mariano/soleprint
Public Access
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dataconvert updates

Browse Source
This commit is contained in:
buenosairesam
2026-09-16 09:33:03 -03:00
parent 5391f50755
commit 24aeadde83
8 changed files with 547 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
soleprint/station/tools/dataconvert/README.md
View File

@@ -7,6 +7,7 @@ plus a `SCHEMA.md` describing every table.
uv run dataconvert.py --input data/ --out-dir seed/ # full seeds
uv run dataconvert.py --input data/ "*.xlsx" --out-dir seed/ # dirs, files, globs, .zip
uv run dataconvert.py --input data/ --out-dir sample/ --max-rows 20 # to understand the data
uv run dataconvert.py --input data/ --out-dir seed/ --sql-format batch --max-file-size 100M # a full dump
uv run dataconvert.py --input export.xlsx --header-row 2 --data-row 6
uv run dataconvert.py --input data/ --config their-exports.json
```
@@ -29,8 +30,12 @@ produced the files. Start from `dataconvert-example.json`.
```json
{
"out_dir": "sample",
"max_rows": 20,
"out_dir": "seed",
"all_text": true,
"keep_folders": true,
"exclude": ["*.xlsx.ods"],
"sql_format": "batch",
"max_file_size": "100M",
"schema": true,
"header_row": 1,
"layouts": [
@@ -44,8 +49,11 @@ produced the files. Start from `dataconvert-example.json`.
Every key is optional.
- **out_dir, max_rows, schema** are the run settings: the same as `--out-dir`,
`--max-rows` and `--no-schema`, and a flag on the command line wins over the file.
- **out_dir, max_rows, schema, sql_format, batch_rows, max_file_size, all_text,
keep_folders, exclude** are the run settings: the same as `--out-dir`, `--max-rows`,
`--no-schema`, `--sql-format`, `--batch-rows`, `--max-file-size`, `--all-text`,
`--keep-folders` and `--exclude`, and a flag on the command line wins over the file.
`--exclude` patterns are added to the file's.
There is no default output directory: without `--out-dir` or an `out_dir` in the
config, the run stops before writing anything.
`out_dir` is relative to the folder the config is in, so `"sample"` beside the data
@@ -67,6 +75,65 @@ such as `max_row` is not silently ignored.
For a one-off, `--header-row 2 --data-row 6` applies one layout to every file in the
run and skips the config's layouts.
## Values as they are: all_text
By default pandas guesses a type per column, and for real exports that guess does
damage: a site code `0012` becomes `12`, `NA` becomes NULL, and a column that is
numbers for the first 200,000 rows and text after that is typed differently in
different rows. `all_text` reads every cell as the string in the file. Only an empty
cell becomes NULL; `NA`, `N/A` and `null` stay text. Converting types is then the
loader's job, where the real schema is decided.
To help decide it, `SCHEMA.md` describes what each text column looks like:
| type column says | meaning |
|---|---|
| `text, integer-like (max 6)` | every value is a whole number |
| `text, code with leading zeros (max 4)` | whole numbers, some with leading zeros: keep as text |
| `text, decimal-like` / `timestamp-like (YYYY-MM-DD hh:mm)` / `date-like (DDMONYYYY)` / `boolean-like` | every value has that shape |
| `text, mostly date-like (DDMONYYYY): 300 others` | 95% or more fit; the others are what needs a rule (partial dates, `UNK`) |
| `text, 67% integer-like` | at least half fit |
| `text (max 40)` | no common shape |
Shapes are judged on up to 200,000 values spread through the column.
## Folders and exclusions
Output is flat by default: a table's file is named after the file or sheet, and two
source folders with the same table name (two studies' `drm_lb.csv`) write into the
same file. `keep_folders` mirrors the source folders under `out_dir` instead, so each
keeps its own `.sql`; `SCHEMA.md` stays one document at the top, with the folder in
each table's heading.
`exclude` skips files by glob. A pattern with no `/` matches the file name at any
depth (`"*.xlsx.ods"`); one with a `/` matches the path under the input folder
(`"Old/*"`). Each skipped file is named in the run's output.
## Full dumps: format and splitting
One `INSERT` per row repeats the table and column list on every line, so the SQL can
be many times the size of the spreadsheets, which are compressed on disk to begin
with. `sql_format` picks how rows are written for PostgreSQL:
| format | shape | load with | re-runnable | size |
|---|---|---|---|---|
| `insert` (default) | one `INSERT ... ON CONFLICT DO NOTHING` per row | any client | yes | largest |
| `batch` | one `INSERT` per `batch_rows` rows (500), same `ON CONFLICT` | any client | yes | about a quarter |
| `copy` | `COPY ... FROM stdin`, tab-separated | `psql` only | no: a second load fails on duplicate keys | smallest |
`max_file_size` (`"100M"`, decimal k/M/G) splits a table's output into `table.001.sql`,
`table.002.sql`, ... once it passes that size. Splits fall between rows (or batches),
never inside one, and every part is its own transaction, so the parts load one at a time
in name order:
```bash
for f in seed/*.sql; do psql -v ON_ERROR_STOP=1 -f "$f" || break; done
```
A table under the limit keeps its plain `table.sql` name. Rows are written to disk as
they are rendered, so a table larger than memory converts; reading the source still
needs it in memory. Samples (`max_rows`) are never split.
## Sampling for a web LLM
Full seed files get large fast, and a model only needs to see the shape of the data.
@@ -90,8 +157,8 @@ what pandas read: a starting point, not DDL. `--no-schema` skips the file.
| `config.py` | `dataconvert.json`: layouts, sheet naming and run settings |
| `readers.py` | files, directories, ZIPs and globs into DataFrames |
| `progress.py` | progress lines: which file is being read, and how far a long table has got |
| `sqlgen.py` | DataFrames into INSERT statements, with the row cap |
| `output.py` | file naming and writing |
| `sqlgen.py` | rows as `insert`, `batch` or `copy` statements, and size estimates |
| `output.py` | file naming, streaming rows to disk, and splitting into parts |
| `schema.py` | `SCHEMA.md` |
The modules import each other by name, so the folder works wherever it is copied:

85
soleprint/station/tools/dataconvert/config.py
View File

@@ -12,8 +12,10 @@ Start from dataconvert-example.json. With neither, every file is read with its
header on row 1.
{
"out_dir": "sample",
"max_rows": 20,
"out_dir": "seed",
"sql_format": "batch",
"batch_rows": 500,
"max_file_size": "100M",
"schema": true,
"header_row": 1,
"layouts": [
@@ -32,15 +34,21 @@ applies to a sheet (or CSV) when the cell at match.row/match.column, trimmed,
is one of match.in. The first layout that matches wins; a sheet none matches
uses the top-level header_row/data_row, which default to 1 and 2.
The run settings (out_dir, max_rows, schema) are the command line's flags with
the same names, and a flag given on the command line wins over the file. out_dir
is relative to the folder the config file is in. Keys starting with _ are
The run settings (out_dir, max_rows, schema, sql_format, batch_rows,
max_file_size, all_text, keep_folders, exclude) are the command line's flags
with the same names, and a flag given on the command line wins over the file;
--exclude patterns are added to the file's. out_dir is relative to the folder
the config file is in; max_file_size takes a number of bytes or a suffixed size
such as "100M" (decimal: k, M, G). Keys starting with _ are
comments; any other unknown key is an error, so a typo is not silently ignored.
"""
import json
import re
from pathlib import Path
from sqlgen import FORMATS
class ConfigError(Exception):
pass
@@ -86,7 +94,9 @@ class Rule:
class Config:
def __init__(self, rules=(), bare_sheet_prefixes=(), source=None, forced=None,
fallback=None, max_rows=None, schema=None, out_dir=None):
fallback=None, max_rows=None, schema=None, out_dir=None,
sql_format=None, batch_rows=None, max_file_size=None,
all_text=None, keep_folders=None, exclude=()):
self.rules = list(rules)
self.bare_sheet_prefixes = tuple(bare_sheet_prefixes)
self.source = source
@@ -99,6 +109,12 @@ class Config:
self.max_rows = max_rows
self.schema = schema
self.out_dir = out_dir
self.sql_format = sql_format
self.batch_rows = batch_rows
self.max_file_size = max_file_size
self.all_text = all_text
self.keep_folders = keep_folders
self.exclude = tuple(exclude)
@property
def probe_rows(self):
@@ -121,6 +137,18 @@ class Config:
said.append(f"max_rows {self.max_rows}")
if self.schema is not None:
said.append(f"schema {'on' if self.schema else 'off'}")
if self.sql_format is not None:
said.append(f"sql_format {self.sql_format}")
if self.batch_rows is not None:
said.append(f"batch_rows {self.batch_rows}")
if self.max_file_size is not None:
said.append(f"max_file_size {self.max_file_size} bytes")
if self.all_text:
said.append("all_text")
if self.keep_folders:
said.append("keep_folders")
if self.exclude:
said.append(f"exclude {', '.join(self.exclude)}")
if self.fallback is not None:
said.append(f"header row {self.fallback.header_row}, data from row {self.fallback.data_row}")
if self.rules:
@@ -129,7 +157,24 @@ class Config:
FILENAME = "dataconvert.json"
KNOWN_KEYS = {"layouts", "bare_sheet_prefixes", "max_rows", "schema", "out_dir", "header_row", "data_row"}
KNOWN_KEYS = {"layouts", "bare_sheet_prefixes", "max_rows", "schema", "out_dir", "header_row", "data_row",
"sql_format", "batch_rows", "max_file_size", "all_text", "keep_folders", "exclude"}
def parse_size(value):
"""Bytes from 100000000, "100000000" or "100M". Decimal, as distill counts: 1k is 1000."""
if isinstance(value, bool):
raise ValueError(value)
if isinstance(value, int):
n = value
else:
m = re.fullmatch(r"\s*(\d+)\s*([kKmMgG]?)[bB]?\s*", str(value))
if not m:
raise ValueError(value)
n = int(m[1]) * {"": 1, "k": 1000, "m": 1000 ** 2, "g": 1000 ** 3}[m[2].lower()]
if n < 1:
raise ValueError(value)
return n
def source_config_path(in_path: Path):
@@ -175,6 +220,28 @@ def load(path=None, forced=None):
raise ConfigError(f"{path}: out_dir must be a path")
out_dir = path.parent / Path(out_dir).expanduser()
sql_format = raw.get("sql_format")
if sql_format is not None and sql_format not in FORMATS:
raise ConfigError(f"{path}: sql_format must be one of {', '.join(FORMATS)}")
batch_rows = raw.get("batch_rows")
if batch_rows is not None and (isinstance(batch_rows, bool) or not isinstance(batch_rows, int) or batch_rows < 1):
raise ConfigError(f"{path}: batch_rows must be a whole number, 1 or more")
max_file_size = raw.get("max_file_size")
if max_file_size is not None:
try:
max_file_size = parse_size(max_file_size)
except ValueError:
raise ConfigError(f'{path}: max_file_size must be a size such as 100000000 or "100M"')
flags = {}
for key in ("all_text", "keep_folders"):
flags[key] = raw.get(key)
if flags[key] is not None and not isinstance(flags[key], bool):
raise ConfigError(f"{path}: {key} must be true or false")
exclude = raw.get("exclude", [])
if not isinstance(exclude, list) or not all(isinstance(p, str) and p for p in exclude):
raise ConfigError(f'{path}: exclude must be a list of patterns such as "*.xlsx.ods"')
fallback = None
if "header_row" in raw or "data_row" in raw:
try:
@@ -183,4 +250,6 @@ def load(path=None, forced=None):
raise ConfigError(f"{path}: header_row and data_row must be row numbers")
return Config(rules, [str(p) for p in prefixes], source=path, forced=forced,
fallback=fallback, max_rows=max_rows, schema=schema, out_dir=out_dir)
fallback=fallback, max_rows=max_rows, schema=schema, out_dir=out_dir,
sql_format=sql_format, batch_rows=batch_rows, max_file_size=max_file_size,
all_text=flags["all_text"], keep_folders=flags["keep_folders"], exclude=exclude)

14
soleprint/station/tools/dataconvert/dataconvert-example.json
View File

@@ -1,9 +1,17 @@
{
"_comment": "Template for one set of spreadsheets. Copy it to dataconvert.json in the folder that holds them, where it is picked up automatically, or pass any file with --config FILE. Every key is optional; keys starting with _ are comments, any other unknown key is an error. Rows and columns count from 1, as the spreadsheet shows them.",
"_run": "The same settings as the command-line flags, and a flag given there wins. out_dir is relative to this file's folder. max_rows samples each table (null for all rows). schema false skips SCHEMA.md.",
"out_dir": "sample",
"max_rows": 20,
"_run": "The same settings as the command-line flags, and a flag given there wins. out_dir is relative to this file's folder. max_rows samples each table (leave it out for all rows). schema false skips SCHEMA.md. sql_format: insert (a statement per row, default), batch (one INSERT per batch_rows rows), or copy (COPY FROM stdin: smallest, psql only, load once). max_file_size splits a table into numbered parts past that size, e.g. \"100M\".",
"out_dir": "seed",
"_values": "all_text reads every cell as the string in the file (only empty cells become NULL), so codes keep leading zeros and NA stays NA; SCHEMA.md then says what each column looks like. keep_folders mirrors the source folders under out_dir, so same-named tables from different folders do not share a file. exclude skips files by glob: no / matches the file name at any depth, with a / the path under the input folder.",
"all_text": true,
"keep_folders": true,
"exclude": ["*.xlsx.ods"],
"sql_format": "batch",
"batch_rows": 500,
"max_file_size": "100M",
"schema": true,
"_rows": "Where the column names and the data are, for sheets no layout below matches. Defaults: 1 and the row after it.",

35
soleprint/station/tools/dataconvert/dataconvert.py
View File

@@ -28,12 +28,20 @@ import argparse
from pathlib import Path
from output import write_tables
from sqlgen import DEFAULT_BATCH_ROWS, FORMATS
from progress import say
import config as cfg
from readers import expand_inputs, iter_sources
from schema import SchemaReport
def size(value):
try:
return cfg.parse_size(value)
except ValueError:
raise argparse.ArgumentTypeError('a size such as 100000000 or "100M"')
def positive_int(value):
n = int(value)
if n < 1:
@@ -49,6 +57,22 @@ def main():
parser.add_argument("--max-rows", type=positive_int, default=None,
help="Write at most N rows per table; the header and SCHEMA.md note the full row count and size (default: the config's max_rows, else all)")
parser.add_argument("--no-schema", action="store_true", help="Do not write SCHEMA.md, whatever the config says")
parser.add_argument("--sql-format", choices=FORMATS, default=None,
help="insert: a statement per row (default). batch: one INSERT per --batch-rows rows. "
"copy: COPY FROM stdin, smallest, psql only, load once")
parser.add_argument("--batch-rows", type=positive_int, default=None,
help=f"Rows per INSERT with --sql-format batch (default {DEFAULT_BATCH_ROWS})")
parser.add_argument("--max-file-size", type=size, default=None,
help='Split a table into numbered parts past this size, e.g. "100M" (default: never split; samples are never split)')
parser.add_argument("--all-text", action="store_true", default=None,
help="Read every column as text, exactly as in the file; only empty cells become NULL. "
"Types are then the loader's business; SCHEMA.md says what each column looks like")
parser.add_argument("--keep-folders", action="store_true", default=None,
help="Mirror the source folders under the output directory, instead of one flat folder "
"where same-named tables from different folders share a file")
parser.add_argument("--exclude", action="append", default=[], metavar="GLOB",
help='Skip matching files, e.g. "*.xlsx.ods"; a pattern with no / matches the file name '
"at any depth. Repeatable, and added to the config's")
parser.add_argument("--header-row", type=positive_int, default=1,
help="Spreadsheet row holding the column names, for every file; overrides the config's layouts (default 1)")
parser.add_argument("--data-row", type=positive_int, default=None,
@@ -107,6 +131,12 @@ def main():
out_dir = Path(args.out_dir) if args.out_dir else config.out_dir
max_rows = args.max_rows if args.max_rows is not None else config.max_rows
schema = not args.no_schema and config.schema is not False
fmt = args.sql_format or config.sql_format or "insert"
batch_rows = args.batch_rows or config.batch_rows or DEFAULT_BATCH_ROWS
max_bytes = args.max_file_size or config.max_file_size
all_text = bool(args.all_text or config.all_text)
keep_folders = bool(args.keep_folders or config.keep_folders)
exclude = tuple(config.exclude) + tuple(args.exclude)
key = out_dir.resolve()
if key not in reports:
@@ -114,8 +144,9 @@ def main():
_, report, caps = reports[key]
caps.add(max_rows)
for source_name, dfs in iter_sources(in_path, config):
write_tables(dfs, out_dir, source_name, max_rows, report if schema else None, config.bare_sheet_prefixes)
for source_name, folder, dfs in iter_sources(in_path, config, all_text, exclude):
write_tables(dfs, out_dir, source_name, max_rows, report if schema else None, config.bare_sheet_prefixes,
fmt, batch_rows, max_bytes, folder if keep_folders else "")
for out_dir, report, caps in reports.values():
if report.entries and out_dir.is_dir():

132
soleprint/station/tools/dataconvert/output.py
View File

@@ -1,11 +1,14 @@
"""
Output: one .sql file per table or sheet, named after it.
Output: one .sql file per table or sheet, named after it, split into numbered
parts when a size limit is set and the table outgrows it.
"""
import glob
import re
from pathlib import Path
from progress import say
from sqlgen import render_table, sanitize_identifier
from sqlgen import DEFAULT_BATCH_ROWS, Shape, render_sample, sanitize_identifier, utf8_len
def table_filename(raw_name: str, source_name: str, sheet_count: int, bare_prefixes=()) -> str:
@@ -20,28 +23,127 @@ def table_filename(raw_name: str, source_name: str, sheet_count: int, bare_prefi
return f"{clean}.sql"
def write_tables(dfs: dict, out_dir: Path, source_name: str, max_rows=None, report=None, bare_prefixes=()):
"""Write individual .sql files per table/sheet into the output directory."""
class TableFile:
"""
Where one table's statements go: <table>.sql, or once it passes max_bytes,
<table>.001.sql, <table>.002.sql, ... in load order.
Several sources can feed the same table and they accumulate: writing resumes
at the last part that exists, as the unsplit file always did.
"""
def __init__(self, out_dir: Path, filename: str, max_bytes=None):
self.out_dir = out_dir
self.stem = filename[:-len(".sql")]
self.max_bytes = max_bytes
self.handle = None
self.touched = []
parts = self.parts()
self.index = int(parts[-1].name[len(self.stem) + 1:-len(".sql")]) if parts else 0
self.size = self.path.stat().st_size if self.path.exists() else 0
@property
def path(self) -> Path:
name = f"{self.stem}.sql" if self.index == 0 else f"{self.stem}.{self.index:03d}.sql"
return self.out_dir / name
def parts(self):
pattern = re.compile(re.escape(self.stem) + r"\.\d{3}\.sql$")
if not self.out_dir.is_dir():
return []
return sorted(p for p in self.out_dir.glob(glob.escape(self.stem) + ".*.sql") if pattern.match(p.name))
def write(self, text: str):
if self.handle is None:
# Created on the first write, so a run that finds nothing to
# convert leaves no empty directory behind.
self.out_dir.mkdir(parents=True, exist_ok=True)
self.handle = open(self.path, "a", encoding="utf-8")
if self.path not in self.touched:
self.touched.append(self.path)
self.handle.write(text)
self.size += utf8_len(text)
def full(self, adding: int) -> bool:
return self.max_bytes is not None and self.size > 0 and self.size + adding > self.max_bytes
def next_part(self):
"""Move on to the next numbered part; the unnumbered file becomes part 001."""
self.close()
if self.index == 0:
base, first = self.path, self.out_dir / f"{self.stem}.001.sql"
base.rename(first)
self.touched = [first if p == base else p for p in self.touched]
self.index = 1
self.index += 1
self.size = 0
def close(self):
if self.handle is not None:
self.handle.close()
self.handle = None
def write_full(shape: Shape, df, table_file: TableFile) -> int:
"""Every row, streamed; a new part, as its own transaction, whenever the limit is reached."""
opening, closing = shape.open(), shape.close()
open_bytes, close_bytes = utf8_len(opening), utf8_len(closing)
if table_file.full(open_bytes + close_bytes):
table_file.next_part()
table_file.write(opening)
written = open_bytes
units_in_part = 0
for unit in shape.units(df):
size = utf8_len(unit)
# Only between units, and never leaving a part empty: a single unit
# bigger than the limit gets a part of its own instead of looping.
if units_in_part and table_file.full(size + close_bytes):
table_file.write(closing)
table_file.next_part()
table_file.write(opening)
written += close_bytes + open_bytes
units_in_part = 0
table_file.write(unit)
written += size
units_in_part += 1
table_file.write(closing)
return written + close_bytes
def write_tables(dfs: dict, out_dir: Path, source_name: str, max_rows=None, report=None, bare_prefixes=(),
fmt="insert", batch_rows=DEFAULT_BATCH_ROWS, max_bytes=None, folder=""):
"""
Write individual .sql files per table/sheet into the output directory, or
into its `folder` subdirectory when the source folders are kept.
"""
target = out_dir / folder if folder else out_dir
for raw_name, df in dfs.items():
if df.empty:
continue
table = sanitize_identifier(raw_name)
filename = table_filename(raw_name, source_name, len(dfs), bare_prefixes)
sql, total, full_bytes, exact = render_table(df, table, max_rows)
out_file = out_dir / filename
# Created on the first file, so a run that finds nothing to convert
# leaves no empty directory behind.
out_dir.mkdir(parents=True, exist_ok=True)
shape = Shape(fmt, table, df.columns, batch_rows)
total = len(df)
exact = max_rows is None or max_rows >= total
# A sample is never split: it is small, and it is not for loading.
table_file = TableFile(target, filename, max_bytes if exact else None)
# Several sources can feed the same table; they accumulate in one file.
mode = "a" if out_file.exists() else "w"
with open(out_file, mode, encoding="utf-8") as f:
f.write(sql)
try:
if exact:
full_bytes = write_full(shape, df, table_file)
else:
text, full_bytes = render_sample(shape, df, max_rows)
table_file.write(text)
finally:
table_file.close()
shown = total if exact else max_rows
parts = len(table_file.parts()) or 1
if report is not None:
report.add(source_name, table, filename, df, total, full_bytes, exact, shown)
report.add(source_name, table, f"{folder}/{filename}" if folder else filename,
df, total, full_bytes, exact, shown, parts)
suffix = "" if exact else f" ({shown} of {total} rows)"
say(f"Generated: {out_file}{suffix}")
for path in table_file.touched:
say(f"Generated: {path}{suffix}")

60
soleprint/station/tools/dataconvert/readers.py
View File

@@ -1,16 +1,18 @@
"""
Readers: files, directories, ZIP archives and wildcards into DataFrames.
Each input becomes zero or more (source_name, {entity_name: DataFrame}) pairs,
one per spreadsheet or CSV. Nothing here knows about SQL or output files.
Each input becomes zero or more (source_name, folder, {entity_name: DataFrame})
triples, one per spreadsheet or CSV, where folder is where the file sits relative
to the input folder. Nothing here knows about SQL or output files.
"""
import fnmatch
import glob
import os
import tempfile
import time
import zipfile
from pathlib import Path
from pathlib import Path, PurePosixPath
import pandas as pd
@@ -19,6 +21,20 @@ from sqlgen import human_bytes
SUPPORTED = {".csv", ".xlsx", ".xls", ".ods"}
# all_text: every cell as the string in the file. Only an empty cell is missing;
# pandas' default list would also turn "NA", "N/A", "null" and "nan" into NULL,
# and in real data "NA" is as often a value as an absence.
TEXT_READ = dict(dtype=str, keep_default_na=False, na_values=[""])
def excluded(rel_path: str, patterns) -> bool:
"""
A pattern with a / is matched against the path under the input folder; one
without, against the file name alone, so "*.xlsx.ods" works at any depth.
"""
name = rel_path.rsplit("/", 1)[-1]
return any(fnmatch.fnmatch(rel_path if "/" in p else name, p) for p in patterns)
def expand_inputs(inputs):
"""Wildcard patterns become the paths they match; everything else passes through."""
@@ -59,10 +75,11 @@ def read_sheet(read, config):
return data, layout
def load_dataframes_from_file(file_path: Path, config) -> dict:
def load_dataframes_from_file(file_path: Path, config, all_text=False) -> dict:
"""Load a file (.csv, .xlsx, .xls, .ods) into {entity_name: DataFrame}."""
ext = file_path.suffix.lower()
dfs = {}
extra = TEXT_READ if all_text else {}
# Said before the read, not after: a big workbook can take minutes inside
# pandas, and this line is what says which file that is.
@@ -74,7 +91,7 @@ def load_dataframes_from_file(file_path: Path, config) -> dict:
# chunk, so one column can come out as numbers in one chunk and
# text in the next, quoted differently row to row in the SQL. Read
# whole, each column gets one type. Costs memory on huge files.
df, layout = read_sheet(lambda **kw: pd.read_csv(file_path, low_memory=False, **kw), config)
df, layout = read_sheet(lambda **kw: pd.read_csv(file_path, low_memory=False, **extra, **kw), config)
dfs[file_path.stem] = df
say(f" {len(df)} rows{layout_note(layout)} ({seconds(start)})")
elif ext in [".xlsx", ".xls", ".ods"]:
@@ -82,7 +99,7 @@ def load_dataframes_from_file(file_path: Path, config) -> dict:
for sheet in xls.sheet_names:
start = time.monotonic()
df, layout = read_sheet(
lambda sheet=sheet, **kw: pd.read_excel(xls, sheet_name=sheet, **kw), config)
lambda sheet=sheet, **kw: pd.read_excel(xls, sheet_name=sheet, **extra, **kw), config)
dfs[sheet] = df
say(f" sheet {sheet}: {len(df)} rows{layout_note(layout)} ({seconds(start)})")
except Exception as e:
@@ -97,20 +114,37 @@ def layout_note(layout):
return f", layout '{layout.name}' (header row {layout.header_row}, data from row {layout.data_row})"
def iter_sources(path: Path, config):
"""Yield (source_name, dfs) for a file, a directory (recursively) or a ZIP archive."""
def iter_sources(path: Path, config, all_text=False, exclude=(), prefix=""):
"""
Yield (source_name, folder, dfs) for a file, a directory (recursively) or a
ZIP archive given as an input. folder is the file's directory under the
input folder (or inside the ZIP), "" at the top.
"""
if path.is_file() and path.suffix.lower() == ".zip":
if excluded(prefix + path.name, exclude):
say(f"skipping {prefix + path.name} (excluded)")
return
with tempfile.TemporaryDirectory() as tmp_dir:
with zipfile.ZipFile(path, "r") as zip_ref:
zip_ref.extractall(tmp_dir)
yield from iter_sources(Path(tmp_dir), config)
yield from iter_sources(Path(tmp_dir), config, all_text, exclude, prefix)
elif path.is_dir():
for root, _, files in os.walk(path):
for root, dirs, files in os.walk(path):
dirs.sort()
for f in sorted(files):
f_path = Path(root) / f
if f_path.suffix.lower() in SUPPORTED:
yield f_path.stem, load_dataframes_from_file(f_path, config)
if f_path.suffix.lower() not in SUPPORTED:
continue
rel = prefix + f_path.relative_to(path).as_posix()
if excluded(rel, exclude):
say(f"skipping {rel} (excluded)")
continue
folder = str(PurePosixPath(rel).parent) if "/" in rel else ""
yield f_path.stem, folder, load_dataframes_from_file(f_path, config, all_text)
elif path.is_file() and path.suffix.lower() in SUPPORTED:
yield path.stem, load_dataframes_from_file(path, config)
if excluded(prefix + path.name, exclude):
say(f"skipping {prefix + path.name} (excluded)")
return
yield path.stem, prefix.rstrip("/"), load_dataframes_from_file(path, config, all_text)

73
soleprint/station/tools/dataconvert/schema.py
View File

@@ -5,8 +5,14 @@ Written for reading, by a person or a web LLM that has to understand the data
before anything else: columns, an inferred type, how many are empty, one
example value, and how big the table really is. Types are inferred from what
pandas read, so they are a starting point, not a DDL.
Text columns, which is every column with all_text, are described by what their
values look like: integer-like, a code with leading zeros, a date in a given
format. Nothing is converted; the point is to have the evidence in one place
when the real schema is decided, in the loader.
"""
import re
from pathlib import Path
import pandas as pd
@@ -14,6 +20,52 @@ import pandas as pd
from sqlgen import human_bytes, sanitize_identifier
EXAMPLE_MAX = 40
# Text shapes are judged on at most this many values, spread through the column.
SHAPE_SAMPLE = 200_000
# A shape that fits this share of the values is reported as "mostly".
MOSTLY = 0.95
SHAPES = (
# (name, pattern), most specific first.
("integer-like", r"[+-]?\d+"),
("decimal-like", r"[+-]?(?:\d+\.\d*|\.\d+|\d+)(?:[eE][+-]?\d+)?"),
("timestamp-like (YYYY-MM-DD hh:mm)", r"\d{4}-\d{2}-\d{2}[ T]\d{2}:\d{2}(?::\d{2}(?:\.\d+)?)?(?:Z|[+-]\d{2}:?\d{2})?"),
("date-like (YYYY-MM-DD)", r"\d{4}-\d{2}-\d{2}"),
("date-like (DDMONYYYY)", r"\d{1,2}[A-Za-z]{3}\d{4}"),
("timestamp-like (DDMONYYYY:hh:mm)", r"\d{1,2}[A-Za-z]{3}\d{4}[: ]\d{2}:\d{2}(?::\d{2})?"),
("date-like (D/M/Y or M/D/Y)", r"\d{1,2}/\d{1,2}/\d{2,4}"),
("boolean-like", r"(?i:y|n|yes|no|true|false)"),
)
def text_shape(values: pd.Series) -> str:
"""What a column of strings looks like, for choosing its real type later."""
text = values.astype(str)
longest = int(text.str.len().max())
if len(text) > SHAPE_SAMPLE:
text = text.iloc[:: len(text) // SHAPE_SAMPLE]
stripped = text.str.strip()
stripped = stripped[stripped != ""]
if stripped.empty:
return f"text, blank (max {longest})"
best = None
for name, pattern in SHAPES:
fits = stripped.str.fullmatch(pattern)
share = fits.mean()
if name == "integer-like" and share and stripped[fits].str.fullmatch(r"[+-]?0\d+").any():
name = "code with leading zeros"
if share == 1:
return f"text, {name} (max {longest})"
if share >= MOSTLY:
others = int((~fits).sum())
return f"text, mostly {name}: {others} other{'s' if others != 1 else ''} (max {longest})"
if best is None or share > best[1]:
best = (name, share)
# Half or more is still worth knowing when the type is being decided: a
# visit number that is "UNS" a third of the time is a decision, not text.
if best is not None and best[1] >= 0.5:
return f"text, {best[1]:.0%} {best[0]} (max {longest})"
return f"text (max {longest})"
def infer_type(series: pd.Series) -> str:
@@ -36,8 +88,9 @@ def infer_type(series: pd.Series) -> str:
return "numeric"
if kinds <= {"datetime", "Timestamp"}:
return "timestamp"
longest = values.astype(str).str.len().max()
return f"text (max {longest})" if kinds == {"str"} else f"mixed ({', '.join(sorted(kinds))})"
if kinds == {"str"}:
return text_shape(values)
return f"mixed ({', '.join(sorted(kinds))})"
def example(series: pd.Series) -> str:
@@ -54,10 +107,11 @@ class SchemaReport:
def __init__(self):
self.entries = []
def add(self, source, table, filename, df, total_rows, full_bytes, exact, shown_rows):
def add(self, source, table, filename, df, total_rows, full_bytes, exact, shown_rows, parts=1):
self.entries.append(dict(
source=source, table=table, filename=filename, df=df,
total_rows=total_rows, full_bytes=full_bytes, exact=exact, shown_rows=shown_rows,
parts=parts,
))
def tables(self):
@@ -77,6 +131,9 @@ class SchemaReport:
m["full_bytes"] += e["full_bytes"]
m["shown_rows"] += e["shown_rows"]
m["exact"] = m["exact"] and e["exact"]
# The part count is read off the disk after each write, so the
# latest one already includes everything written before it.
m["parts"] = max(m["parts"], e["parts"])
return list(merged.values())
def write(self, out_dir: Path, max_rows):
@@ -100,13 +157,19 @@ class SchemaReport:
for t in tables:
size = ("" if t["exact"] else "~") + human_bytes(t["full_bytes"])
kept = "all" if t["exact"] else f"{t['shown_rows']} rows"
lines.append(f"| `{t['table']}` | `{t['filename']}` | {t['total_rows']} | {size} | {kept} |")
file = f"`{t['filename']}`"
if t["parts"] > 1:
stem = t["filename"][:-len(".sql")]
file = f"`{stem}.001.sql` … `{stem}.{t['parts']:03d}.sql` ({t['parts']} parts)"
lines.append(f"| `{t['table']}` | {file} | {t['total_rows']} | {size} | {kept} |")
lines.append("")
for t in tables:
df = t["df"]
sources = ", ".join(f"`{src}`" for src in t["sources"])
lines += [f"## {t['table']}", "", f"From {sources} · {t['total_rows']} rows · {len(df.columns)} columns", ""]
folder = t["filename"].rsplit("/", 1)[0] if "/" in t["filename"] else ""
heading = f"{t['table']} · {folder}" if folder else t["table"]
lines += [f"## {heading}", "", f"From {sources} · {t['total_rows']} rows · {len(df.columns)} columns", ""]
lines += ["| column | source header | type | nulls | example |", "|---|---|---|---:|---|"]
for col in df.columns:
series = df[col]

153
soleprint/station/tools/dataconvert/sqlgen.py
View File

@@ -1,17 +1,35 @@
"""
SQL rendering: DataFrames into schema-agnostic INSERT statements.
With a row cap, only the first rows are rendered, and the size the full output
would have had is estimated from them, so a sample still says how big the real
thing is.
SQL rendering: DataFrames into schema-agnostic PostgreSQL seed statements.
Three shapes, the same rows:
insert one INSERT per row, ON CONFLICT DO NOTHING. Runs in any client and can
be re-run; the column list repeats on every row, so it is the largest.
batch one INSERT per batch_rows rows, still ON CONFLICT DO NOTHING. Same
guarantees as insert, a fraction of the size, and far faster to load.
copy COPY ... FROM stdin, tab-separated. Smallest and fastest, but only psql
runs it, and COPY has no ON CONFLICT: loading it twice fails on the
first duplicate key.
A block is one transaction: open(), units, close(). Units are the pieces a file
may be split between (a row, or a batch), so every part of a split file is a
complete, loadable script. Getting them onto disk is output.py's business, and
nothing here holds a whole table's text, except a capped sample, which is small.
"""
import math
import re
import pandas as pd
from progress import Ticker
FORMATS = ("insert", "batch", "copy")
DEFAULT_BATCH_ROWS = 500
# Rows measured to estimate a full table's size, spread evenly through it: the
# first rows of a table are often shorter (small ids, early dates) than the rest.
MEASURE_ROWS = 1000
def sanitize_identifier(identifier: str) -> str:
"""Sanitize names for SQL tables, columns, and filenames."""
@@ -31,50 +49,107 @@ def sql_value(v) -> str:
return f"'{escaped}'"
def render_table(df: pd.DataFrame, table_name: str, max_rows=None):
"""
Return (sql_text, total_rows, full_bytes, exact).
def copy_value(v) -> str:
"""The same value in COPY's text format: \\N for NULL, backslash escapes."""
if pd.isna(v):
return r"\N"
if isinstance(v, (bool, int)):
return str(v)
if isinstance(v, float):
return str(int(v)) if v.is_integer() else str(v)
return (str(v).replace("\\", "\\\\").replace("\t", "\\t")
.replace("\n", "\\n").replace("\r", "\\r"))
full_bytes is what the file would weigh with every row: measured when every
row was rendered, extrapolated from the average rendered row otherwise.
"""
total = len(df)
if total == 0:
return "", 0, 0, True
def utf8_len(text: str) -> int:
return len(text.encode("utf-8"))
class Shape:
"""How one table's rows are framed in one format."""
def __init__(self, fmt, table_name, columns, batch_rows=DEFAULT_BATCH_ROWS):
if fmt not in FORMATS:
raise ValueError(f"unknown sql format: {fmt}")
self.fmt = fmt
self.table_name = table_name
self.batch_rows = batch_rows
table_ref = f'"{table_name}"'
cols = ", ".join(f'"{sanitize_identifier(c)}"' for c in df.columns)
shown = df if max_rows is None else df.head(max_rows)
rows = []
ticker = Ticker(table_name, len(shown))
# iterrows, not itertuples: it hands values over the way the original tool
# did, and seed files people already load depend on exactly that quoting.
for i, (_, row) in enumerate(shown.iterrows(), 1):
vals = ", ".join(sql_value(v) for v in row)
rows.append(f"INSERT INTO {table_ref} ({cols}) VALUES ({vals}) ON CONFLICT DO NOTHING;\n")
cols = ", ".join(f'"{sanitize_identifier(c)}"' for c in columns)
self.head = f"-- Generated seed data for table: {table_ref}\n"
if fmt == "copy":
self.head += "-- COPY FROM stdin: load with psql. COPY has no ON CONFLICT, so load it once.\n"
self.insert = f"INSERT INTO {table_ref} ({cols}) VALUES"
self.copy = f"COPY {table_ref} ({cols}) FROM stdin;\n"
def open(self, note="") -> str:
text = self.head + note + "BEGIN;\n\n"
return text + self.copy if self.fmt == "copy" else text
def close(self) -> str:
return ("\\.\n" if self.fmt == "copy" else "") + "\nCOMMIT;\n"
def _row(self, values) -> str:
if self.fmt == "copy":
return "\t".join(copy_value(v) for v in values) + "\n"
vals = ", ".join(sql_value(v) for v in values)
if self.fmt == "insert":
return f"{self.insert} ({vals}) ON CONFLICT DO NOTHING;\n"
return f"({vals})"
def _batch(self, rows) -> str:
return f"{self.insert}\n" + ",\n".join(rows) + "\nON CONFLICT DO NOTHING;\n\n"
def units(self, df: pd.DataFrame):
"""The rows as splittable pieces: one per row, or one per batch."""
ticker = Ticker(self.table_name, len(df))
batch = []
# iterrows, not itertuples: it hands values over the way the original
# tool did, and seed files people already load depend on that quoting.
for i, (_, row) in enumerate(df.iterrows(), 1):
text = self._row(row)
if self.fmt == "batch":
batch.append(text)
if len(batch) == self.batch_rows:
yield self._batch(batch)
batch = []
else:
yield text
if i % Ticker.CHECK == 0:
ticker.tick(i)
if batch:
yield self._batch(batch)
head = f"-- Generated seed data for table: {table_ref}\n"
begin, commit = "BEGIN;\n\n", "\nCOMMIT;\n"
rows_bytes = sum(len(r.encode("utf-8")) for r in rows)
fixed = len((head + begin + commit).encode("utf-8"))
exact = len(rows) == total
if exact:
full_bytes = fixed + rows_bytes
else:
full_bytes = fixed + round(rows_bytes / len(rows) * total)
def estimate(self, df: pd.DataFrame) -> int:
"""Bytes one unsplit file of the whole table would take, from rows spread through it."""
total = len(df)
n = min(total, MEASURE_ROWS)
picks = sorted({round(i * (total - 1) / (n - 1)) for i in range(n)}) if n > 1 else [0]
measured = df.iloc[picks]
per_row = sum(utf8_len(self._row(row)) for _, row in measured.iterrows()) / len(measured)
fixed = utf8_len(self.open() + self.close())
if self.fmt == "batch":
batches = math.ceil(total / self.batch_rows)
framing = utf8_len(f"{self.insert}\n") + utf8_len("\nON CONFLICT DO NOTHING;\n\n")
return fixed + round((per_row + 2) * total) + batches * framing
return fixed + round(per_row * total)
def render_sample(shape: Shape, df: pd.DataFrame, max_rows: int):
"""
A capped table as one block of text, with a note of what was left out.
note = ""
if not exact:
Return (text, full_bytes). Small by definition, so it is built in memory:
the note at the top needs the estimate, which needs the rows.
"""
shown = df.head(max_rows)
body = "".join(shape.units(shown))
full_bytes = shape.estimate(df)
note = (
f"-- SAMPLE: first {len(rows)} of {total} rows. The full file would be "
f"-- SAMPLE: first {len(shown)} of {len(df)} rows. The full file would be "
f"~{human_bytes(full_bytes)}; run without --max-rows for all of it.\n"
)
return head + note + begin + "".join(rows) + commit, total, full_bytes, exact
return shape.open(note) + body + shape.close(), full_bytes
def human_bytes(n: int) -> str: