plugin/proxy: decrease health timeouts (#1107)

Turn down the timeouts and numbers a bit:
FailTimeout 10s -> 5s
Future 60s -> 12s
TryDuration 60s -> 16s
The timeout for decrementing the fails in a host: 10s -> 2s

And the biggest change: don't set fails when the error is Timeout(),
meaning we loop for a bit and may try the same server again, but we
don't mark our upstream as bad, see comments in proxy.go. Testing this
with "ANY isc.org" and "MX miek.nl" we see:

~~~
::1 - [24/Sep/2017:08:06:17 +0100] "ANY IN isc.org. udp 37 false 4096" SERVFAIL qr,rd 37 10.001621221s
24/Sep/2017:08:06:17 +0100 [ERROR 0 isc.org. ANY] unreachable backend: read udp 192.168.1.148:37420->8.8.8.8:53: i/o timeout

::1 - [24/Sep/2017:08:06:17 +0100] "MX IN miek.nl. udp 37 false 4096" NOERROR qr,rd,ra,ad 170 35.957284ms

127.0.0.1 - [24/Sep/2017:08:06:18 +0100] "ANY IN isc.org. udp 37 false 4096" SERVFAIL qr,rd 37 10.002051726s
24/Sep/2017:08:06:18 +0100 [ERROR 0 isc.org. ANY] unreachable backend: read udp 192.168.1.148:54901->8.8.8.8:53: i/o timeout

::1 - [24/Sep/2017:08:06:19 +0100] "MX IN miek.nl. udp 37 false 4096" NOERROR qr,rd,ra,ad 170 56.848416ms
127.0.0.1 - [24/Sep/2017:08:06:21 +0100] "MX IN miek.nl. udp 37 false 4096" NOERROR qr,rd,ra,ad 170 48.118349ms
::1 - [24/Sep/2017:08:06:21 +0100] "MX IN miek.nl. udp 37 false 4096" NOERROR qr,rd,ra,ad 170 1.055172915s
~~~

So the ANY isc.org queries show up twice, because we retry internally -
this is I think WAI.

The `miek.nl MX` queries are just processed normally as no backend is
marked as unreachable.

May fix #1035 #486
This commit is contained in:
Miek Gieben
2017-09-24 20:05:36 +01:00
committed by GitHub
parent 148a99442d
commit 2a32cd4159
5 changed files with 41 additions and 15 deletions

View File

@@ -5,6 +5,7 @@ package proxy
import (
"context"
"fmt"
"net"
"sync/atomic"
"time"
@@ -26,9 +27,9 @@ func NewLookupWithOption(hosts []string, opts Options) Proxy {
upstream := &staticUpstream{
from: ".",
HealthCheck: healthcheck.HealthCheck{
FailTimeout: 10 * time.Second,
MaxFails: 3, // TODO(miek): disable error checking for simple lookups?
Future: 60 * time.Second,
FailTimeout: 5 * time.Second,
MaxFails: 3,
Future: 12 * time.Second,
},
ex: newDNSExWithOption(opts),
}
@@ -85,7 +86,7 @@ func (p Proxy) lookup(state request.Request) (*dns.Msg, error) {
}
// duplicated from proxy.go, but with a twist, we don't write the
// reply back to the client, we return it and there is no monitoring.
// reply back to the client, we return it and there is no monitoring to update here.
atomic.AddInt64(&host.Conns, 1)
@@ -96,11 +97,20 @@ func (p Proxy) lookup(state request.Request) (*dns.Msg, error) {
if backendErr == nil {
return reply, nil
}
if oe, ok := backendErr.(*net.OpError); ok {
if oe.Timeout() { // see proxy.go for docs.
continue
}
}
timeout := host.FailTimeout
if timeout == 0 {
timeout = 10 * time.Second
timeout = 2 * time.Second
}
atomic.AddInt32(&host.Fails, 1)
go func(host *healthcheck.UpstreamHost, timeout time.Duration) {
time.Sleep(timeout)
atomic.AddInt32(&host.Fails, -1)