fix: logger metadata can mangle span_ids

[gerbal]

We are deploying Spandex to our elixir services and have discovered in some circumstances it the Elixir Logger can lose precision on very large integers. This has resulted in oddly truncated trace and span ids in our logs like 6123256812116107000. The underlying cause is unclear, but the whole issue can be avoided if the trace_id and span_id are stored as strings in logger metadata.

Is this a useful change for anyone else?

[zachdaniel]

@gerbal so the trace/span ids are in datadog w/o the trailing zeroes? I suspect its related to this https://github.com/spandex-project/spandex_datadog/blob/3344868ac5eca8f7ce6d9558d970edfd0d109073/lib/spandex_datadog/adapter.ex#L24

[gerbal]

The trace_id is actually 6123256812116107452 and what's logged is 6123256812116107000

[zachdaniel]

Yeah, that is really strange...I'd really suggest seeing if you can reproduce the specific issue of integers losing precision when being transferred to logger metadata. That would be an issue w/ Elixir/logger I imagine

[GregMefford]

I didn’t have a chance to try to repro this yet, but if you’re able to give some more details about what versions of Elixir and OTP you’re seeing this on, that would be helpful. I know there are some recent changes to the way Logger works and others may have just not updated/noticed yet if there are issues there. ❤️ 🚀

[zachdaniel]

Yeah, I tried to reproduce it but I couldn't.

[aselder]

The PR I just submitted is a dup of this. I can confirm the behavior, see #126 for details.

[aselder]

For clarity, the issues is with log ingesters (namely datadog), treating JSON integers as JS Numbers (aka float64) and losing precision.