AWS EC2 dns resolution diagnostic

Question

I am using EC2 instances with amazon linux installed (with amazon dns server settings, which comes from DHCP), as well as an RDS database. The EC2 instances are behind ELB and get high traffic. The application that I use is coded with PHP.

The problem is when PHP tries to connect to the RDS database, sometimes it returns the following error:

PHP Warning:  mysqli_connect(): (HY000/2005): Unknown MySQL server host ...

It doesn't happen a lot but sometimes it is geting worse; I'm getting thousands of error events with that message.

Is there any suggestion for diagnosing the problem? I was thinking about dumping all DNS traffic to a file and checking it but servers get really high traffic so it will be hard to track from that file.

Ip:
197171459 total packets received
1 with invalid addresses
0 forwarded
0 incoming packets discarded
197171458 incoming packets delivered
175015443 requests sent out
Icmp:
12528 ICMP messages received
0 input ICMP message failed.
ICMP input histogram:
    destination unreachable: 188
    echo requests: 12340
12559 ICMP messages sent
0 ICMP messages failed
ICMP output histogram:
    destination unreachable: 219
    echo replies: 12340
IcmpMsg:
    InType3: 188
    InType8: 12340
    OutType0: 12340
    OutType3: 219
Tcp:
5231380 active connections openings
3978862 passive connection openings
881 failed connection attempts
6420 connection resets received
17 connections established
191630575 segments received
200105352 segments send out
2797151 segments retransmited
0 bad segments received.
6910 resets sent
Udp:
5577451 packets received
219 packets to unknown port received.
0 packet receive errors
5577700 packets sent
UdpLite:
TcpExt:
172 invalid SYN cookies received
808 resets received for embryonic SYN_RECV sockets
7176788 TCP sockets finished time wait in fast timer
507 packets rejects in established connections because of timestamp
448055 delayed acks sent
2927 delayed acks further delayed because of locked socket
Quick ack mode was activated 2433 times
94865861 packets directly queued to recvmsg prequeue.
16611185 packets directly received from backlog
54150864749 packets directly received from prequeue
2158966 packets header predicted
79141174 packets header predicted and directly queued to user
40780030 acknowledgments not containing data received
56946553 predicted acknowledgments
84 times recovered from packet loss due to SACK data
Detected reordering 4 times using FACK
Detected reordering 11 times using SACK
Detected reordering 69 times using time stamp
70 congestion windows fully recovered
1241 congestion windows partially recovered using Hoe heuristic
TCPDSACKUndo: 13
2491 congestion windows recovered after partial ack
0 TCP data loss events
220 timeouts after SACK recovery
104 fast retransmits
99 forward retransmits
7 retransmits in slow start
2792531 other TCP timeouts
22 times receiver scheduled too late for direct processing
2423 DSACKs sent for old packets
2785871 DSACKs received
5162 connections reset due to unexpected data
921 connections reset due to early user close
135 connections aborted due to timeout
TCPDSACKIgnoredOld: 533
TCPDSACKIgnoredNoUndo: 393
TCPSackShifted: 477
TCPSackMerged: 536
TCPSackShiftFallback: 2709
TCPBacklogDrop: 46
TCPDeferAcceptDrop: 3906058
IpExt:
InOctets: 69400712361
OutOctets: 94841399143

Solution

There is a known AWS bug which causes DNS resolution to sporadically fail:

https://forums.aws.amazon.com/thread.jspa?messageID=330465#330465

You might want to test with persistent connections as that would reduce the frequency at which DNS resolution is performed.

A local DNS cache (e.g. pdns-recursor or dnscache) will reduce the frequency but the RDS hostname records have very short (60 second) TTLs so it will mean that the problem occurs far less frequently but still happens a few times a day.

Comments

[Konrad K.]

We're seeing this too on an EC2 hosted site with only developer traffic + also with RDS as the backend db. Now at least 1 out of 10 queries are ended in error with an "Unknown MySQL server host" message. It looks as if queries taking several seconds to run would be more prone to this than those that execute in less than a second.

[Kevin Lee]

Actually because of AWS dont have that kind of choice because they are giving a hostname and they dont promise that ip will stay same, so it changes, may make a local cache to update them but then it must fail as well. since it need to check it every second or maybe less than second. adding netstat -s output from an appSrv, it seems ok to me