build: Make backtraces useable in release builds

[szokeasaurusrex]

Having panic = "abort" and strip = "symbols" makes backtraces generated with RUST_BACKTRACE=1 completely useless in release builds.

By changing these options to panic = "unwind" and strip = "debuginfo" makes backtraces usable in release builds. This will help us debug issues we are unable to reproduce, since we can ask users to set RUST_BACKTRACE=1 and provide the logs with this option.

This change does somewhat significantly increase release build binary size (from 7.9 MB to 13.9 MB on my system). This is likely acceptable if it improves our ability to debug problems. If folks complain, we can consider distributing a separate "stripped" binary for users sensitive to large binary size.

Closes #2200

[Swatinem]

this looks reasonable.

Ideally, instead of enabling unwinding directly, another alternative would be to add -C force-unwind-tables to the RUSTFLAGS (see https://swatinem.de/blog/unwind-tables/#rust)

Another option to consider would be using debug = "line-tables-only", which would give you a full-fidelity backtrace with inline frames, etc, but I bet that would inflate the binary size even more.

[szokeasaurusrex]

Ideally, instead of enabling unwinding directly, another alternative would be to add -C force-unwind-tables to the RUSTFLAGS (see https://swatinem.de/blog/unwind-tables/#rust)

@Swatinem So, if I understand correctly, with this option we would still have panic = "abort"? Would we expect the binary size to be smaller with this option?

Another option to consider would be using debug = "line-tables-only", which would give you a full-fidelity backtrace with inline frames, etc, but I bet that would inflate the binary size even more.

I think let's just try with the current change first, I think it may already be helpful enough for what I need it for (mostly, I just would like to be able to find where errors originating from libraries are coming from. Usually if the error is emitted in Sentry CLI, I can find the error by searching for the error string).

[Swatinem]

with this option we would still have panic = "abort"? Would we expect the binary size to be smaller with this option?

Yes, that codegen option adds the necessary unwind info to the binary so that backtraces work, but it avoids generating all the code to safely unwind in Rust, like destructors, catch_unwind, etc.

[szokeasaurusrex]

Alright, I went with your suggestion to set the -C force-unwind-tables instead of panic = "unwind". With this change, the release binary size (on my system) only increased to 12.2 MB, instead of 13.9 MB.