"""Write the star schema as Parquet (typed as the ClickHouse DDL declares it) and as CSVWithNames with \\N for NULL.""" from __future__ import annotations import csv import hashlib from pathlib import Path import numpy as np import pandas as pd import pyarrow as pa import pyarrow.parquet as pq MONEY = pa.decimal128(18, 2) RATE = pa.decimal128(5, 4) # table -> [(column, arrow type, scale of the integer it is held as: 2 for cents, 4 for basis points, None for as is)] SCHEMA: dict[str, list[tuple[str, pa.DataType, int | None]]] = { "dim_date": [("date_key", pa.uint32(), None), ("date", pa.date32(), None), ("year", pa.uint16(), None), ("quarter", pa.uint8(), None), ("quarter_name", pa.string(), None), ("month", pa.uint8(), None), ("month_name", pa.string(), None), ("year_month", pa.uint32(), None), ("day_of_week", pa.uint8(), None), ("day_name", pa.string(), None), ("is_weekend", pa.uint8(), None)], "dim_product": [("product_id", pa.uint32(), None), ("product_name", pa.string(), None), ("subcategory", pa.string(), None), ("category", pa.string(), None), ("brand", pa.string(), None), ("list_price", MONEY, 2), ("unit_cost", MONEY, 2)], "dim_customer": [("customer_id", pa.uint32(), None), ("customer_name", pa.string(), None), ("segment", pa.string(), None), ("country", pa.string(), None), ("region", pa.string(), None), ("city", pa.string(), None), ("signup_date", pa.date32(), None)], "dim_store": [("store_id", pa.uint16(), None), ("store_name", pa.string(), None), ("channel", pa.string(), None), ("country", pa.string(), None), ("region", pa.string(), None), ("city", pa.string(), None)], "orders": [("order_id", pa.uint32(), None), ("order_date", pa.date32(), None), ("customer_id", pa.uint32(), None), ("store_id", pa.uint16(), None)], "order_lines": [("order_line_id", pa.uint32(), None), ("order_id", pa.uint32(), None), ("date_key", pa.uint32(), None), ("order_date", pa.date32(), None), ("customer_id", pa.uint32(), None), ("store_id", pa.uint16(), None), ("product_id", pa.uint32(), None), ("quantity", pa.uint16(), None), ("unit_price", MONEY, 2), ("discount_pct", RATE, 4), ("net_amount", MONEY, 2), ("unit_cost", MONEY, 2), ("cost_amount", MONEY, 2)], "returns": [("return_id", pa.uint32(), None), ("order_line_id", pa.uint32(), None), ("order_id", pa.uint32(), None), ("product_id", pa.uint32(), None), ("customer_id", pa.uint32(), None), ("store_id", pa.uint16(), None), ("sale_date", pa.date32(), None), ("sale_date_key", pa.uint32(), None), ("return_date", pa.date32(), None), ("return_date_key", pa.uint32(), None), ("returned_qty", pa.uint16(), None), ("refund_amount", MONEY, 2), ("reason", pa.string(), None)], } NULLABLE = {("dim_customer", "segment"), ("orders", "customer_id"), ("order_lines", "customer_id"), ("returns", "customer_id")} def _decimal_array(ints: np.ndarray, typ: pa.DataType) -> pa.Array: """Exact decimals from whole numbers of the smallest unit (cents, basis points): a decimal128 is that integer, stored as 16 little-endian bytes, read at the type's scale.""" v = np.ascontiguousarray(ints, dtype=np.int64) wide = np.empty((len(v), 2), dtype="> 63 return pa.Array.from_buffers(typ, len(v), [None, pa.py_buffer(wide.tobytes())]) def _decimal_text(ints: np.ndarray, scale: int) -> np.ndarray: v = np.asarray(ints, dtype=np.int64) unit = 10 ** scale whole, frac = np.abs(v) // unit, np.abs(v) % unit sign = np.where(v < 0, "-", "") return np.char.add(np.char.add(np.char.add(sign, whole.astype(str)), "."), np.char.zfill(frac.astype(str), scale)) def to_arrow(name: str, df: pd.DataFrame) -> pa.Table: cols, fields = [], [] for col, typ, scale in SCHEMA[name]: s = df[col] nullable = (name, col) in NULLABLE if not nullable and s.isna().any(): raise ValueError(f"{name}.{col} has NULLs but is declared NOT NULL") if scale is not None: arr = _decimal_array(s.to_numpy(dtype=np.int64), typ) elif pa.types.is_date32(typ): arr = pa.array(pd.to_datetime(s).dt.date, type=typ) elif pa.types.is_string(typ): arr = pa.array(s.astype(object).where(s.notna(), None).tolist(), type=typ) else: arr = pa.array(s.astype("Int64").to_numpy(dtype=object, na_value=None) if nullable else s.to_numpy(dtype=np.int64), type=typ) cols.append(arr) fields.append(pa.field(col, typ, nullable=nullable)) return pa.Table.from_arrays(cols, schema=pa.schema(fields)) def to_csv_frame(name: str, df: pd.DataFrame) -> pd.DataFrame: """The table as the text CSVWithNames carries: decimals at their scale, ISO dates, whole numbers without a point.""" out = {} for col, typ, scale in SCHEMA[name]: s = df[col] if scale is not None: out[col] = _decimal_text(s.to_numpy(dtype=np.int64), scale) elif pa.types.is_date32(typ): out[col] = pd.to_datetime(s).dt.strftime("%Y-%m-%d") elif pa.types.is_string(typ): out[col] = s else: out[col] = s.astype("Int64") return pd.DataFrame(out) def write(tables: dict[str, pd.DataFrame], parquet_dir: Path, csv_dir: Path) -> dict[str, dict]: """Writes both formats and returns, per file, its rows, bytes and SHA-256.""" parquet_dir.mkdir(parents=True, exist_ok=True) csv_dir.mkdir(parents=True, exist_ok=True) manifest: dict[str, dict] = {} for name in SCHEMA: t = to_arrow(name, tables[name]) pq.write_table(t, parquet_dir / f"{name}.parquet", compression="zstd", row_group_size=1_000_000, write_statistics=True, use_dictionary=True) path = csv_dir / f"{name}.csv" to_csv_frame(name, tables[name]).to_csv(path, index=False, na_rep="\\N", quoting=csv.QUOTE_MINIMAL, lineterminator="\n", encoding="utf-8") for p in (parquet_dir / f"{name}.parquet", path): manifest[f"{p.parent.name}/{p.name}"] = {"rows": t.num_rows, "bytes": p.stat().st_size, "sha256": sha256(p)} return manifest def sha256(path: Path) -> str: h = hashlib.sha256() with path.open("rb") as fh: for chunk in iter(lambda: fh.read(1 << 20), b""): h.update(chunk) return h.hexdigest()