Skip to content

development-notes.md

Latest commit

 

History

History
143 lines (110 loc) · 7.49 KB

development-notes.md

File metadata and controls

143 lines (110 loc) · 7.49 KB

Dev Notes

Development

You might want to run cargo install cargo-insta and run cargo insta for reviewing some tests (especially with parsers).

With debug enabled and other potentially useful cases:

cargo test --features zuban_debug
cargo test --test mypylike -- --only-typecheck testGenericMethodReturnType
cargo test --test mypylike -- --only-language-server
cargo test -- --nocapture
RUSTFLAGS="-Z macro-backtrace" cargo test
CARGO_TARGET_DIR=/tmp/cargo_target cargo test

Debugging:

gdb -ex r --args <executable> <arguments...>

On Windows' Github CI the tempdir path may be part of a symlink. Therefore you may want to run tests like this:

TMP=~/ cargo test -p zubanls

Releasing

Releasing is typically done by GitHub actions, which invokes deploy/pypi/zuban/build.sh. For a patch release, simply run ./release.sh, for a minor release ./release.sh minor (minor here comes from cargo release).

Running primer

Mypy uses mypy-primer to test for changes. We can simply check if we panic on some of these projects with:

cargo run --bin primer --release

# Run on all files to see if there are any panics
RUST_BACKTRACE=1 cargo run --bin primer --release -- -- --strict --ignore-excludes-from-config .

Profiling

sudo apt-get install linux-tools-generic
cargo install flamegraph

# Might be needed
sudo sysctl -w kernel.perf_event_paranoid=-1
sudo sysctl -w kernel.kptr_restrict=0

# We need to force-frame-pointers otherwise there are a lot of Unknown frames
RUSTFLAGS="-C force-frame-pointers=yes" flamegraph -- cargo test jedilike --release
# Sometimes the flamegraph lacks some base frames, because only part of the
# stack is copied, use the command below.
# See also https://github.com/flamegraph-rs/flamegraph/issues/193#issuecomment-2119274041
RAYON_NUM_THREADS=1 flamegraph -c 'record -F 100 --call-graph dwarf,50000 -g' -- /some-binary --release
firefox flamegraph.svg

perf report --stdio --call-graph none | less
# Sort by the Self column, because I haven't found out how to do this with perf report
perf report --stdio  --call-graph none | sort -nr -k2 | less

Progress History

Mypy-Like Tests

  • 2022-09-27: 1170 / 2547 (0.08s -> 14625 tests/s 44b39e90299909bb1672a0a64699c42145efda35)
  • 2023-01-20: 1622 / 2989 (0.13s -> 12476 tests/s 914aede21d716bd80b9a790c213b6d54024b8b79)
  • 2023-02-23: 1750 / 3104 (0.27s -> 6481 tests/s a8d9592a77286caa42501c69ea8b7b89c430c4a8) (note: Cache typeshed multiple times)
  • 2023-03-22: 2106 / 3866 (0.30s -> 7262 tests/s d5b94d4a3dcb75fbdb2abb7c6feb41dd7bd287ec)
  • 2023-04-23: 2328 / 4006 (0.31s -> 7510 tests/s 515d90a4a5c12af68cd885db0bf2b79afe88f54d)
  • 2023-05-23: 2881 / 4845 (0.34s -> 8473 tests/s 2653b7365e335a51adc3eb7dcfede26fe7c7d3d0)
  • 2023-06-23: 3411 / 5752 (0.39s -> 8746 tests/s 52f55a6477c9ecb71a76323cb9f8937e8f0c2433)
  • 2023-07-23: 3744 / 6061 (0.48s -> 7800 tests/s 0d94b68467ac9b4f82e393af4a2d462beee12ead) (note: Removed a dbg)
  • 2023-08-23: 4443 / 6770 (0.53s -> 8383 tests/s d306fda4bcd8455212122ddb683168a2d9da850e)
  • 2023-09-23: 4822 / 7390 (0.60s -> 8037 tests/s a46931615583c16ef412f4f00b27d712e1fcd897)
  • 2023-10-23: 5086 / 7397 (0.72s -> 7064 tests/s 8b191e26f20d15c7616adffb19674c42e56bb016)
  • 2023-11-23: 5467 / 7643 (1.00s -> 5467 tests/s eff073096626fbede374f3d9431d2a17c644a2c0)
  • 2023-12-23: 5686 / 7657 (1.02s -> 5574 tests/s 30f78634246781c10af0d098bd521487522d9c98)
  • 2024-01-23: 5836 / 7719 (1.09s -> 5354 tests/s 6de18286ca58e27e8da9ff72a3f39fd86d50ad6)
  • 2024-02-23: 5927 / 7739 (1.05s -> 5645 tests/s d718904967d63839be27a8822cba5a48ea186a6a)
  • 2024-03-23: 6128 / 8248 (1.13s -> 5423 tests/s c1ac728e33fd3034887bf39c34edf3047e903e8a)
  • 2024-04-23: 6414 / 8417 (1.24s -> 5173 tests/s e70c61aa9ec0ce2735dd0890f8803537ae4ba115)
  • 2024-05-23: 6826 / 8674 (0.80s -> 8533 tests/s 98753cd9964134f07e66079a650c1966905a7e40) (note: Reuse typeshed caches
  • 2024-06-23: 7129 / 8778 (1.17s -> 6093 tests/s fad9c939a769eb1129ea73581822f559c9d482ed)
  • 2024-07-23: 7425 / 8955 (1.38s -> 5380 tests/s b770518b210c23cfb116ecd943f7aefc8e95e8e7)
  • 2024-08-23: 7725 / 9082 (1.47s -> 5255 tests/s 1d580b30e1b0712182c9c827ef79edd682994dac)
  • 2024-09-23: 7945 / 9150 (1.51s -> 5262 tests/s 17da15d9c0a208848aad501b83ebb875d1a4bb5f)
  • 2024-10-23: 8143 / 9344 (1.60s -> 5089 tests/s 1540a3062b529d99b04fc5865def01482822872b)
  • 2024-11-23: 8480 / 9668 (1.65s -> 5139 tests/s 0dac1c48911dffc925c03b81a79d3f099bae6ffb) (note: Avoided union replacing)
  • 2025-02-23: 8812 / 10128 (1.25s -> 7050 tests/s b65d5256011792e807562f366168a3d3b7845eaa)
  • 2025-03-23: 8968 / 10369 (1.27s -> 7061 tests/s 4d34f7bc9aaa82ed1b64c54479a29932953a8eb8)
  • 2025-04-23: 9272 / 10582 (1.62s -> 5723 tests/s b068ac600aff0dd46bf9e97c4889339ec99fc763) (note: tests cannot reuse Python versions sometimes)
  • 2025-05-23: 9400 / 10701 (1.71s -> 5497 tests/s 4827fc01e4052bfe9fbbd0535a84906ae9159b7d)
  • 2025-07-23: 9573 / 10859 (1.89s -> 5065 tests/s 5b5c3779cde57d943d8d0f06c3ef1e0142f49da3)
  • 2025-08-23: 9789 / 11063 (2.18s -> 4490 tests/s da3fd1393bc921d66ac143889505c239a30777b9) (note: most things are Sync now)
  • 2025-09-23: 9970 / 11203 (2.28s -> 4373 tests/s 6db6999b5fa2614985f1a84889970eb01fac5587)

Jedi-Like Tests

  • 2025-07-23: 1908 / 2494 77% (0.10s -> 24940 tests/s 5b5c3779cde57d943d8d0f06c3ef1e0142f49da3)

Unsound?

  • Choosing __new__ or __init__ via heuristic
  • Initializing casting Type[Class] to Type[object] and then Type[object]() is ok?
  • Ignoring positional arg names when assigning functions in:
    • Override checks (including multiple inheritance checks)
    • Checks where one side of callables is implicitly typed (i.e. def foo...)
    • Checks of inplace operators (However this does not matter, probably).
    • overlapping checks
  • Sequence[str] :> str have both different __contains__ implementations (see Michi's email)
  • property narrowing with __set__ narrows __get__ (see testSubclassDescriptorsBinder)
  • property setter type can be != getter return type python/mypy#3004
  • variables don't need to actually be initialized:
    • foo: str
    • dataclasses __init__ is overwritten and therefore members are not necessarily initialized.
  • hasattr narrowing uses Any instead of object
  • a: int; callable(a) will narrow a to be a callable that returns Any and takes any parameters
  • Enum value when a = 1 and b = '' will be Any for no reason
  • total_ordering fills functions with an Any return type if the other functions return not None

Questionable, maybe change behind flag?

  • .py not type-checked if .pyi exists
  • Method Overrides param signatures are not type checked if they are not annotated

Mypy: Please change!!

  • Untyped decorated methods are not checked when --check-untyped-defs, see also test untyped_override_check_untyped_defs_mypy_compatible

Development Introduction

  • Test driven, always create a failing test first
  • Working on Parser Trees, see also https://docs.python.org/3/reference/grammar.html and crates/parsa_python
  • Trees are flat in a Vec in Memory
  • Trees and compiler information (Vec<Point>) are separate data structures, it's also flat in memory with the same length as the tree.
  • NodeRef or PointLink is essentially a pointer to both a parser tree terminal/nonterminal and a Point
  • Compiler information is either a type (Type), a special object (Specific) or a redirect
  • There is a concept of incremental compiling with localities