[krystofbe]

Looks like a DNSSEC issue, not a nameserver outage. Validating resolvers SERVFAIL on every .de name with EDE:

RRSIG with malformed signature found for a0d5d1p51kijsevll74k523htmq406bk.de/nsec3 (keytag=33834) dig +cd amazon.de @8.8.8.8 works, dig amazon.de @a.nic.de works. Zone data is intact, DENIC just published an RRSIG over an NSEC3 record that doesn't validate against ZSK 33834. Every validating resolver therefore refuses to answer.

Intermittency fits anycast: some [a-n].nic.de instances still serve the previous (good) signatures, so retries occasionally land on a healthy auth. Per DENIC's FAQ the .de ZSK rotates every 5 weeks via pre-publish, so this smells like a botched rollover.

[qazwsxedchac]

So a single configuration mistake in a single place wiped out external reachability of a major economy. It happened in the evening local time and should be fixable, modulo cache TTLs, by morning. This will limit the blast radius somewhat.

Still, at this level, brittle infrastructure is a political risk. The internet's famous "routing around damage" isn't quite working here. Should make for an interesting post mortem.

[belorn]

I am reminded of the warning that zonemaster gives about putting your domain name servers on a single AS, as is common practice for many larger providers. A lot of people do not want others to see this as a problem since a single AS is a convenient configuration for routing, but it has the downside of being a single point of failure.

Building redundant infrastructure that can withstand BGP and DNS configuration mistakes are not that simple but it can be done.

[walrus01]

As the CPU/RAM resources to run an authoritative-only slave nameserver for a few domains are extremely minimal (mine run at a unix load of 0.01), it's a very wise idea to put your ns3 or something at a totally different service provider on another continent. It costs less than a cup of coffee per month.

[belorn]

For a very long time, the computer club I was in operated a DNS server on a Pentium 75MHz and after the last major hardware upgrade it had a total of 110MB RAM memory and 2G disk space. It worked great except that before the upgrade it tended to run out of ram whenever there was a Linux kernel update, a problem we solved forever by populating all the ram slots with the maximum that the motherboard could handle to that nice 110 MB.

[psd1]

Did you populate the motherboard with the most it could handle, or the most you could assemble from a box of assorted sticks?

Otherwise, 110MB would hint at a fascinating engineering culture at the motherboard manufacturer.

[walrus01]

If I remember right there were certain very early pentium 3 processor competitors from VIA and other non-intel, non-AMD sources (with much worse performance) that had integrated onboard SVGA video, where the video RAM was shared with the system DRAM. Meaning that depending how you configured the video in the BIOS, you could have something like a 128GB RAM server "minus" 16GB RAM withheld for video, with like 112GB usable by the OS.

But if this guy is talking about a pentium 75 MHz (socket 5 CPU) that's a totally different generation of stuff several generations before that.

[account42]

This makes sense for larger providers but just for a small/personal website there is literally zero advantages to having distributed authoritative DNS servers when the webserver is on a single host.

Ironically, denic still requires you to have two separate name servers with different IPs for your domain (which can be worked around by changing the IP of the registered name server afterwards lol), a requirement that all other registries I use have dropped or never had because enforcing such a policy at the registry level makes zero sense.

[walrus01]

For a domain owned by someone in North America, it costs me literally $1.50 a month to have an authoritative only ns3 in Europe on a totally different ISP.

[icedchai]

It depends. Do you also have email or other services for that domain? The advantage is your email doesn't start bouncing when your single host web site / DNS server is down.

[account42]

Email bouncing during rare downtimes is hardly that big of an issue - if its actually important the sender will retry, possibly with a different contact method. And for short downtimes most likely the sender's MTA will just automatically retry a bit later - email is designed to work with temporary failures.

There isn't some magic reliability that everyone needs which just so happens to fall into "not achievable with a single authoritative name server" and "guaranteed with two servers". I'm not saying you should never have more than one, just that isn't the registry's business to decide what kind of availability guarantees you need for your domain.

[icedchai]

It's simple enough to get a secondary DNS server somewhere and put it on $5/month VPS. I use BIND and DNS replication (AXFR/IXFR) handles it.

[ondohotola]

Have you ANY clue about the size of .DE's name server infrastructure?

[throw0101c]

> Have you ANY clue about the size of .DE's name server infrastructure?

Is it more or less than the F-root server run by ISC?

* https://www.isc.org/f-root/

If you want, you can even request your own instance (a 1U Dell):

* https://www.isc.org/froot-process/

Or an instance of ICANN's L-root server, also 1U:

* https://www.dns.icann.org/imrs/

Would .de have more, or less, traffic than some of the root servers?

[icedchai]

Are you following the thread? We're talking about redundancy for a single domain here.

[amiga386]

The single domain here is a ccTLD, and DNS's heirarchical nature means your personal domain's redundant DNS can't mitigate an outage at the ccTLD level.

[deepsun]

On Google cloud it's always four nameservers like

    ns-cloud-c1.googledomains.com
    ns-cloud-c2.googledomains.com
    ns-cloud-c3.googledomains.com
    ns-cloud-c4.googledomains.com

Would not make any sense to do four of them if it's a single AZ. Also, they are geo-aware and routed to your nearest region.

[seabrookmx]

Are you conflating autonomous system (AS) with availability zone (AZ)?

[deepsun]

Uhh, you're right, I totally did. Now I see the parent's point, thank you.

[pocksuppet]

DNS is a centralization risk, yes. Somehow we've decided this is fine. DNSSEC isn't the only issue - your TLD's nameservers could also be offline, or censored in your country.

[skywhopper]

DNS is barely centralized. Is there an alternative global name lookup system that is less centralized without even worse downsides?

[miki123211]

The blockchain.

The only thing a blockchain is good for is achieving decentralized consensus on what value a key points to, which is what DNS is.

An alternative way of looking at this is that acquiring domains must be somewhat expensive by definition; either you enforce it at the system level, or you make it free, but then somebody will inevitably grab all the interesting ones and re-sell them to others. A blockchain is the only way to make decentralized financial infrastructure viable.

[fc417fc802]

GNS is the obvious response here, in addition to the various blockchain based solutions. Nothing that enjoys widespread support or mindshare unfortunately.

Even the current centralized ICANN flavor could be substantially more resilient if it instead handed out key fingerprints and semi-permanent addresses when queried. That way it would only ever need to be used as a fallback when the previously queried information failed to resolve.

[account42]

GP said it was a risk (and it is), not that there are better alternatives. Not all risks can be eliminated easily but you should still be aware of them.

[pocksuppet]

BGP, but the names in question are limited to 128 bits, of which at most 48 will be looked up, and you don't get to choose which 48 bits are assigned to you.

[greatgib]

Normally it should not have been, with cache and all, but that was the past...

Think about what would happen the day that letsencrypt is borken for whatever reason technical or like having a retarded US leader and being located in the wrong country. Taken into account the push of letsencrypt with major web browsers to restrict certificate validities for short periods like only a few days...

[muvlon]

Let's Encrypt has to be down for days before people begin to feel the pain. DNS is very different, it breaks stuff immediately everywhere.

[tharkun__]

No it doesn't. DNS breaks as soon as TTLs run out. It's your choice to set them so low that stuff breaks immediately.

[__float]

What do you recommend then? DNS doesn't usually change that often, but if you mess it up when it does, you're in for some pain if TTLs are high!

[account42]

Unfortunately you can't set DNS TTL arbitrarily high (or low) without some resolvers ignoring your suggestion and using arbitrary values.

[ale42]

This assumes that the host name you want has been recently queried. If it's not cached, good luck...

[cyberax]

Not really? .com and .net are still up

If Let's Encrypt goes down, half of the Internet will become inaccessible in a week.

[akerl_]

Presumably if LetsEncrypt goes down and stays down for a week, the sites that go down are the ones that see that their CA went down and at no point in the week take the option to get certs from a different CA?

[bluejekyll]

I guarantee that there are a ton of sites out there not monitoring their certs.

[throw0101c]

Including Microsoft, Starlink, Github, Cisco:

* https://www.keyfactor.com/blog/2023s-biggest-certificate-out...

[gpvos]

"A ton" being a misspelling of "the vast, vast majority".

[fragmede]

Are there alternative CAs that are anywhere as easy to deal with as Lets encrypt?

[kenniskrag]

acme.sh supports multiple CAs there is even a RFC for CAs that describe the api.

[sllabres]

So it seems we need something like this [1] for IT infrastructure? ;)

[1] https://outerspaceinstitute.ca/crashclock/

[gerdesj]

"The internet's famous "routing around damage" isn't quite working here."

DNS is a look up service that runs on the internet.

Internet routing of IP packets is what the internet does and that is working fine (for a given value of fine).

You remind me of someone using the term "the internet is down" that really means: "I've forgotten my wifi password".

[LastTrain]

Us non pod-people caught his drift.

[eru]

What's a pod-people?

[the8472]

fail-closed protocols have introduced some brittleness. A HTTP 1.0 server from 1999 probably still can service visitors today. A HTTPS/TLS 1.0 server from the same year wouldn't.

[zelon88]

I think I see the point you're making here and I agree.

There is designing something to be fail-closed because it needs to be secure in a physical sense (actually secure, physically protected), and then there's designing something fail-closed because it needs to be secure from an intellectual sense (gatekept, intellectually protected). While most of the internet is "open source" by nature, the complexity has been increased to the point where significant financial and technical investment must be made to even just participate. We've let the gatekeepers raise the gates so high that nobody can reach them. AI will let the gatekeepers keep raising the gates, but then even they won't be able to reach the top. Then what?

I think the point you're trying to make, put another way is in the context of "availability" and "accessibility" we've compromised a lot of both availability and accessibility in the name of security since the dawn of the internet. How much of that security actually benefits the internet, and how much of that security hinders it? How much of it exists as a gatekeeping measure by those who can afford to write the rules?

[account42]

Backwards compatibility is unfortunately not something security folk care about.

[sam_lowry_]

This is why I still run my blog on HTTP/1.1 only.

[account42]

What no HTTP/1.0 for those of us too lazy to type the Host header into telnet???

[sam_lowry_]

Oh, because I host it with a few more sites on my tiny Hetzner cloud server.

[fc417fc802]

You're not wrong but objecting to fail-closed in a security sensitive context is entirely missing the point.

[the8472]

Giving up on robustness is capitulation on a social-technical interface. "users can't be trusted, we can't have any fallbacks, ever" is a problem that needs to be solved.

[walrus01]

It looks like a failed key replacement during a scheduled maintenance event. Normally this sort of thing is thoroughly tested and has multiple eyes on for detailed review and planning before changes get committed, but obviously something got missed.

[wildylion]

.ru [had this](https://habr.com/ru/news/790214/) in 2024. On the other hand, the russian government in itself had been doing a stellar job at breaking the internet.

(I hope I'll live to see them all sentenced to life without parole)

[account42]

Would be interesting to know how something could get missed. You'd think the system was set up so that new keys could not be published without being verified working in a staging system.

[Muromec]

>So a single configuration mistake in a single place wiped out external reachability of a major economy.

And fuck nothing at all happened as a result.

[Our_Benefactors]

Prove it? I’m sure many lifespans were lost to stress

[pinkgolem]

As someone with oncall yesterday it was a fun experience, but you noticed quickly that everything .de was down and then it was just a waiting game.

We had a short discussion about migrating to .com, but decided risk != reward as no one would know the new tld

I assume there are a couple people working for denic who had a stressfull night..

[Woodi]

> So a single configuration mistake in a single place wiped out external reachability of a major economy.

Real world beats sci-fi :) And isn't it why we love IT ? And hate it too, because of "peoples in charge"...

[throw0101c]

> So a single configuration mistake in a single place wiped out external reachability of a major economy.

No different than a bunch of BGP issues we've seen over the years.

And you don't even need DNSSEC for DNS to break things: reminder of the October 2025 AWS outage:

* https://www.akamai.com/blog/security/when-cloud-breaks-lesso...

[lschueller]

I have a bad feeling, that the impact will be quite severe for some services, as monitoring, performance, and security services might get disrupted. and just cleaning up is a big mess.. Worst case, some ot will experience outage and / or damage. But maybe I am just overestimating the severity of this.

[miki123211]

The more interesting question is, could a political adversary do this to a country on purpose, and how hard would that be?

[number6]

There is the kritis (critical infrastructure law) law, which trys to enforce some standards to make things not as brittle.

[otabdeveloper4]

> The internet's famous "routing around damage"

...is only for Pentagon networks and military stuff. It's not for us normal people. (We get Cloudflare and FAANG bullshit instead.)

[zelon88]

This is actually startlingly true.

Every FAANG company has their own fiber backbone. Why invest the internet that everyone uses when you can invest in your own private internet and then sell that instead?

[profmonocle]

It's not like the long-haul fiber not owned by FAANG is a public utility, at least not in most places.

Traffic that goes over "the Internet" traverses some mix of your ISP's fiber, fiber belonging to some other ISP they have a deal with, then fiber belong to some ISP they have a deal with, etc.

All those ISPs are being paid to provide service, they can invest in their own networks.

[account42]

And we all know that ISPs are famous for investing in timely infrastructure upgrades.

[beeforpork]

... wiped out external reachability of a major economy ...

internal reachability (from Germany to .de domains), too... :-)))

[dlopes7]

I love how I work with IT for 20 years and don't understand a single acronym here other than DNSSEC

[icedchai]

I've been in IT 30+ years, been running DNS, web servers, etc. since at least 1994. I haven't bothered with DNSSEC due to perceived operational complexity. The penalty for a screw up, a total outage, just doesn't seem worth the security it provides.

[gerdesj]

That was my experience too until I decided that just running email systems for 30 odd years when HN says that is unnatural piqued my weird or something!

I ran up three new VMs on three different sites. I linked all three systems via a private Wireguard mesh. MariaDB on each VM bound to the wg IP and stock replication from the "primary". PowerDNS runs across that lot. One of the VMs is not available from the internet and has no identity within the DNS. The idea is that if the Eye of Sauron bears down on me, I can bring another DNS server online quite quickly and fiddle the records to bring it online. It also serves as a third authority for replication.

I also deployed https://github.com/PowerDNS-Admin/PowerDNS-Admin which is getting on a bit and will be replaced eventually but works beautifully.

Now I have DNS with DNSSEC and dynamic DNS and all the rest. This is how you start signing a zone and PowerDNS will look after everything else:

  # pdnsutil secure-zone example.co.uk
  # pdnsutil zone set-nsec3 example.co.uk
  # pdnsutil zone rectify example.co.uk

Grab a test zone and work it all out first, it will cost you not a lot and then go for "production".

My home systems are DNSSEC signed.

[qingcharles]

How simple sysadmin was in 1994 with no cryptography on any protocol. Everything could be easily MITM'd. Your credit card number would get jacked left and right in the 90s.

[icedchai]

Nobody was taking credit cards online then. Your telnet sessions were easily sniffed, however.

[qingcharles]

Not in '94, sure. But a couple of years later it was common and SSL was still uncommon, for a bunch of reasons, and also everyone was storing the card numbers in plaintext on their servers too.

Telnet was sniffed. IRC was being sniffed and logged.

[icedchai]

Yes, I worked on some early ecommerce sites. Often, we'd accept credit cards with SSL and then send them out with email (plain text SMTP) to the customer, for manual entry. Very secure.

[account42]

And your mailman can also just open your letters. So what, it mostly doesn't happen in developed countries. Not everything needs an airtight technical solution, we have way less costly ways to deal with unwanted behavior.

[gerdesj]

Cool. Feel free to explain how to tighten things up.

I've just given them part of a recipe for using DNSSEC. I suspect you are not actually human .. qingcharles.

[qingcharles]

I don't even understand what your comment is about, my dude. Given who a recipe? DENIC?

[icedchai]

Look at his previous post. He described how to set up DNSSEC with PowerDNS.

[walrus01]

To be fair, advanced real world knowledge of public/private key PKIs (x.509 or other), things like root CAs, are a fairly esoteric and very specialized field of study. There's people whose regular day jobs are nothing but doing stuff with PKI infrastructure and their depth of knowledge on many other non-PKI subjects is probably surface level only.

[hannob]

I know quite a bit about PKI and X.509, and I can tell you that much: the overlap with how DNSSEC works is limited.

[silisili]

As is the overlap between DNSSEC and DNS itself, to be honest.

I once worked at the level of administering DNSSEC for 300+ TLDs. It's its own world. When that company was winding down, I tried to continue in the field but the most common response (outside of no response, of course), was 'we already have a DNS team/vendor/guy.' And well, then things like this happen. I won't throw stones though, it's a lot to learn and can be incredibly brittle.

[mschuster91]

It's not made easier by the fact that a lot of cryptography is either very old and arcane or it's one hell of a mess of code that doesn't make sense without reading standards.

I had the misfortune of having to dig deep into constructing ASN.1 payloads by hand [1] because that's the only thing Java speaks, and oh holy hell is this A MESS because OF COURSE there's two ways to encode a bunch of bytes (BIT STRING vs OCTET STRING) and encoding ed25519 keys uses BOTH [2].

And ed25519 is a mess in itself. The more-or-less standard implementation by orlp [3] is almost completely lacking any comments explaining what is going on where and reading the relevant RFCs alone doesn't help, it's probably only understandable by reading a 500 pages math paper.

It's almost as if cryptographers have zero interest in interested random people to join the field.

End of rant.

[1] https://github.com/msmuenchen/meshcore-packets-java/blob/mai...

[2] https://datatracker.ietf.org/doc/html/rfc8410#appendix-A

[3] https://github.com/orlp/ed25519/tree/master

[Muromec]

The trick to asn.1 is to generate both parser and serializer from the spec. Elliptic curve math on the other hand is ... yeah, you need to know the math and also know the tricks to code that implements it. Both of those have steep learning curve, but it's hardly because it's a mess or it's old.

[thayne]

The problem with ASN.1 is that it is big and complicated, and you only need a fraction of it for cryptography, and it isn't really used for anything outside of pki anymore.

It wouldn't be as bad if asn.1 had cought on more as a general purpose serialization format and there were ubiquitous decent libraries for dealing with it. But that didn't happen. Probably partly because there are so many different representations of asn.1.

A bespoke serialization specifically for certificates might actually have aged better, if it was well designed.

[jll29]

Assuming there are some libraries for it, would this make a pretty good case for LLM-generated ports of these existing libraries into other languages or onto other OSs/platforms? One implementation could be treated as "the spect".

[pocksuppet]

ASN.1 is protobufs designed by committee. It is a general-purpose serialization format, but there's no good reason to choose it instead of protobufs.

[mschuster91]

> Both of those have steep learning curve, but it's hardly because it's a mess or it's old.

Bitpacking structures used to be important in the 60s. That time has passed, unless you're dealing with LoRa, NFC or other cases of highly constrained bandwidth there are way better options to serialize and deserialize information. It's time to move on, and the complexity of all the legacy garbage in crypto has been the case of many a security vulnerability in the past.

As for the code, it might be personal preference but I'd love to have at least some comments referring back to a specification or original research paper in the code.

[Muromec]

I think you misunderstand the problem asn.1 solves and constrains it works within (both 30 years ago and now). We sure can have a better one now once we learned all the lessons and know what good parts to keep, but this critique of bitpacking is misplaced.

[Avamander]

ASN.1 is not used because of just bitpacking. There are other benefits to ASN.1 and it's probably one of the least problematic parts there.

People who have thought they can do better have made things like PGP. It's one of the worst cryptographic solutions out there. You're free to try as well though.

[tptacek]

The trick to ASN.1 is to serialize/unserialize it backwards.

[dwattttt]

#1 NSA, I get it now!

[cyberax]

X.509 is a deep legacy, but at least at this point it's well tested.

> because that's the only thing Java speaks

No, it most definitely is not. You can just construct a private key directly in BouncyCastle: https://downloads.bouncycastle.org/java/docs/bcprov-jdk18on-...

I'm 100% certain that you also can do that with raw java.security. I did that about 15 years ago with raw RSA/EC keys. You can just directly specify the private exponent for RSA (as a bigint!) or the curve point for EC.

Ditto for ed25519, you can just take the canonical implementation from DJB. And you really really shouldn't do that anyway, please just use OpenSSL or another similar major crypto library.

[mschuster91]

> I'm 100% certain that you also can do that with raw java.security.

I tried that, the problem is Meshcore specific - they do their own weird shit with private and public keys [1]. Haven't figured out how to do the private key import either, because in the C source code (or in python re-implementations) Meshcore just calls directly into the raw ed25519 library to do their custom math... it's a mess.

[1] https://jacksbrain.com/2026/01/a-hitchhiker-s-guide-to-meshc...

[cyberax]

I'm playing with LORA/Meshcore right now (I have an nRF52840 lying around). I'm pretty sure I know how to do that, will take a look.

[Muromec]

I wouldn't recommend touching openssl (the library, command line tools are okay-ish) with anything that breaths life.

[tptacek]

The typical vector for entering cryptography as a professional is called "grad school".

[hathawsh]

Is that actually true, though? Even though it's not really my job, I find myself debugging certificates and keys at least once a month, and that's after automating as much as possible with certbot and cloud certificates. PKI always seems to demand attention.

[walrus01]

In my initial comment, I meant more in terms of complexity and planning from the perspective of the people who are running the public/private key infrastructure on the other side/upstream of what you're doing as a letsencrypt end user.

Broadly similar general concept to the team responsible for the DNSSSEC signing keys for an entire ccTLD.

Yeah a x509 PKI / root CA is a very different thing than DNSSSEC but they have a number of general logical similarities in that the chain of trust ultimately comes down to a "do not fuck this up" single point of failure.

[bflesch]

Don't worry, that's by design ;)

[1vuio0pswjnm7]

Curious how this comment was originally marked as "dead" but is now on top

I saw it at bottom of thread and vouched for it. Usually when I "vouch" nothing happens

I always ignore the RRSIG lines in zone files. To me it's not "DNS data", it's cruft

But DNSSEC has its true believers. I'm just not one of them