Categories
Admin Network Technologies

Monitoring by Zabbix: a working document

Intro
I panned Zabbix in this post: DIY monitoring. But I have compelling reasons to revisit it. I have to say it has matured, but there remain some very frustrating things about it, especially when compared with SiteScope (now owned by Microfocus) which is so much more intuitive.

But I am impressed by the breadth of the user base and the documentation. But learning how to do any specific thing is still an exercise in futility.

I am going to try to structure this post as a problems encountered, and how they were resolved.

Current production version as of this writing?
Answer: 5.0

Zabbix Manual does not work in Firefox
That’s right. I can’t even read the manual in my version of Firefox. Its sections do not expand. Solution: use Chrome

Which database?
You may see references to MYSQL in Zabbix docs. MYSQL is basically dead. what should you do?

Zabbix quick install on Redhat

Answer
Install mariadb which has replaced MYSQL and supports the same commands such as the mysql from the screenshot. On my Redhat instance I have installed these mariadb-related repositories:

mariadb-5.5.64-1.el7.x86_64
mariadb-server-5.5.64-1.el7.x86_64
mariadb-libs-5.5.64-1.el7.x86_64

Terminology confusion
what is a host, a host group, a template, an item, a web scenario, a trigger, a media type?
Answer

Don’t ask me. When I make progress I’ll post it here.

Web scenario specific issues
Can different web scenarios use different proxies?
Answer: Yes, no problem. In really old versions this was not possible. See web scenario screenshot below.

Can the proxy be a variable so that the same web scenario can be used for different proxies?
Answer: Yes. Let’s say you attach a web scenario to a host. In that host’s configuration you can define a “macro” which sets the variable value. e.g., the value of HTTP_PROXY in my example. I think you can do the same from a template, but I’m getting ahead of myself.

Similarly, can you do basic proxy auth and hide the credentials in a MACRO? Answer: I think so. I did it once at any rate. See above screenshot.

Why does my google.com web scenario work whereas my amazon.com scenario not when they’re exactly the same except for the URL?
Answer: some ideas, but the logging information is bad. Amazon does not take to bots hitting it for health check reasons. It may work better to change the agent type to Linux|Chrome, which is what I am trying now. Here’s my original answer: Even with command-line curl I get an error through this proxy. That can’t be good:
$ curl ‐vikL www.amazon.com

...
NSS error -5961 (PR_CONNECT_RESET_ERROR)
* TCP connection reset by peer
* Closing connection 1
curl: (35) TCP connection reset by peer


My amazon.com web scenario is not working (status of 1), yet in dashboard does not return any obvious warning or error or red color. Why? Answer:
no idea. Maybe you have to define a trigger?

Say you’re on the Monitoring|latest data screen. Does the data get auto-updated? Answer: yes, it seems to refresh every 30 seconds.

Why is the official FAQ so useless? Answer: no idea how a piece of software otherwise so feature-rich could have such a useless FAQ.

Zabbix costs nothing. Is it still actively supported? Answer: it seems very actively supported for some reason. Not sure what the revenue model is, however…

Can I force one or more web scenarios to be run immediately? I do this all the time in SiteScope. Answer: I guess not. There is no obvious way.

Suppose you have defined an item. what is the item key? Answer: You define it. Best to make it unique and use contiguous characters. I’m seeing it’s very important…

What is the equivalent to SiteScope’s script monitors? Answer: Either ssh check or external check.

How would you set up a simple PING monitor, i.e., to see if your host is up? Answer: Create an item as a “simple check”, e.g., with the name ping this host, and the key icmpping[{HOST.IP},3]. That can go into a template, by the way. If it succeeded it will return a 1.

I’ve made an error in my script for an external check. Why does Latest data show nothing at all? Answer: no idea. If the error is bad enough Zabbix will disable the item on you, so it’s not really running any longer. But even when it doesn’t do that, a lot of times I simply see no output whatsoever. Very frustrating.

Help! The Latest Data graph’s Y axis only shows 0’s and 5’s. Answer: Another wonderful Zabbix feature, this happens because your Units are too long. Even “per minute” as Units can get you into trouble if it is trying to draw a Y axis with values 22.0 22.5 23.0, etc: you’ll only see the .0’s and the .5’s. Change units to a maximum of seven characters such as “per min.”

Why is the output from an ssh check truncated, where does the rest go? Answer: no idea.

How do you increase the information contained in the zabbix server log? Let’s say your zabbix server is running normally. Then run this command: zabbix_server ‐R log_level_increase
You can run it multiple times to keep increasing the verbosity (log level), I think.

Attempting to use ssh items with key authentication fails with :”Public key authentication failed: Callback returned error” Initially I thought Zabbix was broken with regards to ssh public key authentication. I can get it to work with password. I can use my public/private key to authenticate by hand from command-line as root. Turns out running command such as sudo -u zabbix ssh … showed that my zabbix account did not have permissions to write to its home directory (which did not even exist). I guess this is a case of RTFM, because they do go over all those steps in the manual. I fixed up permissions and now it works for me, yeah.

Where should the scripts for external checks go? In my install it is /usr/lib/zabbix/externalscripts.

Why is the behaviour of triggers inconsistent. sometimes the same trigger has expected behaviour, sometimes not. Answer: No idea. Very frustrating. See more on that topic below.

How do you force a web scenario check when you are using templates? Answer: No idea.

Why do (resolved) Problems disappear no matter how you search for them if they are older than, say, 30 minutes? Answer: No idea. Just another stupid feature I guess.

Why does it say No media defined for user even though user has been set up with email as his media? Answer: no idea.

Why do too many errors disable an ssh check so that you get Status Disabled and have no graceful way to recover? Answer: no idea. It makes sense that Zabbix should not subject itself to too many consecutive errors. But once you’ve fixed the underlying problem the only recovery I can figure is to delete the item and recreate it. or delete the host and re-create it. Not cool.

I heard dependent items are the way to go to parse complex data coming out of a rich text item. How do you do that? Answer: Yes they are. I have gotten them to work and really give me the fine-grained control I’ve always wanted. I hope to show a real-life example soon. To get started creating a dependent item you can right-click on the dots of an item, or create a new item and choose type Dependent Item.

I am looking at Latest data and one item is grayed out and has no data. Why? Answer: almost no idea. This happens to me in a dependent item formed by a regular expression where the regular expression does not match the content. I am trying to make my RegEx more flexible to match both good and error conditions.

Why do my dependent items, when running a Check Now, say Cannot send request: wrong data type, yet they are producing data just fine when viewed through Latest data? Answer: this happens if you ran a Check Now on your template rather than when viewing an individual host. Make sure you select a host before you run Check Now. Actually, even still it does not work, so final answer: no idea.

Why do some regular expressions check out just fine on regex101.com yet produce a match for value of type “string”: pattern does not match error in Zabbix? Answer: Some idea. Fancy regular expressions do not seem to work for some reason.

Every time I add an item it takes the absolute maximum amount of time before I see data, whether or not I run check Now until I turn blue in the face. Why? Answer: no idea. Very frustrating.

If the Zabbix server is in one time zone and I am in another, can I have my view of timestamps customized to my time zone? Otherwise I see all times in the timezone of the Zabbix server. Answer: You are out of luck. The suggestion is to run two GUIs, one in your time zone. But there is a but. Support for this has been announced for v 5.20. Stay tuned…

My DNS queries using net.dns don’t do anything. Why? Answer: no idea. Maybe your host is not running an actual Zabbiox agent? That’ll do it. Forget that net.dns check if you can’t install an agent. Zabbix has no agentless DNS monitor for some strange reason.

A DNS query which returns many address records fails (such as querying an AD domain), though occasionally succeeds. Why? Answer: So your key looks something like this, right? net.dns.record[10.1.2.3,my-AD-domain.net,A,10,2,tcp]. And when you do the query through dig it works fine, right? E.g., dig +tcp my-AD-domain.net @10.1.2.3. And you’ve set the Zabbix response to type text? It seems to be just another Zabbix bug. You may have to use a script instead. Zabbix support has been able to reproduce this bug and they are working on it as we speak.

What does Check/Execute Now really do? Answer: essentially nothing as far as I tell. It certainly doesn’t “check now”, i.e., force the item to be run. However, if you have enough permissions, what you can do when you’re looking at an item for a specific Host is to run Test. Then Get Value. I usually get Permission Denied, however.

I want to show multiple things on a dashboard widget graph like an item plus its baseline (Ed: see references for calculating a baseline). What’s the best way? Answer: You can use the add new data set feature for instance to add your baseline. In your additional data set you put your baselines. Then I like to make the width 2, transparency 0 and fill 0. This will turn it into a thin bold line with a complementary color while not messing too much with the original colors of your items. The interface is squirrely, but, hey, it’s Zabbix, what did you expect?

I have a lot of hosts I want to add to a template. Does that Mass Update feature actually work? Answer: yes. Use it. It will save you time.

Help! I accidentally deleted an entire template. I meant to just delete one of its macros. Is there a revert? Answer: it doesn’t look like it. Hope you remember what you did…

It seems if I choose units in an item which have too many characters, e.g., client connections, the graph (in Latest Data) cuts it off and doesn’t even display the scale? Answer: seems so. It’s a bug. This won’t happen when using vector graphs in Dashboard. The graphs in Latest Data are PNG and limited to short Units, e.g., mbps. Changing to vector graphs has been in the roadmap but then disappeared.

Can I create a baseline? Nope. It’s on the roadmap. However, see this clever idea for building one on your own without too much effort.

I’ve put a few things on the same Dashboard graph. Why don’t they align? There are these big gaps. Zabbix runs the items when it feels like, and the result is gaps in data which Zabbix makes no attempt to conceal at the beginning and end of a graph. You can use Scheduling Intervals on your items to gain some control over this. See this article for details.

Besides cloning the whole thing, how can I change the name of a Dashboard? Answer: If you just click to edit a Dashboard the name appears fixed. However, click on the gear icon and that gives you the option to edit the dashboard name. It’s kind of an undocumented feature.

My SNMP MIB has bytes in/out for an interface when what I really want is bandwidth, i.e., Megabits per second. A little preprocessing on a 64-bit bytes value and you are there (32 bit values may roll over too frequently). See this article for details.

In functions like avg (sec|#num,<time_shift>), why is the time_shift argument so restricted? It can’t be a macro, contain a formula like 1w-30m, or anything semi-sophisticated. It just accepts a dumb literal like 5h? Answer: It’s just another shortcoming in Zabbix. How much did you pay for it? 🙂

I have an SNMP template with items for a hostgroup of dispersed servers. Some work fine. The one in Asia returns a few values, but not all. I am using Bulk Request. Answer (to your implied question!) You must have bad performance to that one. Use a Zabbix proxy with a longer timeout for SNMP requests. I was in that situation and that worked for me.

Riddle: when is 80% not 80%? Answer: when pulling in used storage on a filesystem via SNMP and comparing it to storage size! I had carefully gotten a filesystem 83% full based on the output of df -m. But my trigger, set to go off at 80%, never went off. How could it be? The 83% includes some kind of reserved user space on the filesystem which is not included when you do the calculation directly. So I was at 78% or so in actuality. I changed the trigger to 75%.

My trigger for a DNS item, which relies on a simple diff(), goes off from time-to-time yet the response is the same. Why? Answer: We have seen this behavior for a CNAME DNS item. The response changed the case of the returned FQDN from time-to-time, and that is enough to set off the Zabbix diff()-based trigger! We pre-processed the output with a RegEx to just get the bits we wanted to examine to fix this.

Related question. My diff() trigger for a DNS item does NOT go off when the server actually goes down. What’s up with that? Answer: Although you might expect a suddenly unavailable server constitutes a “difference,” in Zabbix’s contorted view of reality it does not. I recommend an additional trigger using the function nodata().

Does the new feature of login using SAML actually work? Answer: Yes, we are using it in Zabbix v 5.0.

My OIDs for my filesystems keeps shifting around. What to do? Answer: Use low-level discovery. It’s yet another layer of abstraction and confusion, but it’s probably worth it. I intend to write up my approach in my practical Zabbix examples blog post.

After an Zabbix agent item goes bad (no data), Zabbix refuses to test it for a full 30 minutes after it went bad, despite an update interval of 5 minutes. Why? Answer: In one of the worst architectural decisions of all time, Zabbix created the concept of unsupported items. It works something like this: the very moment when you need to be told Hey there’s something wrong here is when Zabbix goes quiet. Your item became unsupported, which is like being in the penalty box for 30 minutes, during which time nothing works like you naively expected it to. Even the fact that your item became unsupported is almost impossible to find out from a trigger. An example of software which treats this situation correctly is Microfocus SiteScope. In Zabbix in version 5.0 there’s a global timeout for all unsupported items. Ours is set to 30 minutes, you see. In some cases that may make sense and prevent Zabbix from consuming too many resources trying to measure things which don’t work. I find it annoying. For DNS, specifically, best to use a key of type net.dns and not net.dns.record. That returns a simple 0 or 1 and does not become unsupported if the dns server can’t be reached. V 5.2 will provide some more options around this issue. For a HTTP agent and I suppose many other items, it’s best to create triggers which use the nodata() function,  which can somewhat compensate for this glaring weakness in Zabbix. If you run Zabbix v 5.2, you should use the new preprocessing rule “Check for not supported value” and then set new value e.g. “Error”. Then the Item won’t become unsupported and can also be used for triggers.

I wish to use a DNS value instead of an IP in net.tcp.service[service,IP,port] because I use geoDNS or round-robin DNS. Can I? Answer: It seems to work, yes.

Can I send alerts to MS teams? Answer: This is obviously a fake question. But the answer is Yes. You set up a Connector in a MS Teams channel. It’s pretty staight forward and it’s pretty cool. I’ll try to publish more in my Zabbix tips post if I have time.

Get a lot of false positives? Answer: Yes! On F5 equipment this one is vexing me:

Resolved: BIG-IP is unreachable via SNMP for 15 minutes

And for others (pool member unavailable for a few minutes) I tried to require two consecutive failures before sending an alert. Basically still working on it.

I have a bunch of HTTP items on this one Zabbix proxy. They all sort of go bad at the same time (false positives) and Zabbix says this agent is unreachable for five minutes around the same time. Answer: Seen that. Short term it may be advisable to create a dependent trigger: https://www.zabbix.com/documentation/5.0/manual/config/triggers/dependencies  Mid-term I am going to ask support about this problem.

Why is the name field truncated in Monitoring | Latest Data, with no possibility to increase it? Answer: If you have Show Details selected you see very few characters. Deselect that.

What, Zabbix version 5.2 RPMs are not available for RHEL 7? Answer: that is correct, unfortunately, as of this writing. You can run as high as v 5.0.7. We are trying to pressure them to provide this compatibility. Lots of people still run Redhat v 7.

Can you send reminder alerts periodically for a problem which persists? Answer: Yes you can. For instance, every four hours. Read all about it in the manual, under Action | Escalations, and look at their examples.

Is Zabbix affected by the same hack that infected SolarWinds? Answer: No idea. Let’s see. Developed in Eastern Europe. Basically, no one’s saying. Let’s hope not.

Is Zabbix stupid enough to send multiple alerts for the same problem? Answer: In a word, yes. If you are unlucky enough to have defined overlapping alert conditions in your various alerts, Zabbix will make no effort to consolidate them.

What does it mean when I look at a host and I see inaccessible template? Answer: Most likely explanation is that you don’t have permission to see that template.

Can the y-axis be drawn in a logarithmic scale in a dashboard graph? I have low values (time for a DNS query) which sometimes soar to high ones. Answer: No. This feature has been requested now for almost 10 years and still is lacking. I will try to make a feature request.

A word about SSH checks and triggers
Through the school of hard knocks I have learned that my ssh check is clipping the output from the executed command. So you know that partial data you see when you look at latest data, and thought it was truncating it for display purposes? Nuh, ah. That’s all you’re getting to go up against in your trigger, which sucks. It’s something like 260 characters. I got lucky in a sense to discover this early by running an ssh check against dns resolution of amazon.com. The response I got varied almost every 60 seconds depending on whether or not the response came out of the dns cache. So this was an excellent testbed to learn about the flakiness of triggers as well as waste an entire day.

Another thing about triggers with a regex. As far as I can tell the logic is reversed. So you think you’re defining the OK condition when you seek to match the output and have it given the value of 1. But instead try to match the desired output for the OK condition, but assign it a value of 0. I guess. Only that approach seems to work for me. And getting the regex to treat multiple lines as a unit was also a little tricky. I think by default it favored testing only against the last line.

So let’s say my output as scraped from Monitoring|Latest Data alternated between either

proxy1&gt;test dns amazon.com
Performing DNS lookup for: amazon.com
 
DNS Response data:
Official Host Name: amazon.com
Resolved Addresses:
  205.251.242.103
  176.32.98.166
  176.32.103.205
Cache TTL: 1, cache HIT
DNS Resolver Response: Success

or

proxy1&gt;test dns amazon.com
Performing DNS lookup for: amazon.com
 
Sending A query for amazon.com to 192.168.135.145.
 
Sending A query for amazon.com to 8.8.8.8.
 
DNS Response data:
Official Host Name: amazon.com
Resolved Addresses:
  20

, then here is my iregexp expression which seems to do the correct thing (treat both of these outcomes as successes):

{proxy1:ssh.run[resolve DNS,1.2.3.4,22,utf-8].iregexp("(?s)((205\.251\.|176\.32\.)|Sending A query.+\s20)")}=0

Note that the (?s) at the beginning helps, I think, to treat the newline character as just another character which matches “.”. I may have an extra set of parentheses around the outermost alternating expression, but I can only experiment so much…

I ran various tests such as to change just one of the numbers to make sure it triggered.

I now think I will get better, i.e., more complete, results if I make the item of type text rather than character, at least that switch definitely helped with another truncated output I was getting from another ssh check. So, yes, now I am capturing all the output. So, note to self, use type text unless you have really brief output from your ssh check.

So with all that gained knowledge, my simplified expression now reads like this:

{proxy:ssh.run[resolve a dns name,1.2.3.4,22,utf-8].iregexp("(205\.251\.|176\.32\.)")}=0

Here’s a CPU trigger. From a show status it focuses on the line:

CPU utilization: 29%

and so if I want to trigger a problem for 95% or higher CPU, this expression works for me:

CPU utilization:\s+([ 1-8]\d|9[0-4])\%

A nice online regular expression checker is https://regexr.com/

And a very simple PING test ssh check item, where the expected resulting line will be:

5 packets transmitted, 5 packets received, 0% packet loss

– for that I used the item wizard, altered what it came up with, and arrived at this:

(({proxy:ssh.run[ping 8.8.8.8,1.2.3.4,22,utf-8].iregexp("[45] packets received")})=0)

So I will accept the results as OK as long as at most one of five packets was dropped.

A lesson learned from SNMP monitoring of F5 devices
My F5 BigIP devices began producing problems as soon as we set up the SNMP monitoring. Something like this:

Node /Common/drj-10_1_2_3 is not available in some capacity: blue (4)

It never seemed to matter until now that my nodes appear blue. But perhaps SNMP is enforcing a best practice and expecting nodes to not be blue, meaning to be monitored. And it turns out you can set up a default monitor for your nodes (I use gateway_icmp). It’s found in Nodes | Default Monitor. I’m not sure why this is not better documented by F5. After this, many legacy nodes turn red so I am cleaning them up… But my conclusion is that I have learned something about my own systems from the act of implementing this monitoring, and that’s a good thing.

To be continued…

References and related
A good commercial solution for infrastructure monitoring: Microfocus SiteScope.

DIY monitoring

The Zabbix manual

A nice online regular expression (RegEx) checker is: https://RegEx101.com/.

Another online regular expression checker is: https://RegExr.com/.

Just to put it out there: If you like Zabbix you may also like Specto. Specto is an open-source tool for monitoring web sites (“synthetic” monitoring). I know one major organization which uses it so it can’t be too bad. https://specto.sourceforge.net/

Since this document is such a mess I’m starting to document some of my interesting items and Zabbix tips in this newer and cleaner post. It includes the baseline calculation formula.

Categories
Admin

The IT Detective Agency: New AIX Server working really slowly

Intro
Don’t ask me why anyone would willingly run IBM AIX, but it happens. And when they do, watch out for network punishment. We dealt with such a case, unfortunately, and we ran into a serious, somewhat obscure network issue and figured out the solution (we think). Maybe someone else can learn from this painful experience. Or maybe we’ll completely forget what we ourselves have done two years from now and find ourselves stepping on the same rake.

The details

So this new AIX server was configured to run an very old application, WebMethods, that makes a lot of database connections as well as connections to external partners for document exchange.

This had been working fine on the old AIX server, but we switched to newer hardware. As much as possible the old configurations were used. Yet this new server just couldn’t keep up with the load. Its queue started building up, connections to the database climbed into the hundreds, and then it just seemed like it was doing nothing at all.

Someone lent me root access so I can join the debugging party. What, no bash shell! Not even a properly configured korn shell. And everything’s just a little different on AIX – nothing is quite how you are accustomed to it. But at least it has tcpdump. I guess they also have their own AIXish utility as well, but I never bothered with that. tcpdump seemed to work. So I quickly began to get a feel from what the application folks were saying about their transfers which weren’t going outbound, and only slowly going inbound. They used port 5443 on one of the interfaces, en5:

# tcpdump -i en5 -n port 5443

And, true, not much was going on.

This went on for a day and things were looking desperate – to the point where we decided to go back to the old hardware! But we never stopped thinking.

Check the traffic to the Oracle database:

# tcpdump -i en0 -n port 1521

No, not that much either.

Try to check system logs, but who knows where those are? The ones I found had absolutely nothing of interest.

Being a DNS guy, I decided to check for DNS traffic:

# tcpdump -i en0 -n udp

(everything else uses tcp so I could get away with this).

Now DNS turned out to be quite chatty – around a dozen entries per second. And a lot of repetition. And a lot of IPv6 queries, labelled as AAAA?. I didn’t like it.

And this jogged my memory. I remembered encountering these IPv6 queries and wanting to turn them off on the old AIX servers. But how to do that??

As in all things, Google (actually DuckDuckGo) is your friend. You modify the /etc/netsvc.conf file. You need an entry like this:

hosts = local4, bind4

To be continued…

Categories
DNS IT Operational Excellence Network Technologies Uncategorized

Google’s DNS Servers Rock!

Intro
DNS is the Domain name Service, the Internet service that converts IP addresses, e.g., 200.54.129.57 into mnemonic names like www.mysite.com.

I tried to run a cache-only DNS server for use by a proxy server. What I found is that certain sites were not accessible on a frequent basis. I think uol.com.br is one of the problem sites (need to check this). It may not mean much to a US audience, but it’s really popular in Brazil!

At some point I happened to learn that Google has a public DNS service. This is worth pondering. No one of any repute has offered a DNS service to that point. There are a host of concerns about security, especially DNS cache poisoning. They blazed a trail, and did it in a way only Google and very few other major infrastructure players could. Not only did they offer a DNS service, they put their DNS servers all over the Internet and created convenient anycast addresses for their servers.

I am no expert on anycast addresses. You can look it up on Wikipedia, however. The essence for my purposes is that with a single IP address you’re going to hit the closest server, network-wise. So no matter where you are some Google DNS server is not far away. Try it. The anycast addresses are 8.8.8.8 and 8.8.4.4. They don’t mind, really! You can ping them. Traceroute to them, whatever. From the Amazon cloud Northeast 8.8.8.8 responds to PINGs in 3.4 ms. That’s really low. Not so low as to make me think they are in the same data center (it is different companies after all), but not far away.

The gold standard for running a DNS service is BIND. I have been running it for many years now and I want to give the Internet Software Consortium their due for providing this wonderful application. Once I got wind of my DNS difficulties as mentioned above, I had to wonder why not everyone else was complaining? They had to be using something else. I ran a flat-out performance test. 5000 queries from an actual proxy log, fed straight to my BIND DNS server, and then to Google’s DNS server 8.8.8.8. I have to dig up the numbers, but Google’s won by quite a bit! This result was actually surprising because you’re always going off-site to the Google DNS server, whereas my server can build up its cache and is right on my network. From where I tested the Google server was about 11 ms away. So 5000 x 11 ms = 55 s. So there is a 55 s handicap from just network considerations alone! Yet it is faster. On the quickest of queries the local server is indeed faster, but what happens is that over the course of real life queries, you always get a few problematic ones which either time out or just seem to take a long time to get back a response. That’s what kills the traditional DNS server and where Google has (obviously) made some optimizations.

And, that’s not all! Google also deals in a more forgiving fashion with broken domain names. I used to get on my high horse and proclaim to others about how broken their DNS servers are – it’s no wonder I can’t resolve their names, which means, by the way, I also cannot get to their web site nor send them email!

It’s effectively like taking yourself off the Internet, or so I thought. Turns out in some cases that’s only true if you’ve constrained yourself to resolving names with BIND. You see, BIND enforces the rules. And I’m a believer in rules. The Internet has about 5,000 technical rules called RFCs. DNS is a topic of many of these rules. The Internet could only have expanded to the size it currently has because all the major players agreed to abide by those rules. What Google has done with their server, in effect, is to say, “Well, if you don’t follow the rules, we’re going to try to work with you anyways.”

Here’s a concrete example. appliedcoatings.org. I guess at some point they’ll actually fix their severely broken DNS, but at the time I write this, August 21, 2011, these comments are valid and their domain is severely broken. In fact, I was amazed that people weren’t jumping up and down screaming at them. I couldn’t even send an email to them. That’s akin to knocking yourself off the Internet, right? Ah, but it all depends on whose DNS servers you are using!

There used to be lots of good free DNS analyzers, like dnsreport.com. You can still find a few around. www.zonecheck.fr, for instance. It shows FAILURE. If it were better written it would show the real problem, which is a lame delegation. But we’re experts, and we don’t need such tools! We will do the queries ourselves and show the lame delegation. We start by learning who are the authoritative nameservers for .ca, the top-level domain used in Canada:

 dig ns ca

; <<>> DiG 9.7.1-P2 <<>> ns ca
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 52928
;; flags: qr rd ra; QUERY: 1, ANSWER: 10, AUTHORITY: 0, ADDITIONAL: 1

;; QUESTION SECTION:
;ca.                            IN      NS

;; ANSWER SECTION:
ca.                     83585   IN      NS      a.ca-servers.ca.
ca.                     83585   IN      NS      c.ca-servers.ca.
ca.                     83585   IN      NS      e.ca-servers.ca.
ca.                     83585   IN      NS      f.ca-servers.ca.
ca.                     83585   IN      NS      j.ca-servers.ca.
ca.                     83585   IN      NS      k.ca-servers.ca.
ca.                     83585   IN      NS      l.ca-servers.ca.
ca.                     83585   IN      NS      m.ca-servers.ca.
ca.                     83585   IN      NS      z.ca-servers.ca.
ca.                     83585   IN      NS      sns-pb.isc.org.

;; ADDITIONAL SECTION:
a.ca-servers.ca.        83594   IN      A       192.228.27.11

Now we ask one of them about the nameservers for appliedcoatings.ca:

 dig ns appliedcoatings.ca @a.ca-servers.ca.

; <<>> DiG 9.7.1-P2 <<>> ns appliedcoatings.ca @a.ca-servers.ca.
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 288
;; flags: qr rd; QUERY: 1, ANSWER: 0, AUTHORITY: 2, ADDITIONAL: 0
;; WARNING: recursion requested but not available

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      NS

;; AUTHORITY SECTION:
appliedcoatings.ca.     86400   IN      NS      sp2.domainpeople.com.
appliedcoatings.ca.     86400   IN      NS      sp1.domainpeople.com.

So far everything's cool. Now, since the authoritative flag (AA) was not present in that response we re-ask that query, but now to one of the nameservers that's supposed to be authoritative for that domain:

dig ns appliedcoatings.ca @sp2.domainpeople.com.

; <<>> DiG 9.7.1-P2 <<>> ns appliedcoatings.ca @sp2.domainpeople.com.
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 24373
;; flags: qr aa rd; QUERY: 1, ANSWER: 2, AUTHORITY: 0, ADDITIONAL: 0
;; WARNING: recursion requested but not available

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      NS

;; ANSWER SECTION:
appliedcoatings.ca.     86400   IN      NS      ns1.domainpeople.com.
appliedcoatings.ca.     86400   IN      NS      ns2.domainpeople.com.

Oh, oh. That's not supposed to happen. We're getting back an entirely different set of nameservers. That's a lame delegation. The domain should be considered completely broken. I think even BIND might be forgiving up to this point. a BIND resolver does these types of quesires to get at the answer. At this point it says, "OK, this is strange, but not necessariily fatal. I will ask my subsequent queries to ns1.domainpeople.com and ns2.domainpeople.com since they are listed as being the nameservers of record.

So now let's get to something useful: looking up the mail exchanger record so we see how to deliver mail to this domain. BIND, which has been fastidiously following the rules, does it as follows:

dig mx appliedcoatings.ca @ns1.domainpeople.com.

; <<>> DiG 9.7.1-P2 <<>> mx appliedcoatings.ca @ns1.domainpeople.com.
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: REFUSED, id: 49996
;; flags: qr rd; QUERY: 1, ANSWER: 0, AUTHORITY: 0, ADDITIONAL: 0
;; WARNING: recursion requested but not available

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      MX

;; Query time: 79 msec
;; SERVER: 204.174.223.72#53(204.174.223.72)
;; WHEN: Sun Aug 21 19:05:43 2011
;; MSG SIZE  rcvd: 36

That's not good. Status is REFUSED. But BIND can even forgive this slight. There is one more nameserver to try after all, right? Last chance query:

dig mx appliedcoatings.ca @ns2.domainpeople.com.

; <<>> DiG 9.7.1-P2 <<>> mx appliedcoatings.ca @ns2.domainpeople.com.
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: REFUSED, id: 44404
;; flags: qr rd; QUERY: 1, ANSWER: 0, AUTHORITY: 0, ADDITIONAL: 0
;; WARNING: recursion requested but not available

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      MX

;; Query time: 72 msec
;; SERVER: 64.40.96.140#53(64.40.96.140)
;; WHEN: Sun Aug 21 19:07:34 2011
;; MSG SIZE  rcvd: 36

Status also REFUSED. Now we are really and truly dead. If you are using a BIND nameserver you have no way to send email to someone@appliedcoatings.ca. But not so with Google!

Of course I don't know how Google wrote their DNS server, but I do think that some of their infrastructure experts write it themselves rather than using open source programs. So with a Google nameserver you will get a response:

dig mx appliedcoatings.ca @8.8.8.8

; <<>> DiG 9.7.1-P2 <<>> mx appliedcoatings.ca @8.8.8.8
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 6901
;; flags: qr rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 0

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      MX

;; ANSWER SECTION:
appliedcoatings.ca.     82805   IN      MX      10 mail.appliedcoatings.ca.

;; Query time: 4 msec
;; SERVER: 8.8.8.8#53(8.8.8.8)
;; WHEN: Sun Aug 21 19:11:14 2011
;; MSG SIZE  rcvd: 57

and just to close the loop and make sure this is a valid host you would do this:

dig mail.appliedcoatings.ca @8.8.8.8

; <<>> DiG 9.7.1-P2 <<>> mail.appliedcoatings.ca @8.8.8.8
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 35190
;; flags: qr rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 0, ADDITIONAL: 0

;; QUESTION SECTION:
;mail.appliedcoatings.ca.       IN      A

;; ANSWER SECTION:
mail.appliedcoatings.ca. 86400  IN      A       66.183.21.181

And we can go the next step and begin an SMTP conversation with that server to make sure it is really operating. After all, if they messed up DNS there's no telling what else they might have gotten wrong.

 telnet  66.183.21.181 25
Trying 66.183.21.181...
Connected to 66.183.21.181.
Escape character is '^]'.
220 mail.appliedcoatings.ca Microsoft ESMTP MAIL Service, Version: 6.0.3790.4675 ready at  Sun, 21 Aug 2011 16:22:04 -0700
HELO localhost
250 mail.appliedcoatings.ca Hello [50.17.188.196]
quit
221 2.0.0 mail.appliedcoatings.ca Service closing transmission channel
Connection closed by foreign host.

Yup. They've got an operating mail server at that IP.

So we can reverse engineer a bit what Google's DNS server must have done behind the scenes to arrive at a valid answer where BIND could not. I'm 100% sure that Google would have also done the query

dig mx appliedcoatings.ca @ns1.domainpeople.com

since that is the right thing to do. But not getting a satisfactory answer (status: REFUSED), what it must do additionally after getting refused a second time by ns2.domainpeople, is to go back to the originally named nameservers sp1 and sp2. Watch what happens in that case:

 dig mx appliedcoatings.ca @sp1.domainpeople.com.

; <<>> DiG 9.7.1-P2 <<>> mx appliedcoatings.ca @sp1.domainpeople.com.
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 10226
;; flags: qr aa rd; QUERY: 1, ANSWER: 1, AUTHORITY: 2, ADDITIONAL: 1
;; WARNING: recursion requested but not available

;; QUESTION SECTION:
;appliedcoatings.ca.            IN      MX

;; ANSWER SECTION:
appliedcoatings.ca.     86400   IN      MX      10 mail.appliedcoatings.ca.

;; AUTHORITY SECTION:
appliedcoatings.ca.     86400   IN      NS      ns1.domainpeople.com.
appliedcoatings.ca.     86400   IN      NS      ns2.domainpeople.com.

;; ADDITIONAL SECTION:
mail.appliedcoatings.ca. 86400  IN      A       66.183.21.181

The AA (authoritative) flag is set in the response. So it's a good response, but sent to the "wrong" nameserver. Nevertheless, it is a response and it gets anyone using that nameserver more functionality than someone using BIND.

Conclusion
So far we've got three advantages speaking favorably for Google's DNS server: it's faster, it's answers are more complete and it's universally available. Wait, there's more! Another nice thing is what it does not do. Some ISPs have a "feature" I call DNS clobbering. In fact it's so annoying I will devote a whole blog post to describing it in more detail. Essentially they take license with DNS and make up answers to some queries! It's true and it's truly annoying. Not all ISPs do this but mine certainly does. So the other nice thing about Google DNS is that it does not do DNS clobbering and it's available for you to use it at home and avoid this annoying feature. You just set your DNS servers rather than have them assigned automatically via DHCP.

Other Resources
I should mention that while researching public DNS servers I was also led to commercial versions of the same thing. I went so far as to test the timings on one of those services and found that it is more distant, round-trip-wise, than Google's anycast server. Stands to reason. Google's got the best Internet access of anyone. They're on all the major highways. The commercial offerings have some additional cool features, however. They can serve as URL filter. So if someone puts in a URL which leads to a malicious site, for example, they can respond with an answer that spares you from going to that infected site. This is a little more crude than URL filtering at the proxy level, since a DNS server has no knowledge of the URI whereas a proxy URL filter does, but it could be quite serviceable. I'm not sure it allows you to pick and choose URL categories to block as with a URL filter (gambling, porn, hacking sites, etc.).

A lot more information on using Google DNS is at http://code.google.com/speed/public-dns/docs/using.html.

September 1 Update - a Crack in the Infrastructure
I now have my first case of a domain name which Google DNS did not resolve correctly, and for no apparent reason. The domain name is forums.tweaktown.com. Here's proof of Google's failure, followed immediately by Amazon's DNS servers' success:

dig forums.tweaktown.com @8.8.8.8

; <<>> DiG 9.7.1-P2 <<>> forums.tweaktown.com @8.8.8.8
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NXDOMAIN, id: 15826
;; flags: qr rd ra; QUERY: 1, ANSWER: 0, AUTHORITY: 1, ADDITIONAL: 0

;; QUESTION SECTION:
;forums.tweaktown.com.          IN      A

;; AUTHORITY SECTION:
tweaktown.com.          116     IN      SOA     ns21.domaincontrol.com. dns.jomax.net. 2011060602 28800 7200 604800 86400

;; Query time: 4 msec
;; SERVER: 8.8.8.8#53(8.8.8.8)
;; WHEN: Thu Sep  1 14:40:50 2011
;; MSG SIZE  rcvd: 106


 dig forums.tweaktown.com

; <<>> DiG 9.7.1-P2 <<>> forums.tweaktown.com
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 52290
;; flags: qr rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 2, ADDITIONAL: 1

;; QUESTION SECTION:
;forums.tweaktown.com.          IN      A

;; ANSWER SECTION:
forums.tweaktown.com.   1885    IN      A       38.101.21.25

;; AUTHORITY SECTION:
tweaktown.com.          1943    IN      NS      ns22.domaincontrol.com.
tweaktown.com.          1943    IN      NS      ns21.domaincontrol.com.

;; ADDITIONAL SECTION:
ns21.domaincontrol.com. 753     IN      A       216.69.185.11

;; Query time: 0 msec
;; SERVER: 172.16.0.23#53(172.16.0.23)
;; WHEN: Thu Sep  1 14:40:55 2011
;; MSG SIZE  rcvd: 122

All BIND servers I tried during this time returned the correct answer.

Is this an isolated incident or a tip of an iceberg of problems? I hope it is a one-off. I'll post updates as I find out more. I am slightly concerned now.

References and related
I finally wrote my own web interface to DNS and published the code I did it with. Check it out here.

A web interface to Google's public DNS service, which will give you more debug information, is https://dns.google.com/