Files
sientia-dataops-laborious_t…/laborious/activities/model_metrics.py

433 lines
18 KiB
Python

from temporalio import activity, workflow
with workflow.unsafe.imports_passed_through():
import time
import traceback
import warnings
from typing import Any
import numpy as np
import pandas as pd
from pandas import DataFrame, Index, Series, to_datetime
from sientia_do.notifications.handlers import CoreNotificationHandler as NotificationHandler
from sientia_do.notifications.models import NotificationLevel
from sientia_do.observability.logger import Logger
from sientia_do.observability.metrics_controller import MetricsController
from sientia_do.observability.sientia_monitoring import SientiaMonitoring
from sientia_do.temporal.constants import DATETIME_FORMAT_WITH_TZ
from sientia_model.analytics.drift_analysis import DriftAnalysis, DriftInsufficientDataError
from laborious import metrics
from laborious.utils.dataframe_debug import build_dataframe_debug_message
warnings.filterwarnings('ignore', category=RuntimeWarning, message='Degrees of freedom <= 0')
warnings.filterwarnings(
'ignore', category=RuntimeWarning, message='invalid value encountered in scalar divide'
)
class ModelMetrics(SientiaMonitoring):
"""
Metrics and statistical analysis activities for the Laborious pipeline.
This class centralizes drift/statistical computations and model-quality
aggregates used by scheduled workflows. Besides producing tabular outputs
for persistence, it also emits operational metrics (count, lag, error)
through ``SientiaMonitoring`` so execution health is observable in runtime.
"""
_MAX_DEBUG_DATAFRAME_ROWS = 100
def __init__(
self,
logger: Logger,
notification_handler: NotificationHandler,
metrics_controller: MetricsController,
):
SientiaMonitoring.__init__(self, logger, notification_handler, metrics_controller)
def close(self) -> None:
"""
Shutdown monitoring resources associated with model metrics activities.
This is invoked during worker teardown to flush/close metric controller
internals and prevent dangling telemetry tasks.
"""
SientiaMonitoring.shutdown(self)
def __del__(self):
self.close()
def _debug_dataframe(self, message: str, data: Any, metadata: dict[str, Any]) -> None:
"""
Log dataframe content only when row count is below the configured threshold
Args:
- message (str): Base log message to identify the dataframe in logs
- data (Any): Dataframe-like payload to be logged
- metadata (dict[str, Any]): Workflow metadata for contextual logging
"""
self.debug(
build_dataframe_debug_message(
message=message,
data=data,
max_rows=self._MAX_DEBUG_DATAFRAME_ROWS,
),
metadata,
)
def _drift_analyze_stage_error(
self,
exc: Exception,
context: str,
metadata: dict[str, Any],
core_labels: dict[str, Any],
) -> None:
"""
Log analyzer failure for a drift stage and increment the analyze error metric.
Args:
- exc (Exception): Failure raised by ``sientia_model``.
- context (str): Short label for the log line (e.g. univariate detection).
- metadata (dict[str, Any]): Workflow metadata for logging.
- core_labels (dict[str, Any]): Tags from ``get_core_labels`` for metrics.
"""
self.error(f'{context}: {exc}', metadata)
self.emit_metric_sync(metric_object=metrics.MODEL_ANALYZE_ERROR_COUNT, tags=core_labels)
def get_drift_metrics(
self,
reference_data: DataFrame,
target_data: DataFrame,
target_name: str,
reference_columns: Index,
drift_metrics: list[str],
chunk_period: str,
metadata: dict[str, Any],
) -> DataFrame:
"""
Compute univariate and multivariate drift outputs and merge them into one dataframe.
The method orchestrates three analysis stages (univariate drift,
multivariate drift, and dataframe projection), emitting lag/count/error
metrics for each stage independently so failures are attributable.
Args:
- reference_data (DataFrame): Baseline dataset representing expected behavior.
- target_data (DataFrame): Current analysis dataset to compare against reference.
- target_name (str): Target column name used by ``DriftAnalysis`` config.
- reference_columns (Index): Feature columns evaluated for drift.
- drift_metrics (list[str]): Enabled univariate methods.
- chunk_period (str): Time bucket granularity used by analysis methods.
- metadata (dict[str, Any]): Workflow metadata for logs and notifications.
Return:
DataFrame: Consolidated drift dataframe from ``get_drift_metrics_dataframe`` using
``method`` / ``value`` (and optional ``threshold``, ``drift_type``), ready for
activity-level formatting before Postgres export.
"""
# ``DriftAnalysis`` uses truthiness checks on ``features`` (e.g. ``if not features``);
# a pandas ``Index`` is ambiguous in boolean context — normalize to a list.
feature_names: list[str] = list(reference_columns)
config = {
'target': target_name,
'prediction': 'prediction',
'timestamp': 'timestamp',
'features': feature_names,
}
drift_analysis = DriftAnalysis(config=config)
self._debug_dataframe(
f'Reference data: Size {reference_data.shape}', reference_data, metadata
)
self._debug_dataframe(f'Target data: Size {target_data.shape}', target_data, metadata)
core_labels = self.get_core_labels(metadata, operation_type='detect_univariate_drift')
start_time = time.time()
try:
univariate_drift = drift_analysis.detect_univariate_drift(
reference_df=reference_data,
analysis_df=target_data,
features=feature_names,
timestamp_col=config['timestamp'],
methods=drift_metrics,
chunk_period=chunk_period,
)
except Exception as e:
if isinstance(e, DriftInsufficientDataError):
raise
self._drift_analyze_stage_error(
e, 'Error detecting univariate drift', metadata, core_labels
)
raise
self.observe_lag_sync(start_time, metrics.MODEL_ANALYZE_LAG, core_labels)
self.emit_metric_sync(metric_object=metrics.MODEL_ANALYZE_COUNT, tags=core_labels)
core_labels = self.get_core_labels(metadata, operation_type='detect_multivariate_drift')
start_time = time.time()
try:
multivariate_drift = drift_analysis.detect_multivariate_drift(
reference_df=reference_data,
analysis_df=target_data,
features=feature_names,
timestamp_col=config['timestamp'],
chunk_period=chunk_period,
)
except Exception as e:
if isinstance(e, DriftInsufficientDataError):
raise
self._drift_analyze_stage_error(
e, 'Error detecting multivariate drift', metadata, core_labels
)
raise
self.observe_lag_sync(start_time, metrics.MODEL_ANALYZE_LAG, core_labels)
self.emit_metric_sync(metric_object=metrics.MODEL_ANALYZE_COUNT, tags=core_labels)
start_time = time.time()
core_labels = self.get_core_labels(metadata, operation_type='get_drift_metrics_dataframe')
try:
drift_df = drift_analysis.get_drift_metrics_dataframe(
univariate_drift=univariate_drift,
multivariate_drift=multivariate_drift,
)
except Exception as e:
if isinstance(e, DriftInsufficientDataError):
raise
self._drift_analyze_stage_error(
e, 'Error building drift metrics dataframe', metadata, core_labels
)
raise
self.observe_lag_sync(start_time, metrics.MODEL_ANALYZE_LAG, core_labels)
self.emit_metric_sync(metric_object=metrics.MODEL_ANALYZE_COUNT, tags=core_labels)
return drift_df
@staticmethod
def _to_naive_utc(series: Series) -> Series:
"""
Parse ``series`` as datetime and return a TZ-naive UTC copy.
``sientia_model.analytics.drift_analysis.DriftAnalysis`` preserves the
timezone of the input dataframe in its outputs, while target rows
loaded from PostgreSQL come in with ``+00:00``. Forcing both sides of
a comparison to TZ-naive UTC keeps ``isin`` / ``floor`` operations
deterministic regardless of how the analyzer constructs its
timestamps.
Args:
- series (Series): Input series containing datetime-parseable values.
Return:
Series: Datetime64 series with ``tz=None`` representing UTC instants.
"""
parsed = to_datetime(series)
if getattr(parsed.dt, 'tz', None) is not None:
parsed = parsed.dt.tz_convert('UTC').dt.tz_localize(None)
return parsed
@activity.defn(name='calculate_drift')
def calculate_drift(self, input_data: dict[str, Any]) -> list[dict]:
"""
Calculate drift metrics for a model.
Args:
input_data (dict[str, Any]): Input data containing:
- metadata (dict): Workflow execution metadata
- model_name (str): Name of the MLFlow model to calculate drift for
- reference_data (pd.DataFrame): Reference data for the model
- target_data (pd.DataFrame): Target data for calculating drift
- target_name (str): Name of the target column
- drift_metrics (list[str]): List of drift metrics to calculate
"""
metadata = input_data['metadata']
model_name = input_data['model_name']
model_id = input_data['model_id']
reference_raw_data = input_data['reference_data']
target_data = DataFrame(input_data['target_data'])
target_name = input_data['target_name']
drift_metrics = input_data['drift_metrics']
chunk_period = input_data['chunk_period']
if chunk_period not in ['min', 's']:
self.error(f'Invalid chunk period: {chunk_period}', metadata)
raise ValueError(f'Invalid chunk period: {chunk_period}, must be "min" or "s"')
self.info(f'Calculating drift for model {model_name}', metadata)
target_data = target_data.pivot(index='timestamp', columns='variable', values='value')
target_data['timestamp'] = target_data.index
# Keep timestamps as datetime: DriftAnalysis._chunk_dataframe relies on
# ``pd.Grouper(freq=...)`` which rejects string timestamp columns.
target_data['timestamp'] = to_datetime(target_data['timestamp'])
target_data = target_data.reset_index(drop=True)
target_data.dropna(inplace=True)
if reference_raw_data is not None:
self.info('Using reference data', metadata)
reference_data = DataFrame(reference_raw_data)
if 'timestamp' in reference_data.columns:
reference_data['timestamp'] = to_datetime(reference_data['timestamp'])
accurate = True
else:
# Get 30% first rows of target_data
self.warning('Using 30% first rows of target data as reference data', metadata)
target_data.sort_values(by='timestamp', ascending=True, inplace=True)
reference_data = target_data.head(int(len(target_data) * 0.3))
accurate = False
self.send_notification(
metadata=metadata,
notification_id='MODEL_METRICS_REFERENCE_DATA_WARNING',
message='Using 30% first rows of target data as reference data',
block='model_metrics',
level=NotificationLevel.WARNING,
attachment_content=reference_data.to_csv(),
)
reference_columns = reference_data.drop(
columns=[target_name, 'timestamp', 'target', 'prediction'], errors='ignore'
).columns
try:
drift_df = self.get_drift_metrics(
reference_data=reference_data,
target_data=target_data,
target_name=target_name,
reference_columns=reference_columns,
drift_metrics=drift_metrics,
chunk_period=chunk_period,
metadata=metadata,
)
except Exception as e:
if isinstance(e, DriftInsufficientDataError):
self.error(str(e), metadata)
notification_id = e.notification_id
notification_message = str(e)
else:
self.error(f'Error getting drift metrics: {e}', metadata)
notification_id = 'MODEL_METRICS_GET_DRIFT_METRICS_ERROR'
notification_message = f'Error getting drift metrics: {e}'
self.send_notification(
metadata=metadata,
notification_id=notification_id,
message=notification_message,
block='model_metrics',
level=NotificationLevel.ERROR,
attachment_content=traceback.format_exc(),
)
raise
# Drop chunks whose floored timestamp does not appear in the analysis window.
# ``DriftAnalysis`` chunks over ``analysis_df``; this only excludes rows that
# do not belong to the current target window (e.g. stray merged reference rows).
target_floor = self._to_naive_utc(target_data['timestamp']).dt.floor(chunk_period)
drift_floor = self._to_naive_utc(drift_df['timestamp']).dt.floor(chunk_period)
drift_df = drift_df[drift_floor.isin(target_floor)]
if drift_df.empty:
self.warning(
'No drift metrics found after dropping rows where timestamp is not in target data',
metadata,
)
return []
# Analyzer emits diagnostic columns that are not stored in ``sientia_data.drift_metrics``.
drift_df = drift_df.drop(columns=['threshold', 'drift_type'], errors='ignore')
drift_df['model_id'] = str(model_id)
drift_df['accurate'] = accurate
# ``timestamp`` is overridden with the most recent target instant so
# every persisted row shares a single business timestamp (the run's
# logical "now"), matching what downstream consumers expect.
latest_target_timestamp = self._to_naive_utc(target_data['timestamp']).max()
drift_df['timestamp'] = (
pd.Timestamp(latest_target_timestamp)
.tz_localize('UTC')
.strftime(DATETIME_FORMAT_WITH_TZ)
)
# ``chunk_start_date`` / ``chunk_end_date`` may carry nanosecond
# precision (beyond ``timestamptz`` microseconds), so serialize as ISO
# text for the ``text`` Postgres columns.
for column in ('chunk_start_date', 'chunk_end_date'):
drift_df[column] = drift_df[column].apply(
lambda value: pd.Timestamp(value).isoformat() if pd.notna(value) else None
)
self._debug_dataframe(f'Drift dataframe: Size {drift_df.shape}', drift_df, metadata)
return drift_df.to_dict(orient='records')
@activity.defn(name='calculate_simple_metrics')
def calculate_simple_metrics(self, input_data: dict[str, Any]) -> list[dict]:
"""
Calculate simple metrics for a model. Metrics available are:
- rmse
- mse
- mae
- r2
- accuracy
- precision
- recall
- f1
Args:
input_data (dict[str, Any]): Input data containing:
- metadata (dict): Workflow execution metadata
- model_id (str): ID of the MLFlow model
- target_data (pd.DataFrame): Target data for calculating metrics, containing target and prediction columns
- metrics (list[str]): List of metrics to calculate
Returns:
dict[Hashable, Any]: Dictionary containing the calculated metrics
"""
metadata = input_data['metadata']
model_id = input_data['model_id']
target_data = DataFrame(input_data['target_data'])
metric_names = input_data['metrics']
interval_minutes = input_data['interval_minutes']
data_size = target_data.shape[0]
output_data = []
diff = target_data['target'] - target_data['prediction']
diff_squared = diff**2
self.info(f'Calculating simple metrics for model {model_id}: {metric_names}', metadata)
for metric in metric_names:
if metric == 'rmse':
output_data.append({'metric': 'rmse', 'value': np.sqrt(np.mean(diff_squared))})
elif metric == 'mse':
output_data.append({'metric': 'mse', 'value': np.mean(diff_squared)})
elif metric == 'mae':
output_data.append({'metric': 'mae', 'value': np.mean(np.abs(diff))})
elif metric == 'r2':
y_true = target_data['target']
y_mean = np.mean(y_true)
ss_res = np.sum(diff_squared)
ss_tot = np.sum((y_true - y_mean) ** 2)
# Evita divisão por zero
if ss_tot == 0:
r2_score = 0.0
else:
r2_score = 1 - (ss_res / ss_tot)
output_data.append({'metric': 'r2', 'value': r2_score})
data = DataFrame(output_data)
data['model_id'] = model_id
data['timestamp'] = target_data['timestamp'].max()
data['data_size'] = data_size
data['interval_minutes'] = interval_minutes
self._debug_dataframe(f'Simple metrics dataframe: Size {data.shape}', data, metadata)
return data.to_dict(orient='records')