Do not truncate Logger timestamps

[binarylogic]

Is there any reason why the precision on Logger timestamps is being truncated? If so, can this be commented / documented somewhere?

A few reasons we should not do this:

1. We're seeing log messages share the same timestamp, thus presenting them out of order in external logging systems.
2. The syslog specification allows up to a microsecond precision: https://tools.ietf.org/html/rfc5424
3. The backend should be solely responsible for removing precision.

Thanks for taking a look!

[josevalim]

In order to not change the output currently logged, let's make this be:

[pad2(hh), ?:, pad2(mi), ?:, pad2(ss), ?., ms |> div(1000) |> pad3()]

If you want to log with higher precision, then you can always call your own format_time function.

This way you get higher precision in the backend but remove the precision only when formatting, as you mentioned.

[binarylogic]

Sure, makes sense. I'll update.

[binarylogic]

Ok, all set.

[fishcakez]

This would be a breaking change, so would have to be opt-in. I would rather we support an option to use a date time struct.

[josevalim]

@fishcakez you are right. We are changing the meaning of the tuple element. I think having an option for using time structs is likely the way to go here.

[binarylogic]

Sure, do you mind explaining why this is breaking? And does this mean Elixir is locked in at millisecond precision for logging, by default?

[fishcakez]

do you mind explaining why this is breaking?

We support custom formatting functions, and these receive the timestamp tuple, which is currently milliseconds. See Logger.Formatter.compile({module, function}). Therefore we can't change the timestamp to be microseconds, instead of milliseconds.

And does this mean Elixir is locked in at millisecond precision for logging, by default?

Yes

[binarylogic]

Thanks. I'll update this to take an option.

[josevalim]

Hi @binarylogic! I am closing this PR as, per the discussion above, it is backwards incompatible. Instead, we should probably move to a NaiveDateTime struct, configured via an option. Thank you!