Files
routesapi/app/config/dynamic_config.py
2026-06-22 17:40:08 +05:30

409 lines
15 KiB
Python

"""
Dynamic Configuration - rider-api
Replaces all hardcoded hyperparameters with DB-backed values.
The ML hypertuner writes optimal values here; services read from here.
Fallback: If DB is unavailable or no tuned values exist, defaults are used.
This means zero risk - the system works day 1 with no data.
"""
import contextvars
import json
import logging
import os
import sqlite3
from datetime import datetime
from typing import Any, Dict, Optional
logger = logging.getLogger(__name__)
# --- DB Path ------------------------------------------------------------------
_DB_PATH = os.getenv("ML_DB_PATH", "ml_data/ml_store.db")
# ---------------------------------------------------------------------------
# Per-request strategy override (async-safe via contextvars).
# Each FastAPI request/asyncio task gets its own copy — no cross-request leaks.
# Usage:
# set_request_strategy("fuel_saver") → override active for this request
# clear_request_strategy() → restore to DB value
# ---------------------------------------------------------------------------
_strategy_override: contextvars.ContextVar[Optional[str]] = contextvars.ContextVar(
"strategy_override", default=None
)
def set_request_strategy(strategy: Optional[str]) -> None:
"""Override ml_strategy for the current async task only (thread-safe)."""
_strategy_override.set(strategy)
def clear_request_strategy() -> None:
"""Remove the per-request strategy override for the current async task."""
_strategy_override.set(None)
# --- Hard Defaults (What the system used before ML) ---------------------------
DEFAULTS: Dict[str, Any] = {
# System Strategy / Prompt
"ml_strategy": "balanced",
# AssignmentService
"max_pickup_distance_km": 10.0,
"max_kitchen_distance_km": 3.0,
"max_orders_per_rider": 12,
"ideal_load": 6,
"workload_balance_threshold": 0.7,
"workload_penalty_weight": 100.0,
"distance_penalty_weight": 2.0,
"preference_bonus": -15.0,
"home_zone_bonus_4km": -3.0,
"home_zone_bonus_2km": -5.0,
"emergency_load_penalty": 3.0, # km penalty per order in emergency assign
# RouteOptimizer
"search_time_limit_seconds": 5,
"avg_speed_kmh": 18.0,
"road_factor": 1.3,
# ClusteringService
"cluster_radius_km": 3.0,
# KalmanFilter
"kalman_process_noise": 1e-4,
"kalman_measurement_noise": 0.01,
# RealisticETACalculator
"eta_pickup_time_min": 3.0,
"eta_delivery_time_min": 4.0,
"eta_navigation_buffer_min": 1.5,
"eta_short_trip_factor": 0.8, # speed multiplier for dist < 2km
"eta_long_trip_factor": 1.1, # speed multiplier for dist > 8km
# EmpiricalETACalculator (learned ETAs from actual delivery times)
"eta_empirical_enabled": True, # False -> instantly revert to the formula
"eta_min_samples": 10, # min history samples a key needs before trust
# (backtest on live 14d data: 10 -> MAE 4.85 vs
# formula 5.73; 20 -> 5.64. 10 wins on held-out.)
"eta_history_days": 14, # rolling window pulled from nearledb
"eta_stat": "median", # "median" or "p75" (p75 = more conservative)
"eta_sync_interval_hours": 6, # autonomous background sync cadence
# Road-aware sequencing (Phase 2). OFF by default: enabling adds a Google
# Directions call (cost + latency) to the route hot path. Results are cached.
# Only the *visiting order* changes; step/ETA metrics stay aerial-based.
"routing_use_road_distance": False, # AGENT-MANAGED (see routing_auto_manage)
"routing_road_cache_ttl_seconds": 86400, # road geometry is stable; cache 24h
"routing_road_max_stops": 25, # Google distance-matrix practical cap
# Autonomous road-sequencing decision agent: measures road-vs-aerial travel
# time on real batches and flips routing_use_road_distance on its own.
"routing_auto_manage": True, # False -> humans own the flag
"routing_auto_enable_gain_pct": 3.0, # enable when mean gain >= this
"routing_auto_disable_gain_pct": 1.0, # disable when mean gain < this (hysteresis)
"routing_eval_sample_batches": 8, # batches measured per cycle (cost bound)
"routing_eval_min_batches": 3, # need >= this evaluated to decide
"routing_eval_interval_hours": 24, # decision cadence
"routing_eval_days": 14, # window sampled from the local mirror
# Learned rider->kitchen affinity (soft steering only; union with curated config).
"rider_affinity_enabled": True, # False -> pure curated config
"rider_affinity_min_deliveries": 10, # learned owner needs >= this many deliveries
"rider_affinity_refresh_hours": 6, # recompute cadence (piggybacks the agent)
}
class DynamicConfig:
"""
Thread-safe, DB-backed configuration store.
Usage:
cfg = DynamicConfig()
max_dist = cfg.get("max_pickup_distance_km")
all_params = cfg.get_all()
"""
_instance: Optional["DynamicConfig"] = None
def __new__(cls) -> "DynamicConfig":
"""Singleton - one config per process."""
if cls._instance is None:
cls._instance = super().__new__(cls)
cls._instance._initialized = False
return cls._instance
def __init__(self):
if self._initialized:
return
self._initialized = True
self._cache: Dict[str, Any] = {}
self._last_loaded: Optional[datetime] = None
self._ensure_db()
self._load()
# --------------------------------------------------------------------------
# Public API
# --------------------------------------------------------------------------
def get(self, key: str, default: Any = None) -> Any:
"""Get a config value. Returns ML-tuned value if available, else default.
For 'ml_strategy' specifically, a per-request ContextVar override takes
precedence so that hypertuning_params requests don't mutate the shared
singleton (thread-safe for concurrent FastAPI requests).
"""
self._maybe_reload()
# Per-request strategy override (async-safe, no cross-request leaks)
if key == "ml_strategy":
override = _strategy_override.get()
if override is not None:
return override
val = self._cache.get(key)
if val is not None:
return val
fallback = default if default is not None else DEFAULTS.get(key)
return fallback
def get_all(self) -> Dict[str, Any]:
"""Return all current config values (ML-tuned + defaults for missing keys)."""
self._maybe_reload()
result = dict(DEFAULTS)
result.update(self._cache)
return result
def set(self, key: str, value: Any, source: str = "manual") -> None:
"""Write a config value to DB (used by hypertuner)."""
try:
os.makedirs(os.path.dirname(_DB_PATH) or ".", exist_ok=True)
conn = sqlite3.connect(_DB_PATH)
conn.execute(
"""
INSERT INTO dynamic_config (key, value, source, updated_at)
VALUES (?, ?, ?, ?)
ON CONFLICT(key) DO UPDATE SET
value=excluded.value,
source=excluded.source,
updated_at=excluded.updated_at
""",
(key, json.dumps(value), source, datetime.utcnow().isoformat()),
)
conn.commit()
conn.close()
self._cache[key] = value
logger.info(f"[DynamicConfig] Set {key}={value} (source={source})")
except Exception as e:
logger.error(f"[DynamicConfig] Failed to set {key}: {e}")
def set_bulk(self, params: Dict[str, Any], source: str = "ml_hypertuner") -> None:
"""Write multiple config values at once (called after each Optuna study)."""
for key, value in params.items():
self.set(key, value, source=source)
logger.info(f"[DynamicConfig] Bulk update: {len(params)} params from {source}")
def reset_to_defaults(self) -> None:
"""Wipe all ML-tuned values, revert to hardcoded defaults."""
try:
conn = sqlite3.connect(_DB_PATH)
conn.execute("DELETE FROM dynamic_config")
conn.commit()
conn.close()
self._cache.clear()
logger.warning("[DynamicConfig] Reset to factory defaults.")
except Exception as e:
logger.error(f"[DynamicConfig] Reset failed: {e}")
# --------------------------------------------------------------------------
# Internal
# --------------------------------------------------------------------------
def _ensure_db(self) -> None:
try:
os.makedirs(os.path.dirname(_DB_PATH) or ".", exist_ok=True)
conn = sqlite3.connect(_DB_PATH)
conn.execute("""
CREATE TABLE IF NOT EXISTS dynamic_config (
key TEXT PRIMARY KEY,
value TEXT NOT NULL,
source TEXT DEFAULT 'manual',
updated_at TEXT
)
""")
conn.execute("""
CREATE TABLE IF NOT EXISTS kitchen_encoding (
kitchen_name TEXT PRIMARY KEY,
label_id INTEGER NOT NULL,
frequency REAL DEFAULT 0.0,
avg_profit REAL DEFAULT 0.0,
order_count INTEGER DEFAULT 0,
updated_at TEXT
)
""")
conn.commit()
conn.close()
except Exception as e:
logger.error(f"[DynamicConfig] DB init failed: {e}")
def _load(self) -> None:
try:
conn = sqlite3.connect(_DB_PATH)
rows = conn.execute("SELECT key, value FROM dynamic_config").fetchall()
conn.close()
self._cache = {}
for key, raw in rows:
try:
self._cache[key] = json.loads(raw)
except Exception:
self._cache[key] = raw
self._last_loaded = datetime.utcnow()
if self._cache:
logger.info(
f"[DynamicConfig] Loaded {len(self._cache)} ML-tuned params from DB"
)
except Exception as e:
logger.warning(
f"[DynamicConfig] Could not load from DB (using defaults): {e}"
)
self._cache = {}
def _maybe_reload(self, interval_seconds: int = 300) -> None:
"""Reload from DB every 5 minutes - picks up new tuned params without restart."""
if self._last_loaded is None:
self._load()
return
delta = (datetime.utcnow() - self._last_loaded).total_seconds()
if delta > interval_seconds:
self._load()
# --- Module-level convenience singleton ---------------------------------------
_cfg = DynamicConfig()
def get_config() -> DynamicConfig:
"""Get the global DynamicConfig singleton."""
return _cfg
__all__ = [
"DynamicConfig",
"get_config",
"set_request_strategy",
"clear_request_strategy",
"get_kitchen_label_id",
"get_kitchen_frequency",
"update_kitchen_stats",
"get_kitchen_avg_profit_smoothed",
]
# --- Kitchen Encoding Persistence ---------------------------------------------
def get_kitchen_label_id(kitchen_name: str) -> int:
"""Get or create persistent label ID for a kitchen."""
try:
conn = sqlite3.connect(_DB_PATH)
row = conn.execute(
"SELECT label_id FROM kitchen_encoding WHERE kitchen_name = ?",
(kitchen_name,),
).fetchone()
if row:
conn.close()
return row[0]
max_id = conn.execute(
"SELECT COALESCE(MAX(label_id), -1) FROM kitchen_encoding"
).fetchone()[0]
new_id = max_id + 1
conn.close()
return new_id
except Exception as e:
logger.warning(f"[KitchenEncoding] Failed to get label_id: {e}")
return hash(kitchen_name.lower().strip()) % 10000
def get_kitchen_frequency(kitchen_name: str) -> float:
"""Get frequency ratio for a kitchen from DB."""
try:
conn = sqlite3.connect(_DB_PATH)
row = conn.execute(
"SELECT frequency FROM kitchen_encoding WHERE kitchen_name = ?",
(kitchen_name,),
).fetchone()
conn.close()
return row[0] if row else 0.0
except Exception:
return 0.0
def update_kitchen_stats(kitchen_name: str, profit: float):
"""Update kitchen stats after order completion."""
try:
conn = sqlite3.connect(_DB_PATH)
row = conn.execute(
"SELECT order_count, avg_profit FROM kitchen_encoding WHERE kitchen_name = ?",
(kitchen_name,),
).fetchone()
if row:
count, avg = row
new_count = count + 1
new_avg = ((avg * count) + profit) / new_count
new_freq = new_count / (
conn.execute(
"SELECT SUM(order_count) FROM kitchen_encoding"
).fetchone()[0]
or 1
)
conn.execute(
"UPDATE kitchen_encoding SET order_count = ?, avg_profit = ?, frequency = ?, updated_at = ? WHERE kitchen_name = ?",
(
new_count,
new_avg,
new_freq,
datetime.utcnow().isoformat(),
kitchen_name,
),
)
else:
conn.execute(
"INSERT INTO kitchen_encoding (kitchen_name, label_id, frequency, avg_profit, order_count, updated_at) VALUES (?, ?, ?, ?, ?, ?)",
(
kitchen_name,
get_kitchen_label_id(kitchen_name),
1.0,
profit,
1,
datetime.utcnow().isoformat(),
),
)
conn.commit()
conn.close()
except Exception as e:
logger.warning(f"[KitchenEncoding] Failed to update stats: {e}")
def get_kitchen_avg_profit_smoothed(
kitchen_name: str, global_avg: float = 40.0, min_samples: int = 5
) -> float:
"""
Get smoothed average profit for a kitchen using Bayesian smoothing.
Reduces noise for kitchens with few orders.
Formula: smoothed = (kitchen_count * kitchen_avg + min_samples * global_avg) / (kitchen_count + min_samples)
This means:
- Kitchen with many samples -> uses its own avg
- Kitchen with few samples -> pulls toward global avg
"""
try:
conn = sqlite3.connect(_DB_PATH)
row = conn.execute(
"SELECT order_count, avg_profit FROM kitchen_encoding WHERE kitchen_name = ?",
(kitchen_name,),
).fetchone()
conn.close()
if row and row[0] > 0:
count, avg = row
if count >= min_samples:
return avg
# Bayesian smoothing
smoothed = ((count * avg) + (min_samples * global_avg)) / (
count + min_samples
)
return smoothed
return global_avg # Unknown kitchen defaults to global avg
except Exception:
return global_avg