Feature Request: Support metric time type

[zqm6133]

Use Case

I need to write multiple timestamp columns (e.g., created_at, updated_at, event_time)
to a PostgreSQL table with timestamptz type. Currently, Telegraf fields only support
string, float64, int64, uint64, and bool. There is no way to pass a native
time.Time value through the metric pipeline so that outputs like outputs.postgresql
can create timestamp with time zone columns for data fields (not just the metric's
time column).

Expected behavior

Allow time.Time as a valid field value type in telegraf.Metric, and have
outputs.postgresql (and potentially other SQL outputs) map it to timestamptz
or timestamp column types.

Actual behavior

Fields with timestamp data must be stored as strings or integers, which requires:

1. Either pre-processing outside Telegraf to convert before insertion
2. Or maintaining a custom fork with a processor plugin that converts string/numeric
   fields to time.Time and patching the metric core + postgresql output

Additional info

This is distinct from the existing timestamp_column_name option in outputs.postgresql,
which only controls the type of the metric's time column. The need here is for
arbitrary fields to carry timestamp values.

[srebhan]

Thanks for reaching out @zqm6133!

We cannot add a new field-type without breaking all output plugins so this is out of question. But what we could do is to add a timestamp_fields option to the outputs.postgresql plugin where you can specify the fields that carry a timestamp. Ideally, we can define the field to carry a nanosecond timestamp (ns epoch in UTC).

Would that work for you?

[zqm6133]

Thanks for response @srebhan

Yes, a timestamp_fields option on outputs.postgresql works for my core need — real timestamptz / timestamp columns without a new metric field type. Agreed that not breaking every output plugin is the right call.

One practical point on the value format, though. In practice the timestamp data I receive almost never arrives as an ns epoch in UTC. It's usually:

• a unix millisecond epoch (typical of JSON / MQTT payloads), or
• an RFC3339 string like 2024-05-06T08:30:00+08:00.

If timestamp_fields only accepts ns-epoch-UTC, every user has to convert beforehand.

Could the option carry a per-field source format instead of assuming ns epoch? Something like:

 [[outputs.postgresql]]
   [outputs.postgresql.timestamp_fields]
     "*_time"   = "unix_ms" # all *_time fields
     created_at = "2006-01-02T15:04:05Z07:00" # exact name wins over glob
     "*"        = "unix_ns" # optional catch-all

The values unix / unix_ms / unix_us / unix_ns and Go time layouts are exactly what internal.ParseTimestamp already handles (it's used for parsing the metric's own timestamp), so this reuses existing, tested code rather than
adding new parsing. It could default to unix_ns in UTC when only a field name is given, to match what you described.