Problem parsing timestamps in CSV parser

[graemeyeo]

Relevant telegraf.conf:

[[inputs.file]]
  files = ["data2s.csv"]
  data_format = "csv"
  csv_header_row_count = 0
  csv_column_names = ["pid","time","msg"]
  csv_timestamp_column = "time"
  csv_timestamp_format = "unix"

System info:

Telegraf 1.10.3
Ubuntu 16.04 LTS

Steps to reproduce:

CSV contains Unix timestamps with fractional seconds with a large numver of decimal places.

Input csv file:

2108,1551129661.95456123352050781250,0105000008000000cb2e392df52c9c2d
2108,1551129662.14315605163574218750,0405000008000000527bec7ab47a20bf
2108,1551129662.14344882965087890625,02050000080000006a2f312ed62ca02d

Expected behavior:

Timestamps would be parsed as Unix epoch seconds.fractional_seconds at down to nanosecond precision.

Actual behavior:

The timestamps were being parsed without giving an error message, but were incorrect when imported to the database.

Additional info:

It turned out that the ParseTimestampWithLocation function in internal/internal.go was receiving a timestamp string with an exponent e.g. 1.5515752371385763e+09 . This was confusing the part of the function which splits the string into timeInt and timeFractional.

Fix:

The fix was to change one line in plugins/parsers/csv/parser.go inside the function parseTimestamp. I changed the line:
tStr := fmt.Sprintf("%v", recordFields[timestampColumn])
to
tStr := fmt.Sprintf("%.9f", recordFields[timestampColumn])

This retains (up to) nanosecond precision, but forces it not to give an exponent.

EDIT:
In my haste, I forgot to test this with integer timestamps. Of course forcing Sprintf to try to parse as a float causes integers to be parsed incorrectly. The final solution is slightly longer:

import "reflect"

...

var tStr string
k := reflect.ValueOf(recordFields[timestampColumn]).Kind()
if k == reflect.Float32 || k == reflect.Float64 {
    tStr = fmt.Sprintf("%.9f", recordFields[timestampColumn])
} else {
    tStr = fmt.Sprintf("%d", recordFields[timestampColumn])
}