[EPIC] Migrate static type checking to ty

This is a multi-stage epic, not a starter task. Stage 3 is the most approachable piece: pick one suppressed ty rule such as invalid-method-override and remove its # ty: ignore comments in pyiceberg/. Start by running make typecheck-ty and uv run ty check to reproduce the diagnostics. Done means the rule stays at error and the check passes with the comments removed.

Feature Request / Improvement

Why

ty is a modern, Rust-based type checker built for fast feedback in CI and editors

• ty is faster, I timed three cold runs (runs without a cache) on the 91 package files, in the same environment. ty took 0.37 to 0.44 s, and mypy took 6.28 to 6.37 s

• ty includes a language server, so editors show the same diagnostics (errors and warnings) as CI
• Its type system narrows types with intersections and skips branches that type inference shows are unreachable, mypy just assumes Any

The mypy setup we have is broken, and that makes parts of this migration easier than it looks. It passes because its pre-commit environment holds four stub packages, so Pydantic, PyArrow and the rest are Any there. Run the same mypy against the real environment, mypy --python-executable .venv/bin/python, and it reports 561 errors in 35 files.

There is no clean baseline to preserve. Matching mypy's green would mean pointing ty at the same stub environment, which drops its count from 470 to 104. The libraries have to be hidden to reproduce the number we currently call passing. ty reaches the real figure about fifteen times faster.

Prior work

#3964 added the ty configuration, and #3965 fixed Pydantic subclass constructors and test annotations, which cut the count from 3,547 to 655 diagnostics at the time. #4207 adds ty as a pre-commit hook that runs in its own stub environment, and reports on pull requests without blocking them

Baseline

I measured these numbers at commit 12a8f3b1 with ty 0.0.84, mypy 1.18.2 and Python 3.11, after make install. Both checkers now start on the 91 files in pyiceberg/. The rows that say "before stage 1" are what the checker reports on the code itself, with no suppressions; the rows after them are only green because stage 1 added the comments below, so they say nothing about the code being clean

CI does not run the project-environment check at all. make typecheck-ty runs the stub-only hook, which has six rules switched off, so the 470 and the 403 comments are a local measurement until something runs uv run ty check

Check	Command	Result
CI mypy (pre-commit environment)	uv run prek run mypy -a	Passes on 91 files, because this environment cannot see the libraries
mypy (project environment)	uv run mypy	561 errors in 35 files
ty (project environment), before stage 1	uv run ty check	470 diagnostics in 45 files
ty (project environment), stage 1	uv run ty check	Passes, covering the 470 with 403 # ty: ignore comments
ty (stub-only environment), before stage 1	ty check against the types group venv	104 diagnostics, 3 of them artifacts of the missing libraries
ty (stub-only environment), stage 1	make typecheck-ty	Passes, with the 403 comments plus the hook's six rule deltas
ty in a tests/ scope	uv run ty check -c 'src.include = ["pyiceberg", "tests"]'	403 further diagnostics, all in tests/. Out of scope for now

The two environments are not interchangeable: the 104 and the 470 describe different code, so suppression comments written against one environment do not hold in the other

Settings

This table shows which rule enforces each setting in [tool.mypy] after mypy is gone

mypy setting	mypy default	Replacement	Wired
disallow_untyped_defs	false	Ruff ANN0 and ANN2 with mypy-init-return = true	No, stage 2
disallow_any_generics	false	ty missing-type-argument	Yes
warn_unused_ignores	false	ty unused-ignore-comment and unused-type-ignore-comment	Yes
warn_redundant_casts	false	ty redundant-cast	Yes
warn_unreachable	false	ty redundant-condition, partly	Yes
implicit_reexport = false	true	None. ty has no equivalent rule among its 254	No, decision needed
ignore_missing_imports = true	false	ty allowed-unresolved-imports, per module	Yes
no_implicit_optional = true	true	ty invalid-parameter-default	Redundant: mypy already defaults to true

Seven of the eight lines are deviations from a mypy default; no_implicit_optional is not one of them, and running mypy 1.18.2 with an empty configuration shows it already errors on def f(x: int = None). Its ty mirror therefore enforces nothing the project did not already enforce

By default, ty 0.0.84 fails on warnings (terminal.error-on-warning = true). I kept that default, so the rules above stay at "error" and the existing diagnostics are suppressed one site at a time instead of downgrading a rule and hiding every future error of that kind

Plan

Stage 1: Parity (done, #4207)

• Scope both checkers to the same files: files = ["pyiceberg"] for mypy, src.include = ["pyiceberg"] for ty, both at Python 3.11
• Give each ty rule that replaces a mypy flag the same level mypy used, and name the flag next to it
• Run ty in its own pre-commit environment, which holds only ty and the four stub packages, so both checkers see the same third-party types. ty picks the environment up from the VIRTUAL_ENV the hook runner exports
• Report ty on every pull request without blocking: the typecheck-ty job sets continue-on-error and stays out of python-ci-required
• Add make typecheck-ty. The hook is manual-stage, so make lint does not run it
• Suppress the 470 diagnostics in place, one # ty: ignore[rule] per site, rather than a ratchet that downgrades the 12 rules

Stage 2: Drop mypy

• Keep tests/ out of the ty scope for now. It adds 403 diagnostics, but the two biggest families are the ones the package already carries: invalid-argument-type 120 and unresolved-attribute 116. Sampling them shows BytesIO against the OutputStream protocol, which is the same finding as pyiceberg/avro/file.py:260, and attribute access on X | None unions, which is the same as the package's 53 unresolved-attribute sites. So this is package debt at test sites, not mock noise, and it is deferred rather than relaxed. I looked at ten other projects for comparison, five of the most-starred Python ones and five ASF ones: seven either leave tests/ out of the checker or include it with the expensive checks switched off for test modules, and none checks tests at full strictness, so no established strategy exists to copy here
• Add the three files outside the package that mypy checked before stage 1: setup.py, dev/provision.py and mkdocs/gen_doc_stubs.py
• Set analysis.respect-type-ignore-comments = false. The 121 # type: ignore comments in pyiceberg/ hide 65 diagnostics, so fix those or convert each comment to a # ty: ignore[rule] first

• Add Ruff's ANN0 and ANN2 rules with flake8-annotations.mypy-init-return = true. They flag three functions today, and they are the three that mypy already exempts: expressions/literals.py:90 and cli/console.py:44 carry # type: ignore[no-untyped-def], utils/deprecated.py:31 carries a bare # type: ignore. So this converts three suppressions rather than annotating new code. deprecated needs ParamSpec
• Decide the implicit_reexport gap, or accept losing the check
• Remove the six rule deltas from the ty hook, which exist only because the stub environment cannot see the real libraries
• Make ty a required check: drop continue-on-error, add the job to python-ci-required, and sync .asf.yaml. The if: github.event_name == 'pull_request' guard has to go first, because python-ci-required also runs on merge_group and a job that is skipped there fails its every-result-is-success loop
• Delete the mypy hook with its stub packages, [tool.mypy], the mypy pin and the mypy mentions in AGENTS.md. Keep the mypy-boto3-* packages, because catalog/glue.py and catalog/dynamodb.py import them for type checking

Stage 3: Fix one suppressed rule at a time

Pick a rule, delete its suppressions, fix the code, and leave the rule at "error". One rule per pull request. unused-ignore-comment already fails when a site is fixed and its comment stays behind, so nothing can be forgotten

Rule	Sites
invalid-method-override	104
missing-type-argument	93
invalid-argument-type	75
unresolved-attribute	53
invalid-return-type	44
no-matching-overload	36
invalid-assignment	12
unsupported-operator	5
missing-argument	2
invalid-type-form	1
not-subscriptable	1
call-non-callable	1
possibly-missing-submodule (warns)	1
pydantic-discarded-extra-argument (warns)	1

The 12 error-level rules are suppressed on 429 sites across 403 comments, because 26 comments name two rules. Taking invalid-method-override first clears 104 of them and needs nothing from the libraries, which makes it the cheapest rule to start with

• Replace the 56 bare # type: ignore comments. Fix the code, or use # type: ignore[code] # ty: ignore[rule]. In 17 of them, as_bound returns type[...] in a class that has a field named type. builtins.type[...] fixes these in both checkers
• Then bring tests/ in with path-level relaxation, not another directive wave: [[tool.ty.overrides]] with include = ["tests"] and [tool.ty.overrides.rules], removing one rule per pull request. The 56 # type: ignore comments in tests/ that ty already reports as unused can go in one mechanical commit
• Delete redundant-cast, redundant-condition, the two unused-ignore rules and the missing-override-decorator = "ignore" line from the config once nothing there needs watching

Stage 4: Raise strictness

Apply the configuration ty recommends for strict mode, and fix what it reports

[tool.ty.rules]
blanket-ignore-comment = "error"
dynamic-function-decorator-return = "error"
missing-type-argument = "error"
possibly-unresolved-reference = "warn"
unsound-assignment = "error"
unsound-return-statement = "error"
unsound-yield = "error"
unsupported-dynamic-base = "warn"

# NOTE: the following rules are known to have a significant number of false positives,
# which is mostly unavoidable. Enable them at your own risk!
division-by-zero = "warn"
possibly-missing-attribute = "warn"
possibly-missing-import = "warn"

[tool.ty.analysis]
strict-equality-semantics = true
strict-generic-narrowing = true

[tool.ruff.lint]
extend-select = ["ANN", "PYI", "PGH003"]
preview = true

• redundant-condition-strict reports 16, unsound-assignment 21, dynamic-function-decorator-return 50 and unsound-return-statement 175 diagnostics in pyiceberg/. Turn each one on only after its diagnostics are fixed
• possibly-unresolved-reference, possibly-missing-attribute, possibly-missing-import and unsound-yield report nothing in pyiceberg/ today

• Add @override when you change an override, then turn on missing-override-decorator, which reports 1,033 diagnostics

Done when

• ty is the required check for pyiceberg/ and the three files outside it. tests/ joins once stage 3 has cleared the package
• No suppression comments are left, respect-type-ignore-comments is false and blanket-ignore-comment is on
• The stage 4 configuration is in place, and error-on-warning is still at its default
• mypy, its configuration and its CI step are gone

Decisions needed

• ty has no rule that matches mypy's implicit_reexport = false. Can we accept losing that check? It is what makes mypy report an import as a missing name unless the module re-exports it explicitly
• Should mypy keep its current 214-file scope in this PR? Tradeoff: parity is easier to reason about, but narrowing it silently drops
  tests/ from the only check that runs today
• Delete mypy before or after the suppressions are cleared? Tradeoff: one checker sooner for contributors, versus a longer safety net

cc @rambleraptor @kevinjqliu