[Python 3.10 issues] Internal Changes Exposed Bugs (PEP 626 & friends)

Symptoms you might see

• Debuggers/profilers/tracers behaving oddly (breakpoints miss lines, coverage mismatches).
   

• Tools relying on bytecode offsets break (e.g., assumptions about frame.f_lasti).
   

• Custom instrumentation that inspects frames/bytecode becomes inaccurate.

Why this happens

Python 3.10 adopted precise line number tables (PEP 626), changing how the interpreter maps bytecode to source lines. As part of this:

• frame.f_lasti effectively became an instruction offset rather than a raw byte offset (the interpreter switched to wordcode previously; 3.10 tightened semantics). See What’s New in 3.10 → Tracebacks & line numbers and the frame object docs.
   

• Line number tables are now more accurate for exceptions and tracing, which is great for users, but breaks tools that depended on old, informal behaviors.

Fix / migration patterns

1) Don’t rely on internal bytecode details. Use supported APIs.

• Prefer the dis module to reason about instructions/offsets instead of hardcoding offsets.
   

• For source locations, rely on frame.f_lineno and code object attributes (co_filename, co_firstlineno, co_positions() in 3.11+) rather than decoding internals yourself. See inspect.

2) Update tracing/debugging logic for PEP 626 semantics.

• If you implement sys.settrace/sys.setprofile, expect more trace events and more accurate line hops. Review the tracing hooks docs and adjust filters so you don’t over-count coverage or miss branches.

3) Use higher‑level libraries where possible.

• For code rewriting/analysis, prefer ast/LibCST over bytecode poking. Bytecode is not a stable interface; the AST is.

4) Rebuild & retest C extensions and any frame/bytecode integrations.

• If a C extension touches frame internals or code objects, recompile on 3.10 and audit against the 3.10 C-API changes.

Example: tracing tool that assumed byte offsets

How to avoid pain next time

• Treat bytecode as unstable. If you must inspect it, go through dis and keep logic resilient to offset shifts.
   

• Add CI on new Python versions early (alphas/betas) so tracing/coverage tools surface changes before GA.
   

• Prefer AST/source-level instrumentation for portability across versions.