pyproject.toml

# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES.
# All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# Workspace root for the structured-data-models-client monorepo. This builds no package; each
# independently released distribution lives under packages/ with its own
# pyproject.toml:
#
#   packages/structured-data-models-client            - the public client SDK (pure python)
#   packages/structured-data-models-api               - Universal TFM wire types (pure python)
#   packages/structured-data-models-pql               - the Predictive Query Language (pure python)
#   packages/structured-data-models-connectors        - shared data-source connectors (pure python)
#   packages/structured-data-models-relational-engine - the relational driver (native build)
#
# This file carries only shared tooling configuration. Each package's
# [tool.ruff] extends this file, so the lint baseline below applies to every
# first-party tree and a package cannot drop rules by shadowing the block.

# Not a distribution. This exists so that `uv lock` has a workspace root to
# resolve from, which is what gives nSpect's open source scan something to
# inventory: a Python project with no lock file is reported as having zero
# open source components. Nothing is built or published from here, and
# nothing installs from the lock.
[project]
name = "workspace"
version = "0"
requires-python = ">=3.10"

[tool.uv.workspace]
members = ["packages/*"]

[tool.uv.sources]
structured-data-models-api = { workspace = true }
structured-data-models-pql = { workspace = true }
structured-data-models-client = { workspace = true }
structured-data-models-connectors = { workspace = true }
structured-data-models-relational-engine = { workspace = true }

[tool.ruff]
line-length = 80
target-version = "py310"
# The generated parser is ANTLR's output and the generated bindings are the generator's output, and their bytes are
# recorded in PROVENANCE.json. Letting the formatter rewrite them would break
# that record on every run.
# Prose, not source. A fenced snippet in a bug report is written to be read --
# aligned comments comparing two behaviours, a query left on one line to show
# its shape, and formatting it as a module rewrites the thing being
# illustrated.
extend-exclude = [
    "packages/structured-data-models-pql/src/predictive_query/grammar",
    "packages/structured-data-models-api/src/sdm_api/generated/tfm_api.py",
    "*.md",
]

[tool.ruff.lint]
select = ["E4", "E7", "E9", "F", "I", "B", "UP"]

[tool.ruff.lint.per-file-ignores]
# Each cell imports what it uses so it can be rerun on its own (E402). The
# names Databricks injects are declared in examples/databricks/ruff.toml.
"examples/databricks/*.ipynb" = ["E402"]

[tool.ruff.format]
quote-style = "single"

# Static type checking. Every distribution ships `py.typed`, which is a promise
# made to other people's type checkers, so one has to run here, or the
# promise is only as good as the last person who happened to check.
#
# It is introduced the way a checker can be introduced to an existing tree
# without rewriting it: the whole repository is checked, and the modules that
# do not pass yet are listed below with the count they carried on the day the
# job was added. New code is checked from the first commit; the list is a
# backlog to retire, not a permanent exemption, and it can only shrink because
# `warn_unused_ignores` and this file's own test will report an entry that is
# no longer needed.
[tool.mypy]
# Deliberately not pinned to the 3.10 floor: numpy's bundled stubs use 3.12
# syntax, and a pinned older version makes mypy reject them before it reads a
# line of this repository. The interpreter CI runs on decides instead.
ignore_missing_imports = true
warn_unused_ignores = true
warn_redundant_casts = true
no_implicit_optional = true
exclude = [
    # ANTLR's output, regenerated from the grammar rather than edited. Excluded
    # from ruff for the same reason; the two lists move together.
    "packages/structured-data-models-pql/src/predictive_query/grammar/",
]

# --- Baseline recorded 2026-09-17 -------------------------------------------
# 37 errors across 11 hand-written modules in the four pure-python packages,
# plus 62 across 21 in the relational engine. Retire an entry by fixing its
# module and deleting its line.
#
# `ignore_errors` is blunt: it hides a *new* error in a listed module as well as
# the recorded ones. That is the cost of adopting a checker on a tree this size
# without rewriting it, and it is why the list is dated and counted rather than
# open-ended. The counts are what make a growing entry visible.
[[tool.mypy.overrides]]
module = [
    "sdm_client.materializers.relational",   # 12
    "sdm_client.request",                    #  5
    # --- arrived with the ranking/forecast decomposition, 2026-09-17 --------
    # decompose.py's eight are all "Item None of X has no attribute Y" on
    # `Aggregation | None` and `DateOffsetRange | None`. They may each be
    # guarded by something mypy cannot see, but they are the shape a real
    # AttributeError takes, and they are in new code, worth reading before
    # this entry is retired rather than after.
    "sdm_client.decompose",                  #  8
    "sdm_client.query",                      #  3
    "sdm_client.wire.keys",                  #  1
    "sdm_client.wire.dtypes",                #  3
    "sdm_client.core.warnings",              #  1
    "sdm_client.adapters.relational",        #  1
    "predictive_query.validator.join_validator",  # 4
    "sdm_connectors.sql",                    #  1
    "sdm_connectors.reader",                 #  1
    # --- relational engine: 62 errors across 21 modules --------------------
    "sdm_relational_engine.rfm.base.table",          # 12
    "sdm_relational_engine.rfm.graph",               # 10
    "sdm_relational_engine.rfm.rfm",                 #  9
    "sdm_relational_engine.rfm.backend.sqlite.sampler",     # 2
    "sdm_relational_engine.rfm.backend.snow.sampler",       # 2
    "sdm_relational_engine.rfm.backend.duckdb.table",       # 1
    "sdm_relational_engine.rfm.backend.duckdb.sampler",     # 2
    "sdm_relational_engine.rfm.backend.databricks.sampler", # 2
    "sdm_relational_engine.client.client",           #  2
    "sdm_relational_engine.client.rfm",              #  1
    "sdm_relational_engine.rfm.backend.sqlite.table",    # 1
    "sdm_relational_engine.rfm.backend.snow.table",      # 1
    "sdm_relational_engine.rfm.backend.local.table",     # 1
    "sdm_relational_engine.rfm.backend.local.sampler",   # 1
    "sdm_relational_engine.rfm.backend.databricks.table",# 1
    # --- arrived with the engine cleanup on main, 2026-09-21 ----------
    "sdm_relational_engine.client.engine_access",    #  2
    "sdm_relational_engine.client.state",            #  1
    "sdm_relational_engine.rfm.inference.payload",   #  1
    "sdm_relational_engine.rfm.inference.task",      #  1
]
ignore_errors = true

# `warn_unused_ignores` is on because a suppression that stopped being needed is
# debt that hides the next real error. But "needed" depends on which version of
# a dependency is installed: the ignores below cover a pydantic v1/v2 fallback
# import and two third-party stubs, so they are unused under one resolution and
# required under another. Silencing only `unused-ignore` here keeps every other
# check running in these modules, rather than exempting them wholesale.
[[tool.mypy.overrides]]
module = [
    "predictive_query.schema.typing",
    "predictive_query.schema.json_serde",
    "predictive_query.ast.date_offset_range",
    "predictive_query.query",
    "sdm_client.core.transport",
]
disable_error_code = ["unused-ignore"]

# Coverage is measured to be read, not to be hit. There is no percentage gate
# here on purpose: a number can be raised by tests that assert nothing, and the
# useful signal is which branches no test reaches at all. Branch coverage,
# because line coverage counts an `if` as covered when only one side ever runs.
#
# What it found when it was first turned on, and what the number is for:
# `predictive_query.rewriting`: 401 statements, exported from that package's
# `__all__`, reached by no test and imported by nothing in this repository.
[tool.coverage.run]
branch = true
source_pkgs = ["sdm_api", "sdm_client", "sdm_connectors", "predictive_query"]
omit = [
    # ANTLR's output. Excluded from ruff and mypy for the same reason.
    "*/predictive_query/grammar/*",
]

[tool.coverage.report]
skip_covered = true
exclude_also = [
    "if TYPE_CHECKING:",
    "raise NotImplementedError",
    "@overload",
]


# Registered here as well as per package, so that running the whole workspace
# from the repository root neither warns about unknown marks nor silently
# selects a test that opens a socket.
[tool.pytest.ini_options]
markers = [
    "live_nim: needs a running NIM",
    "live_warehouse: needs a real Snowflake or Databricks warehouse",
    "socket: binds a local socket; deselect where that is not permitted",
    "gpu: needs a CUDA GPU",
]