Problems using UDP influxdb output

[mcfedr]

Looks like telegraf is splitting a udp message and this means its not getter parsed properly by influxdb.

This has started happening since update to telegraf 1.3

Relevant telegraf.conf:

[[outputs.influxdb]]
  urls = ["udp://influxdb:8089"]
  database = "telegraf"
  retention_policy = ""
  write_consistency = "any"
  timeout = "5s

[[inputs.system]]

System info:

Telegraf 1.3 docker image
Influxdb 1.2 docker image

Log message

influxdb_1    | [I] 2017-05-29T09:25:00Z Failed to parse points: unable to parse 'system,host=proxy-docker uptime_format="5 days 1496049900000000000': unbalanced quotes service=udp
influxdb_1    | [I] 2017-05-29T09:25:00Z Failed to parse points: unable to parse 'system,host=proxy-docker   1:26" 1496049900000000000': invalid field format service=udp

[oplehto]

I am seeing something which may be related with the socket writer plugin with UDP: Unrelated metrics are getting randomly dropped when there are measurements active which require a split into multiple UDP packets. The amount of drops seems to correlate with the amount of splits required.

I haven't had time to look in any deeper but I suspect that this loop breaks prematurely:

telegraf/plugins/outputs/influxdb/client/udp.go

Line 75 in f74687d

func (c *udpClient) WriteStream(r io.Reader, contentLength int) (int, error) {

[sebito91]

The splitter was updated before the v1.3.0 release here: #2795. The splits actually take place in the influxdb Write function (the conn in the udp.go file).

Split function -- https://github.com/influxdata/telegraf/blob/master/metric/metric.go#L220

[danielnelson]

I think maybe we shouldn't determine the buffer length until after splitting https://github.com/influxdata/telegraf/blob/master/plugins/outputs/influxdb/influxdb.go#L189

[sebito91]

I see what happened, it was refactored slightly on #2799 since we merged #2795! @danielnelson, didn't realize we actually merged that...I think we built with just #2795 but will confirm in the morning.

[danielnelson]

I'm pretty sure this is it, working on PR now.

[danielnelson]

The invalid points was fixed in 1.3.1, but there was still an edge case where long point are not split correctly and they would be dropped. This is fixed in 1.3.2.