Gotcha. Yeah, the rescue lets that test run; after 425k lookups, it never
got slow.

Here's a bit of the strace:
https://gist.github.com/cholick/88c756760faca77208f8


On Wed, Nov 4, 2015 at 11:59 AM, Amit Gupta <agupta(a)pivotal.io> wrote:

Hey Matt,

I wanted to keep using the uaa.SYSTEM_DOMAIN domain, not the internal
domain, for that experiment. I do expect the TCPSocket.open to fail when
talking to 127.0.0.1, what I wanted to know is, in the presence of no other
nameservers, does it eventually start to fail slow again, or does this
behaviour happen only when there are other nameservers. I imagine the
TCPSocket.open is blowing up on the first iteration in the loop and exiting
the script? My bad, can you replace:

TCPSocket.open("--UAA-DOMAIN--", 80).close

with

TCPSocket.open("--UAA-DOMAIN--", 80).close rescue nil

for the experiment with only 127.0.0.1 listed amongst the nameservers?

Yes, something about the move from the first to second nameserver seems
weird. I have seen strace of one case where it times out polling the FD of
the socket it opened to talk to 127.0.0.1, but in one of your straces it
looked like the poll timeout was on polling the FD for the socket for
8.8.8.8. The fact that the problem persists is interesting too, it seems
like it's not just a one-off race condition where someone messed up with FD
it was supposed to be polling.

Thanks,
Amit

On Wed, Nov 4, 2015 at 11:41 AM, Matt Cholick <cholick(a)gmail.com> wrote:

Ah, I misunderstood.

Consul isn't configured as a recursive resolver, so for a test with only
the 127.0.0.1 in resolve.conf I changed the url in the ruby loop to
"uaa.service.cf.internal", which is what uaa is registering for in consul.

I ran through 225k lookups and it never got slow. Here's a bit of the
strace:
https://gist.github.com/cholick/38e02ce3f351847d5fa3

Bother versions of that test definitely pointing to the move from the
first to the second nameserver in ruby, when the first nameserver doesn't
know the address.


On Tue, Nov 3, 2015 at 11:43 PM, Amit Gupta <agupta(a)pivotal.io> wrote:

I looked at the strace, I see you did indeed mean "loop without
resolver on localhost". If you try it with *only* a resolver on localhost,
do you get the eventually consistent DNS slowdown?

On Tue, Nov 3, 2015 at 8:33 PM, Amit Gupta <agupta(a)pivotal.io> wrote:

Thanks Matt!

When you say "the loop without the resolver on local host" did you
mean "the loop with only a resolver on local host"? Sorry if my setup
wasn't clear, but my intention was to only have 127.0.0.1 in
etc/resolv.conf.


On Tuesday, November 3, 2015, Matt Cholick <cholick(a)gmail.com> wrote:

Here are the results of the ruby loop with strace:
https://gist.github.com/cholick/e7e122e34b524cae5fa1

As expected, things eventually get slow. The bash version of the loop
with a new vm each time didn't get slow.

For the loop without a resolver on localhost, it never did get slow.
Though it's hard to prove with something so inconsistent, it hadn't
happened after 100k requests. Here's some of the strace:
https://gist.github.com/cholick/81e58f58e82bfe0a1489

On the final loop, with the SERVFAIL resolver, the issue did
manifest. Here's the trace of that run:
https://gist.github.com/cholick/bd2af46795911cb9f63c

Thanks for digging in on this.


On Mon, Nov 2, 2015 at 6:53 PM, Amit Gupta <agupta(a)pivotal.io> wrote:

Okay, interesting, hopefully we're narrowing in on something.
There's a couple variables I'd like to eliminate, so I wonder if you could
try the following. Also, feel free at any point to let me know if you are
not interesting in digging further.

Try all things as sudo, on one of the CCs.

1. It appears that the problem goes away when the CC process is
restarted, so it feels as though there's some sort of resource that the
ruby process is not able to GC, leading to this problem to show up
eventually, and then go away when restarted. I want to confirm this by
trying two different loops, one where the loop is in bash, spinning up a
new ruby process each time, and one where the loop is in ruby.

* bash loop:

while true; do time /var/vcap/packages/ruby-VERSION/bin/ruby
-r'net/protocol' -e 'TCPSocket.open("--UAA-DOMAIN--", 80).close'; done

* ruby loop

/var/vcap/packages/ruby-VERSION/bin/ruby -r'net/protocol' -e '1.step
do |i|; t = Time.now; TCPSocket.open("--UAA-DOMAIN--", 80).close; puts
"#{i}: #{(1000*(Time.now - t)).round}ms"; end'

For each loop, it might also be useful to run `strace -f -p PID >
SOME_FILE` to see what system calls are going on before and after.

2. Another variable is the interaction with the other nameservers.
For this experiment, I would do `monit stop all` to take one of
your CC's out of commission, so that the router doesn't load balance to it,
because it will likely fail requests given the following changes:

* monit stop all && watch monit summary # wait for all the
processes to be stopped, then ctrl+c to stop the watch
* monit start consul_agent && watch monit summary # wait for
consul_agent to be running, then ctrl+c to stop the watch
* Remove nameservers other than 127.0.0.1 from /etc/resolv.conf
* Run the "ruby loop", and see if it still eventually gets slow
* When it's all done, put the original nameservers back in
/etc/resolv.conf, and `monit restart all`

Again, strace-ing the ruby loop would be interesting here.

3. Finally, consul itself. Dmitriy (BOSH PM) has a little DNS
resolver that can be run instead of consul, that will always SERVFAIL (same
as what you see from consul when you nslookup something), so we can try
that:

* Modify `/var/vcap/bosh/etc/gemrc` to remove the `--local` flag
* Run `gem install rubydns`
* Dump the following into a file, say `/var/vcap/data/tmp/dns.rb`:

#!/usr/bin/env ruby

require "rubydns"

RubyDNS.run_server(listen: [[:udp, "0.0.0.0", 53], [:tcp, "0.0.0.0",
53]]) do
otherwise do |transaction|
transaction.fail!(:ServFail)
end
end

* monit stop all && watch monit summary # and again, wait for
everything to be stopped
* Run it with `ruby /var/vcap/data/tmp/dns.rb`. Note that this
command, and the previous `gem install`, use the system gem/ruby,
not the ruby package used by CC, so it maintains some separation. When
running this, it will spit out logs to the terminal, so one can keep an eye
on what it's doing, make sure it all looks reasonable
* Make sure the original nameservers are back in the
`/etc/resolv.conf` (i.e. ensure this experiment is independent of the
previous experiment).
* Run the "ruby loop" (in a separate shell session on the CC)
* After it's all done, add back `--local` to `
/var/vcap/bosh/etc/gemrc`, and `monit restart all`

Again, run strace on the ruby process.

What I hope we find out is that (1) only the ruby loop is affected,
so it has something to do with long running ruby processes, (2) the problem
is independent of the other nameservers listed in /etc/resolv.conf,
and (3) the problem remains when running Dmitriy's DNS-FAILSERVer instead
of consul on 127.0.0.1:53, to determine that the problem is not
specific to consul.

On Sun, Nov 1, 2015 at 5:18 PM, Matt Cholick <cholick(a)gmail.com>
wrote:

Amit,
It looks like consul isn't configured as a recursive resolver. When
running the above code, resolving fails on the first nameserver and the
script fails. resolv-replace's TCPSocket.open is different from the code
http.rb (and thus api) is using. http.rb is pulling in 'net/protocol'. I
changed the script, replacing the require for 'resolv-replace' to
'net/protocol' to match the cloud controller.

Results:

3286 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4 ms |
dns_close: 0 ms
3287 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3288 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 6 ms |
dns_close: 0 ms
3289 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3290 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3291 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3292 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3293 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 5 ms |
dns_close: 0 ms
3294 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 2008 ms |
dns_close: 0 ms
3295 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4010 ms |
dns_close: 0 ms
3296 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4010 ms |
dns_close: 0 ms
3297 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4006 ms |
dns_close: 0 ms
3298 -- ip_open: 2 ms | ip_close: 0 ms | dns_open: 4010 ms |
dns_close: 0 ms
3299 -- ip_open: 3 ms | ip_close: 0 ms | dns_open: 4011 ms |
dns_close: 0 ms
3300 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4010 ms |
dns_close: 0 ms
3301 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4011 ms |
dns_close: 0 ms
3302 -- ip_open: 1 ms | ip_close: 0 ms | dns_open: 4010 ms |
dns_close: 0 ms

And the consul logs, though there's nothing interesting there:
https://gist.github.com/cholick/03d74f7f012e54c50b56


On Fri, Oct 30, 2015 at 5:51 PM, Amit Gupta <agupta(a)pivotal.io>
wrote:

Yup, that's what I was suspecting. Can you try the following now:

1. Add something like the following to your cf manifest:

...
jobs:
...
- name: cloud_controller_z1
...
properties:
consul:
agent:
...
log_level: debug
...

This will set the debug level for the consul agents on your CC job
to debug, so we might be able to see more for its logs. It only sets it on
the job that matters, so when you redeploy, it won't have to roll the whole
deployment. It's okay if you can't/don't want to do this, I'm not sure how
much you want to play around with your environment, but it could be helpful.

2. Add the following line to the bottom of your /etc/resolv.conf

options timeout:4

Let's see if the slow DNS is on the order of 4000ms now, to pin
down where the 5s is exactly coming from.

3. Run the following script on your CC box:

require 'resolv-replace'

UAA_DOMAIN = '--CHANGE-ME--' # e.g. 'uaa.run.pivotal.io'
UAA_IP = '--CHANGE-ME-TOO--' # e.g. '52.21.135.158'

def dur(start_time, end_time)
"#{(1000*(end_time-start_time)).round} ms"
end

1.step do |i|
ip_start = Time.now
s = TCPSocket.open(UAA_IP, 80)
ip_open = Time.now
s.close
ip_close = Time.now

dns_start = Time.now
s = TCPSocket.open(UAA_DOMAIN, 80)
dns_open = Time.now
s.close
dns_close = Time.now

ip_open_dur = dur(ip_start, ip_open)
ip_close_dur = dur(ip_open, ip_close)
dns_open_dur = dur(dns_start, dns_open)
dns_close_dur = dur(dns_open, dns_close)

puts "#{"%04d" % i} -- ip_open: #{ip_open_dur} | ip_close:
#{ip_close_dur} | dns_open: #{dns_open_dur} | dns_close: #{dns_close_dur}"
end

You will need to first nslookup (or otherwise determine) the IP
that the UAA_DOMAIN resolves to (it will be some load balancer, possibly
the gorouter, ha_proxy, or your own upstream LB)

4. Grab the files in /var/vcap/sys/log/consul_agent/

Cheers,
Amit

On Fri, Oct 30, 2015 at 4:29 PM, Matt Cholick <cholick(a)gmail.com>
wrote:

Here's the results:

https://gist.github.com/cholick/1325fe0f592b1805eba5

The time all between opening connection and opened, with the
corresponding ruby source in http.rb's connect method:

D "opening connection to #{conn_address}:#{conn_port}..."

s = Timeout.timeout(@open_timeout, Net::OpenTimeout) {
TCPSocket.open(conn_address, conn_port, @local_host, @local_port)
}
s.setsockopt(Socket::IPPROTO_TCP, Socket::TCP_NODELAY, 1)
D "opened"

I don't know much ruby, so that's as far I drilled down.

-Matt