path: root/bin/lookup-server

#!/usr/bin/env python3
# vim:tabstop=4 softtabstop=4 shiftwidth=4 textwidth=160 smarttab expandtab colorcolumn=160
#
# Copyright (C) 2021 Daniel Friesel
#
# SPDX-License-Identifier: AGPL-3.0-only

import argparse
import psycopg2
import aiohttp
from aiohttp import web
from datetime import datetime, timedelta
import dateutil.parser
from geopy.distance import distance
import json
import logging
import numpy as np
import os
import pytz
import sys

max_distance = 50


class Stop:
    def __init__(self, eva, name, coord, ts):
        self.eva = eva
        self.name = name
        self.coord = coord
        self.when = ts

    def to_json(self):
        return self.eva, self.name, self.when.strftime("%H:%M")


class Train:
    def __init__(self, train):
        self.stopovers = train["previousStopovers"]
        self.tripId = train["tripId"]
        try:
            self.train_type, self.line_no = train["line"]["name"].split()
        except ValueError:
            self.train_type = train["line"]["name"]
            self.line_no = ""
        self.train_no = train["line"]["fahrtNr"]
        self.request_eva = int(train["stop"]["id"])

        if train["when"]:
            self.arrival = dateutil.parser.parse(train["when"])
        else:
            self.arrival = None

        if train["delay"] is not None:
            self.delay = timedelta(seconds=train["delay"])
        else:
            self.delay = timedelta()

        # preferred candidate for position estimation?
        self.preferred = False

        # previous/next stop and progress between those
        self.prev_stop = None
        self.next_stop = None
        self.progress_ratio = None

        # location and distance to requested position
        self.coarse_location = None
        self.location = None
        self.distance = None

        self.prepare_stopovers()

    def prepare_stopovers(self):
        """
        Parse arrival/departure into datetimes and add delay.

        /arrivals results have delay information ("when" ≠ "plannedWhen"). "previousStopovers" do not ("departure" == "plannedDeparture" in all cases).
        """
        for stopover in self.stopovers:
            if (
                stopover["arrival"]
                and stopover["plannedArrival"]
                and stopover["arrival"] != stopover["plannedArrival"]
            ):
                # arrival should be realtime. Didn't observe this case in practice yet.
                stopover["arrival"] = dateutil.parser.parse(stopover["arrival"])
            elif stopover["plannedArrival"]:
                # no realtime data available
                stopover["arrival"] = (
                    dateutil.parser.parse(stopover["plannedArrival"]) + self.delay
                )

            if (
                stopover["departure"]
                and stopover["plannedDeparture"]
                and stopover["departure"] != stopover["plannedDeparture"]
            ):
                stopover["departure"] = dateutil.parser.parse(stopover["departure"])
            elif stopover["plannedDeparture"]:
                stopover["departure"] = (
                    dateutil.parser.parse(stopover["plannedDeparture"]) + self.delay
                )

    def set_coarse_location(self, lat, lon):
        now = datetime.now(pytz.utc)
        train_evas = None
        legs = list()
        leg_distances = list()

        # includes train["stop"] -- but with arrival instead of departure
        # FIXME stopovers do not have realtime data :(
        for i, stopover in enumerate(self.stopovers):
            ts = None
            if stopover["departure"]:
                ts = stopover["departure"]
            elif stopover["arrival"]:
                ts = stopover["arrival"]

            if i == 0:
                if ts > now and (ts - now).seconds > 300:
                    # train has not departed from its initial stop yet and is not about to depart
                    return
                continue

            prev_eva = int(self.stopovers[i - 1]["stop"]["id"])
            this_eva = int(stopover["stop"]["id"])

            prev_coord = (
                self.stopovers[i - 1]["stop"]["location"]["latitude"],
                self.stopovers[i - 1]["stop"]["location"]["longitude"],
            )
            this_coord = (
                stopover["stop"]["location"]["latitude"],
                stopover["stop"]["location"]["longitude"],
            )

            # TODO normalize coordinates (distance is skewed towards longitude)
            prev_c = np.array(prev_coord)
            this_c = np.array(this_coord)
            req_c = np.array((lat, lon))
            min_dist = np.abs(
                np.cross(prev_c - this_c, this_c - req_c)
            ) / np.linalg.norm(prev_c - this_c)

            legs.append((prev_eva, this_eva))
            leg_distances.append(min_dist)

            # start with origin. (planned)arrival is always null in a previousStopovers list except for the last entry
            # (which is the stop where arrivals were requested)
            if ts and ts > now and not train_evas:
                train_evas = (prev_eva, this_eva)
                train_stops = (
                    self.stopovers[i - 1]["stop"]["name"],
                    stopover["stop"]["name"],
                )
                train_coords = (prev_coord, this_coord)

                # XXX known bug: we're saving departure at i-1 and (possibly) departure at i. For a more accurate coarse position estimate later on,
                # we need to track departure at i-1 and arrival at i. But we don't always have it.
                train_times = (self.stopovers[i - 1]["departure"], ts)

        if not train_evas:
            return

        if not train_times[0]:
            return

        closest_leg_index = np.argmin(leg_distances)
        closest_evas = legs[closest_leg_index]

        if train_evas != closest_evas:
            return

        self.set_stops(
            Stop(train_evas[0], train_stops[0], train_coords[0], train_times[0]),
            Stop(train_evas[1], train_stops[1], train_coords[1], train_times[1]),
        )

        # the time (i.e., number of minutes) the train needs to travel to reach the requested position
        # might be a better metric than raw distance.
        self.distance = distance(self.coarse_location, (lat, lon)).km

    def set_stops(self, prev_stop, next_stop):
        self.prev_stop = prev_stop
        self.next_stop = next_stop

        now = datetime.now(pytz.utc)

        self.progress_ratio = 1 - (
            (next_stop.when.timestamp() - now.timestamp())
            / (next_stop.when.timestamp() - prev_stop.when.timestamp())
        )
        self.progress_ratio = max(0, min(1, self.progress_ratio))

        if self.progress_ratio == 0:
            self.location = self.coarse_location = prev_stop.coord
        elif self.progress_ratio == 1:
            self.location = self.coarse_location = next_stop.coord
        else:
            ratio = self.progress_ratio
            self.coarse_location = (
                next_stop.coord[0] * ratio + prev_stop.coord[0] * (1 - ratio),
                next_stop.coord[1] * ratio + prev_stop.coord[1] * (1 - ratio),
            )

        if distance(prev_stop.coord, next_stop.coord).km < 20:
            # do not request polyline if the train is between stops less than 20km apart. This speeds up requests
            # (and reduces transport.rest load) at a hopefully low impact on accuracy.
            self.location = self.coarse_location

        if next_stop.eva == self.request_eva:
            # we can compare departure at previous stop with arrival at this stop. this is most accurate for position estimation.
            self.preferred = True

    def merge_with(self, instance):
        # might be useful in the future
        pass

    def to_json(self):
        return {
            "line": f"{self.train_type} {self.line_no}",
            "train": f"{self.train_type} {self.train_no}",
            "tripId": self.tripId,
            "location": self.coarse_location,
            "distance": round(self.distance, 1),
            "likelihood": round(100 - self.distance * 100 / max_distance, 0),
            "stops": [self.prev_stop.to_json(), self.next_stop.to_json()],
        }


headers = {
    "Access-Control-Allow-Origin": "*",
    "Content-Type": "application/json; charset=utf-8",
}

conn = psycopg2.connect(
    dbname=os.getenv("GEOLOOKUP_DBNAME", "geo_to_stations"),
    user=os.getenv("GEOLOOKUP_DBUSER", "geo_to_stations"),
    password=os.getenv("GEOLOOKUP_DBPASS"),
    host=os.getenv("GEOLOOKUP_DBHOST", "localhost"),
)
conn.autocommit = True
conn.set_session(readonly=True)

db_rest_api = os.getenv("GEOLOOKUP_DB_REST_API", "https://v5.db.transport.rest")

arrivals_request_count = 0
polyline_request_count = 0


async def set_location(train):
    trip_id = train.tripId
    line = f"{train.train_type} {train.line_no}"
    url = f"{db_rest_api}/trips/{trip_id}?lineName={line}&polyline=true"
    return
    logging.debug(f"Requesting polyline for {line}: {url}")
    global polyline_request_count
    polyline_request_count += 1
    async with aiohttp.ClientSession() as session:
        async with session.get(url) as response:
            content = await response.text()
            content = json.loads(content)


def is_in_transit(train):
    return 0 < train.progress_ratio < 1


async def handle_stats(request):
    response = {
        "arrivals_request_count": arrivals_request_count,
        "polyline_request_count": polyline_request_count,
    }
    return web.Response(body=json.dumps(response), headers=headers)


async def handle_search(request):
    try:
        lat = float(request.query.get("lat"))
        lon = float(request.query.get("lon"))
    except TypeError:
        return web.HTTPBadRequest(text="lat/lon are mandatory")
    except ValueError:
        return web.HTTPBadRequest(text="lat/lon must be floating-point numbers")

    global conn

    lut_lat = round(lat * 1000)
    lut_lon = round(lon * 1000)

    evas = set()

    try:
        with conn.cursor() as cur:
            cur.execute(
                "select stations from stations where lat between %s and %s and lon between %s and %s",
                (lut_lat - 3, lut_lat + 3, lut_lon - 3, lut_lon + 3),
            )
            for eva_list in cur.fetchall():
                evas.update(eva_list[0])
    except psycopg2.OperationalError as e:
        # Database may have been restarted
        logging.warning(f"Cannot handle train request: {e}")
        conn = psycopg2.connect(
            dbname=os.getenv("GEOLOOKUP_DBNAME", "geo_to_stations"),
            user=os.getenv("GEOLOOKUP_DBUSER", "geo_to_stations"),
            password=os.getenv("GEOLOOKUP_DBPASS"),
            host=os.getenv("GEOLOOKUP_DBHOST", "localhost"),
        )
        conn.autocommit = True
        conn.set_session(readonly=True)
        response = {
            "error": "database connection lost",
            "evas": list(),
            "trains": list(),
        }
        return web.Response(body=json.dumps(response), headers=headers)
    except psycopg2.InterfaceError as e:
        # Database may be down permanently. Notify supervisor that we can't work.
        logging.error(f"Cannot handle train request: {e}")
        sys.exit(1)

    if not evas:
        response = {"evas": list(), "trains": list()}
        return web.Response(body=json.dumps(response), headers=headers)

    arrivals = list()
    candidates = list()

    # deliberately not parallelized to minimize load on transport.rest
    for eva in evas:
        logging.debug(f"Requesting arrivals at {eva}")
        global arrivals_request_count
        arrivals_request_count += 1
        async with aiohttp.ClientSession() as session:
            async with session.get(
                f"{db_rest_api}/stops/{eva}/arrivals?results=40&duration=120&stopovers=true&bus=false&subway=false&tram=false"
            ) as response:
                content = await response.text()
                content = json.loads(content)
                arrivals.append(content)

    for train_list in arrivals:
        for train in train_list:
            for stop in train["previousStopovers"]:
                if (
                    int(stop["stop"]["id"]) in evas
                    and stop["stop"]["id"] != train["stop"]["id"]
                ):
                    candidates.append(Train(train))
                    break

    logging.debug(
        f"{len(candidates)} trains travel between at least two requested evas"
    )

    for train in candidates:
        train.set_coarse_location(lat, lon)

    candidates = list(filter(lambda train: train.coarse_location, candidates))
    logging.debug(f"{len(candidates)} trains have a coarse location")

    candidates = sorted(
        candidates, key=lambda train: 0 if train.preferred else train.distance
    )

    # remove duplicates. for now, we keep the preferred version, or the one with the lowest estimated distance.
    # later on, we'll need to request polylines and perform accurate calculations.
    # TODO polyline requests are not needed for trains currently located at a station (ratio == 0 / == 1)
    # It should also be fine to skip them if the distance between stops[0] and stops[1] is less than ~ 20km
    # Wenn sich ein Zug gerade an einem Bahnhof befindet (ratio == 0 / == 1) und mehrere km entfernt ist kann man ihn auch direkt ganz rausfiltern
    seen = dict()
    trains = list()

    for train in candidates:
        if train.preferred:
            trains.append(train)
            seen[train.train_no] = train
        elif train.train_no not in seen:
            trains.append(train)
            seen[train.train_no] = train
        else:
            seen[train.train_no].merge_with(train)

    logging.debug(f"{len(trains)} trains remain after deduplication")

    # If a train's coarse location is more than max_distance km away, it's fine (polyline-based) location is highly unlikely to be much closer
    trains = list(filter(lambda train: train.distance < max_distance, trains))
    need_fine = list(filter(lambda train: not train.location, trains))
    need_fine = list(filter(is_in_transit, trains))
    logging.debug(f"{len(need_fine)} trains need a polyline")

    for train in trains:
        await set_location(train)

    trains = sorted(trains, key=lambda train: train.distance)
    trains = list(map(lambda train: train.to_json(), trains[:10]))

    response = {"evas": list(evas), "trains": trains}

    return web.Response(body=json.dumps(response, ensure_ascii=False), headers=headers)


if __name__ == "__main__":

    parser = argparse.ArgumentParser(
        description="geolocation to train estimation service"
    )
    parser.add_argument(
        "--log-level",
        metavar="LEVEL",
        choices=["debug", "info", "warning", "error"],
        default="warning",
        help="Set log level",
    )
    parser.add_argument("--port", type=int, metavar="PORT", default=8080)
    parser.add_argument("--prefix", type=str, metavar="PATH", default="/")
    args = parser.parse_args()

    if args.log_level:
        numeric_level = getattr(logging, args.log_level.upper(), None)
        if not isinstance(numeric_level, int):
            print(f"Invalid log level: {args.log_level}", file=sys.stderr)
            sys.exit(1)
        logging.basicConfig(level=numeric_level)

    app = web.Application()
    app.add_routes(
        [
            web.get(f"{args.prefix}search", handle_search),
            web.get(f"{args.prefix}stats", handle_stats),
        ]
    )
    web.run_app(app, host="localhost", port=args.port)