piker/piker/storage/_audit.py

653 lines
18 KiB
Python

# piker: trading gear for hackers
# Copyright (C) Tyler Goodlet (in stewardship for pikers)
'''
Read-only OHLCV persistence qualification.
'''
from datetime import (
UTC,
datetime,
)
from hashlib import sha256
import os
from pathlib import Path
import shutil
from stat import S_ISREG
from typing import BinaryIO
import numpy as np
import polars as pl
from piker.data import def_iohlcv_fields
AUDIT_SCHEMA: str = 'piker.nativedb.audit/v1'
_value_fields: tuple[str, ...] = (
'open',
'high',
'low',
'close',
'volume',
)
def _utc_str(timestamp: float|int|None) -> str|None:
'''
Render an epoch timestamp as deterministic UTC.
'''
if timestamp is None:
return None
try:
dt: datetime = datetime.fromtimestamp(timestamp, tz=UTC)
except (
OSError,
OverflowError,
ValueError,
):
return None
return dt.isoformat().replace('+00:00', 'Z')
def _number(value: float|int) -> float|int:
'''
Convert a NumPy number to a JSON-native scalar.
'''
if isinstance(value, (int, np.integer)):
return int(value)
value = float(value)
return int(value) if value.is_integer() else value
def _sha256(file: BinaryIO) -> str:
'''
Hash and rewind one already-open file snapshot.
'''
digest = sha256()
for chunk in iter(
lambda: file.read(1024 * 1024),
b'',
):
digest.update(chunk)
file.seek(0)
return digest.hexdigest()
def _numeric_counts(series: pl.Series) -> dict[str, int|None]:
'''
Count null and non-finite values without coercing text.
'''
null_count: int = series.null_count()
if not series.dtype.is_numeric():
return {
'null': null_count,
'nan': None,
'positive_infinity': None,
'negative_infinity': None,
'nonfinite': None,
}
numeric: pl.Series = series.cast(pl.Float64, strict=False)
nan_count: int = int(numeric.is_nan().sum() or 0)
pos_inf_count: int = int(
(numeric == float('inf')).sum() or 0
)
neg_inf_count: int = int(
(numeric == float('-inf')).sum() or 0
)
cast_null_count: int = max(
0,
numeric.null_count() - null_count,
)
return {
'null': null_count,
'nan': nan_count,
'positive_infinity': pos_inf_count,
'negative_infinity': neg_inf_count,
'nonfinite': (
null_count
+ nan_count
+ pos_inf_count
+ neg_inf_count
+ cast_null_count
),
}
def _numeric_values(
series: pl.Series,
) -> tuple[np.ndarray, np.ndarray]:
'''
Preserve integer precision and return an explicit valid mask.
'''
if series.dtype.is_integer():
valid: np.ndarray = series.is_not_null().to_numpy()
values: np.ndarray = series.fill_null(0).to_numpy()
return values, valid
values = series.to_numpy()
try:
valid = np.isfinite(values)
except TypeError:
values = series.cast(pl.Float64).to_numpy()
valid = np.isfinite(values)
return values, valid
def _empty_gaps(
period_s: int,
verifiable: bool = False,
) -> dict:
'''
Return the stable gap report shape for unauditable timestamps.
'''
return {
'basis': 'sorted_unique_valid_timestamps',
'expected_period_s': period_s,
'verifiable': verifiable,
'count': 0,
'aligned_count': 0,
'misaligned_count': 0,
'subperiod_count': 0,
'missing_samples_total': 0,
'details_count': 0,
'details_truncated': False,
'intervals': [],
'subperiod_intervals': [],
}
def _audit_timestamps(
df: pl.DataFrame,
period_s: int,
max_gaps: int,
) -> tuple[dict, dict]:
'''
Audit physical timestamp ordering and chronological gaps.
'''
if 'time' not in df.columns:
return (
{
'present': False,
'numeric': False,
'strictly_increasing': False,
},
_empty_gaps(period_s),
)
series: pl.Series = df['time']
counts: dict[str, int|None] = _numeric_counts(series)
if not series.dtype.is_numeric():
return (
{
'present': True,
'numeric': False,
**counts,
'strictly_increasing': False,
},
_empty_gaps(period_s),
)
values, finite = _numeric_values(series)
valid: np.ndarray = values[finite]
unique: np.ndarray = np.unique(valid)
adjacent_valid: np.ndarray = finite[:-1] & finite[1:]
deltas: np.ndarray = np.diff(values)
physical_deltas: np.ndarray = deltas[adjacent_valid]
zero_delta_count: int = int(np.count_nonzero(
physical_deltas == 0
))
negative_delta_count: int = int(np.count_nonzero(
physical_deltas < 0
))
first: float|int|None = None
last: float|int|None = None
if values.size:
if finite[0]:
first = _number(values[0])
if finite[-1]:
last = _number(values[-1])
minimum: float|int|None = None
maximum: float|int|None = None
if valid.size:
minimum = _number(np.min(valid))
maximum = _number(np.max(valid))
chronological: np.ndarray = unique[unique > 0]
gap_deltas: np.ndarray = np.diff(chronological)
gap_indexes: np.ndarray = np.flatnonzero(
gap_deltas > period_s
)
gap_values: np.ndarray = gap_deltas[gap_indexes]
integer_gaps: bool = np.issubdtype(
gap_values.dtype,
np.integer,
)
if integer_gaps:
aligned: np.ndarray = gap_values % period_s == 0
else:
aligned = np.isclose(
gap_values % period_s,
0,
)
aligned_count: int = int(np.count_nonzero(aligned))
if integer_gaps:
missing_total: int = sum(
int(delta) // period_s - 1
for delta in gap_values[aligned]
)
else:
missing_total = sum(
int(round(float(delta) / period_s)) - 1
for delta in gap_values[aligned]
)
subperiod_indexes: np.ndarray = np.flatnonzero(
gap_deltas < period_s
)
intervals: list[dict] = []
for gap_index in gap_indexes[:max_gaps]:
left: float = chronological[gap_index]
right: float = chronological[gap_index + 1]
delta: float = right - left
period_multiple: bool = bool(np.isclose(
delta % period_s,
0,
))
missing_samples: int|None = None
if period_multiple:
missing_samples = (
int(delta) // period_s - 1
if integer_gaps
else int(round(float(delta) / period_s)) - 1
)
left_value: float|int = _number(left)
right_value: float|int = _number(right)
intervals.append({
'left_timestamp': left_value,
'right_timestamp': right_value,
'left_utc': _utc_str(left_value),
'right_utc': _utc_str(right_value),
'delta_s': _number(delta),
'period_multiple': period_multiple,
'missing_samples': missing_samples,
'classification': 'unclassified',
})
subperiod_intervals: list[dict] = []
remaining_details: int = max(0, max_gaps - len(intervals))
for step_index in subperiod_indexes[:remaining_details]:
left = _number(chronological[step_index])
right = _number(chronological[step_index + 1])
subperiod_intervals.append({
'left_timestamp': left,
'right_timestamp': right,
'left_utc': _utc_str(left),
'right_utc': _utc_str(right),
'delta_s': _number(right - left),
})
gap_count: int = gap_indexes.size
gaps: dict = {
'basis': 'sorted_unique_valid_timestamps',
'expected_period_s': period_s,
'verifiable': chronological.size >= 2,
'count': gap_count,
'aligned_count': aligned_count,
'misaligned_count': gap_count - aligned_count,
'missing_samples_total': missing_total,
'subperiod_count': subperiod_indexes.size,
'details_count': (
len(intervals)
+ len(subperiod_intervals)
),
'details_truncated': bool(
gap_count > len(intervals)
or
subperiod_indexes.size > len(subperiod_intervals)
),
'intervals': intervals,
'subperiod_intervals': subperiod_intervals,
}
timestamps: dict = {
'present': True,
'numeric': True,
'first_in_file': first,
'last_in_file': last,
'minimum': minimum,
'maximum': maximum,
'minimum_utc': _utc_str(minimum),
'maximum_utc': _utc_str(maximum),
**counts,
'zero': int(np.count_nonzero(valid == 0)),
'negative': int(np.count_nonzero(valid < 0)),
'fractional': int(np.count_nonzero(
valid != np.floor(valid)
)),
'unique': unique.size,
'duplicate_excess': valid.size - unique.size,
'adjacent_zero_delta': zero_delta_count,
'negative_delta': negative_delta_count,
'non_positive_delta': (
zero_delta_count
+ negative_delta_count
),
'strictly_increasing': bool(
values.size > 0
and
values.size == valid.size
and
np.all(np.diff(values) > 0)
),
}
return timestamps, gaps
def _audit_index(df: pl.DataFrame) -> dict:
'''
Verify persisted indexes match physical row positions.
'''
if 'index' not in df.columns:
return {
'present': False,
'numeric': False,
'canonical_from_zero': False,
}
series: pl.Series = df['index']
counts: dict[str, int|None] = _numeric_counts(series)
if not series.dtype.is_numeric():
return {
'present': True,
'numeric': False,
'dtype': str(series.dtype),
**counts,
'canonical_from_zero': False,
}
values, finite = _numeric_values(series)
expected: np.ndarray = np.arange(df.height)
mismatch_count: int = int(np.count_nonzero(
~finite
|
(values != expected)
))
valid: np.ndarray = values[finite]
unique_count: int = np.unique(valid).size
return {
'present': True,
'numeric': True,
'dtype': str(series.dtype),
'first': (
_number(values[0])
if values.size and finite[0]
else None
),
'last': (
_number(values[-1])
if values.size and finite[-1]
else None
),
**counts,
'duplicate_excess': valid.size - unique_count,
'non_unit_step': int(np.count_nonzero(
np.diff(valid) != 1
)),
'row_position_mismatch': mismatch_count,
'contiguous': bool(
values.size > 0
and
valid.size == values.size
and
np.all(np.diff(values) == 1)
),
'canonical_from_zero': bool(
values.size > 0
and
values.size == expected.size
and
mismatch_count == 0
),
}
def audit_ohlcv_frame(
df: pl.DataFrame,
fqme: str,
period_s: int,
max_gaps: int = 100,
) -> dict:
'''
Inspect a raw persisted frame without normalizing evidence.
'''
if period_s < 1:
raise ValueError('Audit period must be positive')
if max_gaps < 0:
raise ValueError('Maximum gap details must be non-negative')
expected_columns: list[str] = [
name
for name, _ in def_iohlcv_fields
]
expected_dtypes: dict[str, str] = {
name: str(pl.Int64 if field_type is int else pl.Float64)
for name, field_type in def_iohlcv_fields
}
actual_dtypes: dict[str, str] = {
name: str(dtype)
for name, dtype in df.schema.items()
}
missing: list[str] = sorted(
set(expected_columns).difference(df.columns)
)
extra: list[str] = sorted(
set(df.columns).difference(expected_columns)
)
column_order_ok: bool = df.columns == expected_columns
dtypes_ok: bool = all(
actual_dtypes.get(name) == dtype
for name, dtype in expected_dtypes.items()
)
timestamps, gaps = _audit_timestamps(
df,
period_s,
max_gaps,
)
index: dict = _audit_index(df)
nulls: dict[str, int|None] = {}
nans: dict[str, int|None] = {}
infinities: dict[str, int|None] = {}
all_values_finite: bool = True
for field in _value_fields:
if field not in df.columns:
nulls[field] = None
nans[field] = None
infinities[field] = None
all_values_finite = False
continue
counts = _numeric_counts(df[field])
nulls[field] = counts['null']
nans[field] = counts['nan']
pos_inf: int|None = counts['positive_infinity']
neg_inf: int|None = counts['negative_infinity']
infinities[field] = (
None
if pos_inf is None or neg_inf is None
else pos_inf + neg_inf
)
if counts['nonfinite'] != 0:
all_values_finite = False
schema_ok: bool = bool(
not missing
and
not extra
and
column_order_ok
and
dtypes_ok
)
violations: list[str] = []
if df.is_empty():
violations.append('empty_frame')
if missing:
violations.append('missing_columns')
if extra:
violations.append('extra_columns')
if not column_order_ok:
violations.append('column_order')
if not dtypes_ok:
violations.append('canonical_dtypes')
if not all_values_finite:
violations.append('nonfinite_ohlcv')
if timestamps.get('nonfinite') != 0:
violations.append('nonfinite_timestamps')
if (
timestamps.get('zero', 0)
or
timestamps.get('negative', 0)
):
violations.append('nonpositive_timestamps')
if timestamps.get('fractional', 0):
violations.append('fractional_timestamps')
if timestamps.get('duplicate_excess', 0):
violations.append('duplicate_timestamps')
if not timestamps.get('strictly_increasing', False):
violations.append('timestamp_order')
if not index.get('canonical_from_zero', False):
violations.append('index_not_canonical')
if gaps.get('misaligned_count', 0):
violations.append('misaligned_time_gap')
if gaps.get('subperiod_count', 0):
violations.append('subperiod_time_step')
warnings: list[str] = []
if gaps['count']:
warnings.append('positive_time_gaps_unclassified')
structural_ok: bool = not violations
gap_free: bool|None = (
gaps['count'] == 0
if gaps['verifiable']
else None
)
return {
'audit_schema': AUDIT_SCHEMA,
'fqme': fqme,
'period_s': period_s,
'schema': {
'columns': df.columns,
'dtypes': actual_dtypes,
'expected_columns': expected_columns,
'expected_dtypes': expected_dtypes,
'missing_columns': missing,
'extra_columns': extra,
'column_order_ok': column_order_ok,
'canonical_dtypes_ok': dtypes_ok,
'canonical': schema_ok,
},
'rows': {
'count': df.height,
'empty': df.is_empty(),
},
'timestamps': timestamps,
'index': index,
'values': {
'null_by_column': nulls,
'nan_by_column': nans,
'infinity_by_column': infinities,
'all_finite': all_values_finite,
},
'gaps': gaps,
'result': {
'structural_ok': structural_ok,
'gap_free': gap_free,
'qualification_ok': bool(structural_ok and gap_free),
'violations': violations,
'warnings': warnings,
},
}
def audit_ohlcv_parquet(
path: Path,
fqme: str,
period_s: int,
max_gaps: int = 100,
snapshot: Path|None = None,
) -> dict:
'''
Read and audit one Parquet file without touching storage state.
'''
path = path.expanduser().absolute()
descriptor: int = os.open(
path,
os.O_RDONLY
| os.O_CLOEXEC
| os.O_NOFOLLOW
| os.O_NONBLOCK,
)
stat = os.fstat(descriptor)
if not S_ISREG(stat.st_mode):
os.close(descriptor)
raise ValueError('Audit source must be a regular file')
with os.fdopen(descriptor, 'rb') as file:
checksum: str = _sha256(file)
if snapshot is not None:
snapshot_is_symlink: bool = snapshot.is_symlink()
snapshot = snapshot.resolve()
if snapshot == path:
raise ValueError(
'Snapshot path can not replace source'
)
if snapshot_is_symlink:
raise ValueError(
'Snapshot path can not be a symlink'
)
file.seek(0)
created: bool = False
try:
with snapshot.open('xb') as snapshot_file:
created = True
os.fchmod(snapshot_file.fileno(), 0o444)
shutil.copyfileobj(file, snapshot_file)
except BaseException:
if created:
snapshot.unlink(missing_ok=True)
raise
file.seek(0)
frame: pl.DataFrame = pl.read_parquet(file)
report: dict = audit_ohlcv_frame(
frame,
fqme,
period_s,
max_gaps,
)
generated_at: float = datetime.now(tz=UTC).timestamp()
report['generated_at_utc'] = _utc_str(generated_at)
report['source'] = {
'path': str(path),
'size_bytes': stat.st_size,
'mtime_ns': stat.st_mtime_ns,
'sha256': checksum,
}
return report