File size: 12,336 Bytes
2af1d9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c1f89bf
 
 
 
 
 
2af1d9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
02537fe
c1f89bf
 
 
 
 
 
 
02537fe
2af1d9f
 
 
 
 
 
 
 
c1f89bf
2af1d9f
 
 
c1f89bf
2af1d9f
 
 
 
02537fe
a04d2bd
 
 
 
 
 
 
02537fe
2af1d9f
 
 
 
 
 
 
 
 
 
 
a04d2bd
 
 
 
 
 
c1f89bf
 
 
 
 
 
 
2af1d9f
 
 
 
 
 
 
 
 
 
 
02537fe
 
2af1d9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a04d2bd
02537fe
a04d2bd
 
 
c1f89bf
 
 
a04d2bd
02537fe
a04d2bd
 
 
 
 
 
c1f89bf
 
a04d2bd
 
 
 
 
 
 
 
 
 
2af1d9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
"""Supabase read-only helpers + static monitoring artefacts loaders.



DB queries hit ``predictions_log`` (production data) and never touch the

test table. Static artefacts (proba_reference, feature_importance,

drift_report JSON snapshot) live under ``dashboard/static/`` and are

loaded once per session.

"""

from __future__ import annotations

import json
import os
from datetime import datetime, timedelta, timezone
from pathlib import Path
from typing import Any

import pandas as pd
import streamlit as st
from sqlalchemy import Engine, create_engine, text

PROD_TABLE = "predictions_log"
STATIC_DIR = Path(__file__).parent / "static"

# Cap on rows returned by fetch_recent — the Business tab only renders the
# 50 most recent ones, and the proba histogram in the Operational tab is
# representative beyond a few thousand. Without a LIMIT, a 7-day window at
# production QPS would materialise tens of thousands of rows in memory.
RECENT_ROW_LIMIT = 5000


@st.cache_resource(show_spinner=False)
def get_engine() -> Engine:
    url = os.getenv("DATABASE_URL")
    if not url:
        # Local dev convenience: try the repo's database/.env file. Silent
        # no-op on the HF Space where python-dotenv may not be installed and
        # the URL is configured via Space secrets.
        try:
            from dotenv import load_dotenv

            env_path = Path(__file__).resolve().parents[1] / "database" / ".env"
            if env_path.exists():
                load_dotenv(env_path)
                url = os.getenv("DATABASE_URL")
        except ImportError:
            pass
    if not url:
        raise RuntimeError(
            "DATABASE_URL is not set. Configure it as a Space secret "
            "(read-only role recommended) or in database/.env for local dev."
        )
    return create_engine(url, pool_size=2, max_overflow=2, pool_pre_ping=True, future=True)


@st.cache_data(ttl=60, show_spinner=False)
def fetch_recent(hours: int) -> pd.DataFrame:
    """Wide DataFrame of recent rows (capped at ``RECENT_ROW_LIMIT``).



    JSONB ``features`` stays as a Python dict. The cap prevents large

    windows on a busy day from materialising tens of thousands of rows in

    Streamlit memory — Business / Operational tabs only consume the latest

    few hundred anyway.

    """
    since = datetime.now(tz=timezone.utc) - timedelta(hours=hours)
    sql = text(
        f"""

        SELECT timestamp, sk_id_curr, client_known, latency_ms, status_code,

               error_message, probability_default, decision, threshold,

               model_version, features

        FROM {PROD_TABLE}

        WHERE timestamp >= :since

        ORDER BY timestamp DESC

        LIMIT :lim

        """
    )
    with get_engine().connect() as conn:
        df = pd.read_sql(sql, conn, params={"since": since, "lim": RECENT_ROW_LIMIT})
    return df


@st.cache_data(ttl=60, show_spinner=False)
def fetch_summary(hours: int) -> dict:
    """Aggregate KPIs computed in SQL to keep the dashboard responsive.



    Returns p50/p95 of total ``latency_ms`` plus the étape-4 breakdown

    (``feature_assembly_ms``, ``inference_ms``, ``inference_cpu_ms``). The

    breakdown columns are NULL on legacy rows logged before the

    instrumentation was added — PERCENTILE_CONT ignores NULLs natively.

    """
    since = datetime.now(tz=timezone.utc) - timedelta(hours=hours)
    sql = text(
        f"""

        SELECT

            COUNT(*)                                                AS total,

            SUM(CASE WHEN status_code = 200 THEN 1 ELSE 0 END)      AS ok,

            SUM(CASE WHEN status_code != 200 THEN 1 ELSE 0 END)     AS errors,

            SUM(CASE WHEN decision = 'GRANTED' THEN 1 ELSE 0 END)   AS granted,

            SUM(CASE WHEN decision = 'REFUSED' THEN 1 ELSE 0 END)   AS refused,

            SUM(CASE WHEN client_known = false THEN 1 ELSE 0 END)   AS unknowns,

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY latency_ms) AS p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY latency_ms) AS p95,

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY feature_assembly_ms) AS asm_p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY feature_assembly_ms) AS asm_p95,

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY inference_ms)        AS inf_p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY inference_ms)        AS inf_p95,

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY inference_cpu_ms)    AS inf_cpu_p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY inference_cpu_ms)    AS inf_cpu_p95,

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY db_log_ms)           AS db_log_p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY db_log_ms)           AS db_log_p95,

            -- plumbing_ms is computed at log time in the API as

            -- (latency_ms - feature_assembly_ms - inference_ms). We read it

            -- directly here — no arithmetic in the dashboard SQL.

            PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY plumbing_ms)         AS plumbing_p50,

            PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY plumbing_ms)         AS plumbing_p95,

            AVG(probability_default)                                AS avg_proba

        FROM {PROD_TABLE}

        WHERE timestamp >= :since

        """
    )
    with get_engine().connect() as conn:
        row = conn.execute(sql, {"since": since}).mappings().one()
    return dict(row)


@st.cache_data(ttl=60, show_spinner=False)
def fetch_volume_by_hour(hours: int) -> pd.DataFrame:
    since = datetime.now(tz=timezone.utc) - timedelta(hours=hours)
    sql = text(
        f"""

        SELECT date_trunc('hour', timestamp) AS hour,

               COUNT(*)                       AS total,

               SUM(CASE WHEN status_code != 200 THEN 1 ELSE 0 END) AS errors,

               PERCENTILE_CONT(0.5)  WITHIN GROUP (ORDER BY latency_ms) AS p50,

               PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY latency_ms) AS p95

        FROM {PROD_TABLE}

        WHERE timestamp >= :since

        GROUP BY 1

        ORDER BY 1

        """
    )
    with get_engine().connect() as conn:
        return pd.read_sql(sql, conn, params={"since": since})


@st.cache_data(ttl=60, show_spinner=False)
def fetch_latency_breakdown(hours: int) -> pd.DataFrame:
    """Hourly p50 of each timing component (étape 4).



    Returns columns: hour, total_p50, feature_assembly_p50, inference_p50,

    inference_cpu_p50, db_log_p50. NULL columns from legacy rows are

    skipped by PERCENTILE_CONT, so hours predating the instrumentation

    surface as NaN.

    """
    since = datetime.now(tz=timezone.utc) - timedelta(hours=hours)
    sql = text(
        f"""

        SELECT date_trunc('hour', timestamp) AS hour,

               PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY latency_ms)          AS total_p50,

               PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY feature_assembly_ms) AS feature_assembly_p50,

               PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY inference_ms)        AS inference_p50,

               PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY inference_cpu_ms)    AS inference_cpu_p50,

               PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY db_log_ms)           AS db_log_p50

        FROM {PROD_TABLE}

        WHERE timestamp >= :since AND status_code = 200

        GROUP BY 1

        ORDER BY 1

        """
    )
    with get_engine().connect() as conn:
        return pd.read_sql(sql, conn, params={"since": since})


@st.cache_data(ttl=60, show_spinner=False)
def fetch_proba_distribution(limit: int) -> list[float]:
    """Return the last ``limit`` successful prediction probabilities (most recent first)."""
    sql = text(
        f"""

        SELECT probability_default FROM {PROD_TABLE}

        WHERE status_code = 200 AND probability_default IS NOT NULL

        ORDER BY timestamp DESC

        LIMIT :lim

        """
    )
    with get_engine().connect() as conn:
        rows = conn.execute(sql, {"lim": limit}).all()
    return [float(r.probability_default) for r in rows]


# ---------------------------------------------------------------------------
# Static artefacts (built offline by scripts/build_monitoring_artefacts.py
# and scripts/generate_drift_report.py). Cached for the lifetime of the
# Streamlit process — refresh by restarting the Space.
# ---------------------------------------------------------------------------


def _load_json(path: Path) -> dict | None:
    if not path.exists():
        return None
    return json.loads(path.read_text(encoding="utf-8"))


@st.cache_resource(show_spinner=False)
def load_proba_reference() -> dict | None:
    """Histogram + raw samples of training-time probability_default."""
    return _load_json(STATIC_DIR / "proba_reference.json")


@st.cache_resource(show_spinner=False)
def load_feature_importance() -> dict | None:
    """Top-K features by SHAP mean(|value|) on the reference dataset."""
    return _load_json(STATIC_DIR / "feature_importance.json")


@st.cache_resource(show_spinner=False)
def load_drift_report_json() -> dict | None:
    """JSON snapshot saved alongside the HTML by generate_drift_report.py."""
    return _load_json(STATIC_DIR / "drift_report.json")


def parse_drift_results(report_json: dict | None) -> dict[str, dict[str, Any]]:
    """Parse the Evidently 0.7+ JSON snapshot.



    Schema observed in 0.7.x::



        {

          "metrics": [

            {

              "metric_name": "ValueDrift(column=AMT_INCOME_TOTAL,...)",

              "config": {"column": "AMT_INCOME_TOTAL", "method": "K-S p_value",

                          "threshold": 0.05, "type": "evidently:metric_v2:ValueDrift"},

              "value": 7.06e-07

            },

            {

              "metric_name": "DriftedColumnsCount(...)",

              "value": {"count": 40.0, "share": 0.052}

            },

            ...

          ],

          "tests": [...]

        }



    For each ValueDrift entry: drift detected when ``value < threshold``

    (the value is a p-value, lower = more drift).



    Returns: ``{feature_name: {"detected": bool, "score": float, "stattest": str}}``.

    """
    results: dict[str, dict[str, Any]] = {}
    if not report_json:
        return results

    for metric in report_json.get("metrics", []) or []:
        config = metric.get("config", {}) or {}
        # Only ValueDrift entries are per-column; skip the aggregate
        # DriftedColumnsCount and any other metric type.
        if config.get("type", "").endswith(":ValueDrift") is False:
            continue
        column = config.get("column")
        if not isinstance(column, str):
            continue
        threshold = float(config.get("threshold", 0.05))
        method = config.get("method") or "—"
        raw_value = metric.get("value")
        score: float | None
        detected: bool | None
        if isinstance(raw_value, (int, float)):
            score = float(raw_value)
            detected = score < threshold
        else:
            score = None
            detected = None
        results[column] = {
            "detected": detected,
            "score": score,
            "stattest": str(method),
        }
    return results


@st.cache_data(show_spinner=False)
def load_drift_summary(_report_json: dict | None) -> dict | None:
    """Extract the Evidently dataset-level drift verdict from the JSON.



    Reads the ``DriftedColumnsCount`` metric in the snapshot. Returns

    ``None`` if absent (older Evidently versions or empty report).

    """
    if not _report_json:
        return None
    for metric in _report_json.get("metrics", []) or []:
        config = metric.get("config", {}) or {}
        if config.get("type", "").endswith(":DriftedColumnsCount"):
            value = metric.get("value") or {}
            return {
                "count": int(value.get("count", 0)),
                "share": float(value.get("share", 0.0)),
                "threshold": float(config.get("drift_share", 0.5)),
            }
    return None