or whether to create a “masked” domain name such as
search.someagency.gov. This was entirely in the customer’s control; they could simply create a CNAME in their own DNS zone to point
search.usa.gov if they wanted a masked domain name.
We knew that we wanted to avoid a “Big Bang” migration event in which all of our customers were suddenly pointed to the new AWS hosting infrastructure, so we went one step further by asking search.usa.gov customers to CNAME their agency-specific hostnames to customer-specific CNAMEs. This would allow us to shift traffic around on a customer-by-customer basis - both to the new infrastructure, but also back to the old infrastructure if the need arose:
$ORIGIN someagency.gov. search IN CNAME someagency.sites.infr.search.usa.gov
Once a fair number of our customers had these CNAMEs in place, we were able to switch them over to our new AWS infrastructure one-at-a-time rather than switching everyone at once:
$ORIGIN sites.infr.search.usa.gov. earlyadopteragency IN CNAME aws.search.usa.gov anotherdaringagency IN CNAME aws.search.usa.gov notchangingnowagency IN CNAME old-hosting-provider.search.usa.gov
The snippet above is a bit of a lie, however. We used wildcard DNS records at first to direct the majority of our customers to our previous hosting provider, and then later to our AWS-hosted site:
$ORIGIN sites.infr.search.usa.gov. earlyadopteragency IN CNAME aws.search.usa.gov anotherdaringagency IN CNAME aws.search.usa.gov ; All customers not listed above will go to the old hosting provider * IN CNAME old-hosting-provider.search.usa.gov
$ORIGIN sites.infr.search.usa.gov. * IN CNAME aws.search.usa.gov ; All customers not listed below will go to the new infrastructure notchangingnowagency IN CNAME old-hosting-provider.search.usa.gov
$ORIGIN sites.infr.search.usa.gov. ; Everyone goes to the new infrastructure * IN CNAME aws.search.usa.gov
At any step in this process, we were always able to go back to our zone file and add a customer-specific CNAME to direct traffic as needed for that customer, but eventually the migration came to an end and we were left with just the single wildcard record
*.sites.infr.search.usa.gov, pointing to our AWS infrastructure.
In Amazon Web Services (AWS), the problem of unreliable servers is solved by Elastic Load Balancing (ELB). An ELB containing one or more servers is presented to the world as a single hostname - say,
usasearch-elb.ec2.aws.com - and requests are routed to individual servers in the ELB pool based on health and capacity.
For practical reasons, ELBs for human-visible web sites are almost always hidden behind CNAME records in the DNS:
~ host search.usa.gov search.usa.gov is an alias for usasearch-elb.ec2.aws.com. usasearch-elb.ec2.aws.com has address 220.127.116.11
This is a fantastic and invaluable abstraction: what appears to be a single hostname called
search.usa.gov is actually a multi-datacenter, self-healing, auto-scaling pool of servers.
However, it runs afoul of one critical restriction of the DNS: the fact the top-most entry in a DNS zone (known as the “zone apex”) cannot be a CNAME. So if you want to add this ELB CNAME to the
service.gov zone, you’ll have no problem:
$ORIGIN service.gov. ; search.usa.gov is a CNAME to the ELB hostname search IN CNAME usasearch-elb.ec2.aws.com.
but if you only control the
search.usa.gov zone (which is the situation we found ourselves in), you’re out of luck:
$ORIGIN search.usa.gov ; This following line is not a valid DNS configuration @ IN CNAME usasearch-elb.ec2.aws.com.
There are numerous vendor-specific solutions to this problem, typically called ANAME or ALIAS records. They work around the zone apex problem by allowing you to configure the zone apex entry as though it were a CNAME, but presenting the answer to the caller as though it were an A record:
$ORIGIN search.usa.gov ; '@' means "the zone apex", i.e. search.usa.gov @ 60 IN ALIAS usasearch-elb.ec2.aws.com.
# Here are the current IP addresses for our ELB ~ host usasearch-elb.ec2.aws.com usasearch-elb.ec2.aws.com has address 18.104.22.168 usasearch-elb.ec2.aws.com has address 22.214.171.124 usasearch-elb.ec2.aws.com has address 126.96.36.199 # Notice that there are no mentions of CNAMEs in this answer ~ host search.usa.gov search.usa.gov has address 188.8.131.52 search.usa.gov has address 184.108.40.206 search.usa.gov has address 220.127.116.11
This sleight-of-hand is accomplished by dynamically resolving the A record request at the authoritative nameserver by looking at the underlying ALIAS and getting its current value. By design, the answer issued by the authoritative nameserver has a short (60-second) TTL, because the “correct” answer for
usasearch-elb.ec2.aws.com could change suddenly in an environment where servers pop in and out of existence without warning. At the cost of additional traffic to the authoritative nameservers, incorrect results are quickly flushed out of the global DNS cache when the ALIAS lookup results change.
With ANAME/ALIAS records (I’ll just call them ALIAS records for the rest of this post) at our disposal, we could easily satisfy our first two requirements. However, doing so required picking a DNS provider with ALIAS support - and no government-approved hosted DNS providers who support ALIAS also support DNSSEC. Therefore we knew we needed something more.
(As an aside: if we had been willing to forgo the ELBs and their capability for dynamic server replacement, the DNSSEC requirement would have been no problem. We could have have chosen instead to build proxy servers with static IP addresses so that our DNS record contents would never change. This would have made it easy to satisfy requirements #1 and #3 - hence the earlier comment about a “pick two out of three” situation - but we would have lost the ability to dynamically scale or replace servers using Elastic Load Balancers.)
To find the solution, it was valuable to look at why DNS providers don’t provide DNSSEC support along with ALIAS records. The design of DNSSEC - and the cryptographic assurance it provides about DNS record values - requires taking all of the records in a zone file and computing a cryptographic hash of them, called an RRSIG. Computationally, this is a very expensive operation, especially compared to the cost of answering a single DNS request. Therefore RRSIG calculations are done when the contents of the zone change, not on-the-fly while answering requests.
Behind-the-scenes ALIAS expansion throws a wrench into the works here. If the result of a lookup can be different at different times (such as when the list of IP addresses for the
usasearch-elb.ec2.aws.com ALIAS changes), then the RRSIG itself might need to be recomputed at any moment. In the general sense, this is not practical for DNS services that may be serving hundreds of thousands or millions of requests per minute.
After we came to this realization, we had to rummage around the Internet for help. A lot. This discussion thread told us that this issue was at the forefront of the minds of some Very Smart People working on IETF-related projects, and so we pursued it to the people who were proposing a solution: PowerDNS, an open-source DNS software provider based in The Netherlands.
Peter van Dijk, one of PowerDNS’ software developers, was quick to clarify the realizations that we’d had:
Peter then suggested a very simple improvement to the PowerDNS software that could solve our problem: adding the ability for a DNS server to expand ALIAS records (which it already supported) into A records during a zone transfer, or AXFR.
This small change, called outgoing-axfr-expand-alias and available beginning in PowerDNS Authoritative 4.0.0, allows one server (the so-called “DNS master”) to be the only one that knew about our ALIAS records. Then, every minute, “DNS Slaves” acting as the authoritative nameservers for
search.usa.gov would initiate an AXFR of the zone from the DNS master, and would receive a copy of the zone file containing the most up-to-date values for those ALIAS records, expressed as A records. The result of this AXFR would then be compared with the current contents of the zone on the slave server. If the contents had not changed - almost always the case - then no action would be taken. If the contents had changed, then the zone be would reloaded entirely and re-signed using existing DNSSEC signing features, with notification sent to sysadmins.
Overall, the process looks like this:
You can see the script that we wrote to request the AXFR from the master server and compare its contents to the current slave server zone file here. It’s quite simple, and relies on straightforward zone management features already built into the PowerDNS software. [Strictly speaking, this script is no longer necessary and was developed as a precaution against the possibility of a failed AXFR “emptying” the zone on the slave server. In the released version of PowerDNS 4.0.0, a standard zone slaving configuration will have such protection enabled automatically.]
Remember that bit earlier about how we used wildcard DNS records to simplify the process of creating customer-specific CNAMEs? You might have expected that to cause a problem with respect to DNSSEC and dynamic records. However, as it turns out, wildcard records don’t present the same problem to DNSSEC ALIAS records present.
Let’s take a closer look at what happened in our two-step zone updating process. First, the master server contains a wildcard ALIAS record that points customer sites to an ELB CNAME:
$ORIGIN search.usa.gov. *.sites.infr IN ALIAS usasearch-elb.ec2.aws.com
After outbound ALIAS expansion during a zone AXFR, that becomes
$ORIGIN search.usa.gov. *.sites.infr IN A 18.104.22.168 *.sites.infr IN A 22.214.171.124 *.sites.infr IN A 126.96.36.199
Those records, although wildcards, are no longer dynamic. And DNSSEC supports signing wildcard records, making these results as valid as any other A record that might appear after ALIAS expansion.
So, thanks to some tricky multi-tier design, our solution now works as follows:
For search.usa.gov agency customers with DNSSEC-signed zones of their own, this setup allows them to select their own customer-specific CNAME, delegate it to
search.usa.gov, and operate with confidence that the entire chain of DNS resolution is signed with DNSSEC and safe from cache poisoning or man-in-the-middle attacks.