Fix problem with graphite talking to closed connections (#2171)
We were having problems with telegraf talking to carbon-relay-ng using the graphite output. When the carbon-relay-ng server restarted the connection the telegraf side would go into CLOSE_WAIT but telegraf would continue to send statistics through the connection. Reading around it seems you need to a read from the connection and see a EOF error. We've implemented this and added a test that replicates roughly the error we were having. Pair: @whpearson @joshmyers
This commit is contained in:
parent
a505123e60
commit
d7a8bb2214
|
@ -2,6 +2,7 @@ package graphite
|
|||
|
||||
import (
|
||||
"errors"
|
||||
"io"
|
||||
"log"
|
||||
"math/rand"
|
||||
"net"
|
||||
|
@ -71,6 +72,31 @@ func (g *Graphite) Description() string {
|
|||
return "Configuration for Graphite server to send metrics to"
|
||||
}
|
||||
|
||||
// We need check eof as we can write to nothing without noticing anything is wrong
|
||||
// the connection stays in a close_wait
|
||||
// We can detect that by finding an eof
|
||||
// if not for this, we can happily write and flush without getting errors (in Go) but getting RST tcp packets back (!)
|
||||
// props to Tv via the authors of carbon-relay-ng` for this trick.
|
||||
func checkEOF(conn net.Conn) {
|
||||
b := make([]byte, 1024)
|
||||
conn.SetReadDeadline(time.Now().Add(10 * time.Millisecond))
|
||||
num, err := conn.Read(b)
|
||||
if err == io.EOF {
|
||||
log.Printf("E! Conn %s is closed. closing conn explicitly", conn)
|
||||
conn.Close()
|
||||
return
|
||||
}
|
||||
// just in case i misunderstand something or the remote behaves badly
|
||||
if num != 0 {
|
||||
log.Printf("I! conn %s .conn.Read data? did not expect that. data: %s\n", conn, b[:num])
|
||||
}
|
||||
// Log non-timeout errors or close.
|
||||
if e, ok := err.(net.Error); !(ok && e.Timeout()) {
|
||||
log.Printf("E! conn %s checkEOF .conn.Read returned err != EOF, which is unexpected. closing conn. error: %s\n", conn, err)
|
||||
conn.Close()
|
||||
}
|
||||
}
|
||||
|
||||
// Choose a random server in the cluster to write to until a successful write
|
||||
// occurs, logging each unsuccessful. If all servers fail, return error.
|
||||
func (g *Graphite) Write(metrics []telegraf.Metric) error {
|
||||
|
@ -91,13 +117,13 @@ func (g *Graphite) Write(metrics []telegraf.Metric) error {
|
|||
|
||||
// This will get set to nil if a successful write occurs
|
||||
err = errors.New("Could not write to any Graphite server in cluster\n")
|
||||
|
||||
// Send data to a random server
|
||||
p := rand.Perm(len(g.conns))
|
||||
for _, n := range p {
|
||||
if g.Timeout > 0 {
|
||||
g.conns[n].SetWriteDeadline(time.Now().Add(time.Duration(g.Timeout) * time.Second))
|
||||
}
|
||||
checkEOF(g.conns[n])
|
||||
if _, e := g.conns[n].Write(batch); e != nil {
|
||||
// Error
|
||||
log.Println("E! Graphite Error: " + e.Error())
|
||||
|
@ -110,6 +136,7 @@ func (g *Graphite) Write(metrics []telegraf.Metric) error {
|
|||
}
|
||||
// try to reconnect
|
||||
if err != nil {
|
||||
log.Println("E! Reconnecting: ")
|
||||
g.Connect()
|
||||
}
|
||||
return err
|
||||
|
|
|
@ -43,7 +43,8 @@ func TestGraphiteOK(t *testing.T) {
|
|||
var wg sync.WaitGroup
|
||||
// Start TCP server
|
||||
wg.Add(1)
|
||||
go TCPServer(t, &wg)
|
||||
t.Log("Starting server")
|
||||
go TCPServer1(t, &wg)
|
||||
// Give the fake graphite TCP server some time to start:
|
||||
time.Sleep(time.Millisecond * 100)
|
||||
|
||||
|
@ -51,6 +52,7 @@ func TestGraphiteOK(t *testing.T) {
|
|||
g := Graphite{
|
||||
Prefix: "my.prefix",
|
||||
}
|
||||
|
||||
// Init metrics
|
||||
m1, _ := metric.New(
|
||||
"mymeasurement",
|
||||
|
@ -72,29 +74,58 @@ func TestGraphiteOK(t *testing.T) {
|
|||
)
|
||||
|
||||
// Prepare point list
|
||||
metrics := []telegraf.Metric{m1, m2, m3}
|
||||
metrics := []telegraf.Metric{m1}
|
||||
metrics2 := []telegraf.Metric{m2, m3}
|
||||
err1 := g.Connect()
|
||||
require.NoError(t, err1)
|
||||
// Send Data
|
||||
t.Log("Send first data")
|
||||
err2 := g.Write(metrics)
|
||||
require.NoError(t, err2)
|
||||
|
||||
// Waiting TCPserver
|
||||
wg.Wait()
|
||||
t.Log("Finished Waiting for first data")
|
||||
var wg2 sync.WaitGroup
|
||||
// Start TCP server
|
||||
time.Sleep(time.Millisecond * 100)
|
||||
wg2.Add(1)
|
||||
go TCPServer2(t, &wg2)
|
||||
time.Sleep(time.Millisecond * 100)
|
||||
//Write but expect an error, but reconnect
|
||||
g.Write(metrics2)
|
||||
err3 := g.Write(metrics2)
|
||||
t.Log("Finished writing second data, it should have failed")
|
||||
//Actually write the new metrics
|
||||
|
||||
require.NoError(t, err3)
|
||||
t.Log("Finished writing third data")
|
||||
wg2.Wait()
|
||||
g.Close()
|
||||
}
|
||||
|
||||
func TCPServer(t *testing.T, wg *sync.WaitGroup) {
|
||||
tcpServer, _ := net.Listen("tcp", "127.0.0.1:2003")
|
||||
func TCPServer1(t *testing.T, wg *sync.WaitGroup) {
|
||||
defer wg.Done()
|
||||
conn, _ := tcpServer.Accept()
|
||||
tcpServer, _ := net.Listen("tcp", "127.0.0.1:2003")
|
||||
conn, _ := (tcpServer).Accept()
|
||||
reader := bufio.NewReader(conn)
|
||||
tp := textproto.NewReader(reader)
|
||||
data1, _ := tp.ReadLine()
|
||||
assert.Equal(t, "my.prefix.192_168_0_1.mymeasurement.myfield 3.14 1289430000", data1)
|
||||
conn.Close()
|
||||
tcpServer.Close()
|
||||
}
|
||||
|
||||
func TCPServer2(t *testing.T, wg *sync.WaitGroup) {
|
||||
defer wg.Done()
|
||||
tcpServer, _ := net.Listen("tcp", "127.0.0.1:2003")
|
||||
conn2, _ := (tcpServer).Accept()
|
||||
reader := bufio.NewReader(conn2)
|
||||
tp := textproto.NewReader(reader)
|
||||
data2, _ := tp.ReadLine()
|
||||
assert.Equal(t, "my.prefix.192_168_0_1.mymeasurement 3.14 1289430000", data2)
|
||||
data3, _ := tp.ReadLine()
|
||||
assert.Equal(t, "my.prefix.192_168_0_1.my_measurement 3.14 1289430000", data3)
|
||||
conn.Close()
|
||||
conn2.Close()
|
||||
tcpServer.Close()
|
||||
}
|
||||
|
|
Loading…
Reference in New Issue