[ruby-core:70398] [Ruby trunk - Bug #11443] DNS name resolution takes twice as long as it should when primary name server is unavailable

Issue #11443 has been updated by Patrick Robinson.


Doing an strace I can see the DNS lookup occurs twice (with and without a failing name server):

~~~
[pid 32129] socket(PF_INET, SOCK_DGRAM|SOCK_NONBLOCK, IPPROTO_IP) = 7
[pid 32129] connect(7, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("1.1.1.1")}, 16) = 0
[pid 32129] poll([{fd=7, events=POLLOUT}], 1, 0) = 1 ([{fd=7, revents=POLLOUT}])
[pid 32129] sendto(7, "\271\254\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\34\0\1", 28, MSG_NOSIGNAL, NULL, 0) = 28
[pid 32129] poll([{fd=7, events=POLLIN}], 1, 3000) = 0 (Timeout)
[pid 32129] socket(PF_INET, SOCK_DGRAM|SOCK_NONBLOCK, IPPROTO_IP) = 8
[pid 32129] connect(8, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("8.8.8.8")}, 16) = 0
[pid 32129] poll([{fd=8, events=POLLOUT}], 1, 0) = 1 ([{fd=8, revents=POLLOUT}])
[pid 32129] sendto(8, "\271\254\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\34\0\1", 28, MSG_NOSIGNAL, NULL, 0) = 28
[pid 32129] poll([{fd=8, events=POLLIN}], 1, 2000) = 1 ([{fd=8, revents=POLLIN}])
[pid 32129] ioctl(8, FIONREAD, [56])    = 0
[pid 32129] recvfrom(8, "\271\254\201\200\0\1\0\1\0\0\0\0\6google\3com\0\0\34\0\1\300\f\0\34"..., 1024, 0, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("8.8.8.8")}, [16]) = 56
[pid 32129] close(7)                    = 0
[pid 32129] close(8)                    = 0
[pid 32129] stat("/etc/resolv.conf", {st_mode=S_IFREG|0644, st_size=226, ...}) = 0
[pid 32129] socket(PF_INET, SOCK_DGRAM|SOCK_NONBLOCK, IPPROTO_IP) = 7
[pid 32129] connect(7, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("1.1.1.1")}, 16) = 0
[pid 32129] poll([{fd=7, events=POLLOUT}], 1, 0) = 1 ([{fd=7, revents=POLLOUT}])
[pid 32129] sendto(7, "\366N\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\1\0\1", 28, MSG_NOSIGNAL, NULL, 0) = 28
[pid 32129] poll([{fd=7, events=POLLIN}], 1, 3000) = 0 (Timeout)
[pid 32129] socket(PF_INET, SOCK_DGRAM|SOCK_NONBLOCK, IPPROTO_IP) = 8
[pid 32129] connect(8, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("8.8.8.8")}, 16) = 0
[pid 32129] poll([{fd=8, events=POLLOUT}], 1, 0) = 1 ([{fd=8, revents=POLLOUT}])
[pid 32129] sendto(8, "\366N\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\1\0\1", 28, MSG_NOSIGNAL, NULL, 0) = 28
[pid 32129] poll([{fd=8, events=POLLIN}], 1, 2000) = 1 ([{fd=8, revents=POLLIN}])
[pid 32129] ioctl(8, FIONREAD, [220])   = 0
[pid 32129] recvfrom(8, "\366N\201\200\0\1\0\f\0\0\0\0\6google\3com\0\0\1\0\1\300\f\0\1"..., 1024, 0, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("8.8.8.8")}, [16]) = 220
[pid 32129] close(7)                    = 0
[pid 32129] close(8)                    = 0
~~~

----------------------------------------
Bug #11443: DNS name resolution takes twice as long as it should when primary name server is unavailable
https://bugs.ruby-lang.org/issues/11443#change-53801

* Author: Patrick Robinson
* Status: Open
* Priority: Normal
* Assignee: 
* ruby -v: ruby 2.2.2p95 (2015-04-13 revision 50295) [x86_64-linux]
* Backport: 2.0.0: UNKNOWN, 2.1: UNKNOWN, 2.2: UNKNOWN
----------------------------------------
On linux /etc/resolv.conf can be configured to set the timeout of a nameserver before trying the next one (defaults to 5 seconds). So in the event of a failed name server we expect the time taken to resolve a hostname will be timeout + query time. However testing shows it takes (2 * timeout) + query time. Example resolv.conf to replicate the problem of a failed name server:

~~~
nameserver 1.1.1.1
nameserver 8.8.8.8
nameserver 8.8.4.4
~~~

Example ruby code to test:

```
require 'net/protocol'
require 'uri'
require 'benchmark'
uri = URI.parse("http://google.com")
Benchmark.measure { TCPSocket.new(uri.host, uri.port) }.real
```

Example benchmark with a above resolv.conf:

```
irb(main):005:0> Benchmark.measure { TCPSocket.new(uri.host, uri.port) }.real
=> 10.593100978
```

Example benchmark with timeout set to 3 in resolv.conf (options timeout:3):

```
irb(main):005:0> Benchmark.measure { TCPSocket.new(uri.host, uri.port) }.real
=> 6.262138267
```

and without the invalid nameserver:

```
irb(main):005:0> Benchmark.measure { TCPSocket.new(uri.host, uri.port) }.real
=> 0.243385728
```

Example C code that proves the resolver is behaving as expected:

```
#include <stdio.h>
#include <netdb.h>

int main(int argc, char *argv[])
{
    struct hostent *hstnm;
    if (argc != 2) {
        fprintf(stderr, "usage: %s hostname\n", argv[0]);
        return 1;
    }
    hstnm = gethostbyname (argv[1]);
    if (!hstnm)
        return 1;
    printf ("Name: %s\n", hstnm->h_name);
    return 0;
}
```

Benchmark of C code with failed nameserver:

~~~
$ time ./gethostbyname google.com
Name: google.com

real	0m5.189s
user	0m0.000s
sys	0m0.000s
~~~

Benchmark of C code without failed nameserver:

~~~
$ time ./gethostbyname google.com
Name: google.com

real	0m0.040s
user	0m0.000s
sys	0m0.000s
~~~



-- 
https://bugs.ruby-lang.org/