def get_destinations(dfs, roam_dist=110, earth_radius=6372.795 * 1000):
"""
Applies DBSCAN to extract the unique stop locations from a pyspark DataFrame
:param x: DataFrame with ['id_client', 'latitude', 'longitude', "from", "to"]. Coordinates are in degrees.
:param roam_dist: The stop location size in meters.
:param earth_radius: The radius of the earth.
:param group_results: If True, it groups by the cluster's location and id_client.
:return: (pyspark DataFrame) If group_results=True: ['id_client', 'clatitude', 'clongitude', 'time_spent', 'frequency']
(pyspark DataFrame) If group_results=False: ['id_client', 'latitude', 'longitude', 'clatitude', 'clongitude', 'from', 'to']
"""
@pandas_udf(
"userId string, state string, latitude double, longitude double, begin timestamp, end timestamp, clusterId integer",
PandasUDFType.GROUPED_MAP)
def get_destinations(df):
"""
Applies DBSCAN to stop locations
:param x: 2D numpy array with latitude and longitude.
:param from_to_array: 2D numpy array with from and to timestamps.
:param roam_dist: The stop location size in meters.
:param earth_radius: The radius of the earth.
:return: (pandas DataFrame) ['latitude', 'longitude', 'clatitude', 'clongitude', 'from', 'to', 'time_spent']
"""
db = DBSCAN(eps=roam_dist / earth_radius,
min_samples=1,
algorithm='ball_tree',
metric='haversine')
df["clusterId"] = db.fit_predict(df[['latitude', 'longitude']])
return df
dfs = dfs.withColumn('latitude', F.radians('latitude'))
dfs = dfs.withColumn('longitude', F.radians('longitude'))
stops_dfs = dfs.groupby('userId', 'state').apply(get_destinations)
stops_dfs = stops_dfs.withColumn('latitude', F.degrees('latitude'))
stops_dfs = stops_dfs.withColumn('longitude', F.degrees('longitude'))
w = Window().partitionBy('userId', 'clusterId')
stops_dfs = stops_dfs.withColumn('clusterLatitude',
F.mean('latitude').over(w))
stops_dfs = stops_dfs.withColumn('clusterLongitude',
F.mean('longitude').over(w))
stops_dfs = stops_dfs.drop('latitude').drop('longitude')
return stops_dfs
def add_solar_features(df):
return (df.withColumn(
"declination_angle",
radians(-23.45 * cos(((2 * pi) / 365) * (dayofyear("date") + 10))),
).withColumn("diff_local_time_UTC", timezone_from_date("date")).withColumn(
"d", (2 * pi * dayofyear("date")) / 365).withColumn(
"equation_of_time",
-7.655 * sin(col("d")) + 9.873 * sin(2 * col("d") + 3.588),
).drop("d").withColumn(
"time_correction",
4 * (col("loc_long") - (15 * col("diff_local_time_UTC"))) +
col("equation_of_time"),
).withColumn(
"local_solar_hour",
col("hour") + 0.5 + col("time_correction") / 60).withColumn(
"hour_angle", 0.2618 * (col("local_solar_hour") - 12)).drop(
"diff_local_time_UTC",
"equation_of_time",
"time_correction",
"local_solar_hour",
).withColumn(
"solar_elevation",
degrees(
asin(
sin("declination_angle") *
sin(radians("loc_lat")) +
cos("declination_angle") *
cos(radians("loc_lat")) * cos("hour_angle"))),
).drop("declination_angle", "hour_angle"))
def add_solar_features(df):
return \
(df
.withColumn('declination_angle',
radians(-23.45
* cos(((2 * pi)/365) * (dayofyear('date') + 10))))
.withColumn('diff_local_time_UTC', timezone_from_date('date'))
.withColumn('d', (2 * pi * dayofyear('date')) / 365)
.withColumn('equation_of_time',
-7.655 * sin(col('d'))
+ 9.873 * sin(2 * col('d') + 3.588))
.drop('d')
.withColumn('time_correction',
4 * (col('loc_long') - (15 * col('diff_local_time_UTC')))
+ col('equation_of_time'))
.withColumn('local_solar_hour',
col('hour') + 0.5 + col('time_correction') / 60)
.withColumn('hour_angle', 0.2618 * (col('local_solar_hour') - 12))
.drop('diff_local_time_UTC', 'equation_of_time', 'time_correction',
'local_solar_hour')
.withColumn('solar_elevation',
degrees(asin(sin('declination_angle')
* sin(radians('loc_lat'))
+ cos('declination_angle')
* cos(radians('loc_lat'))
* cos('hour_angle'))))
.drop('declination_angle', 'hour_angle'))
def calculate_bearing_degrees(latitude_1, longitude_1, latitude_2,
longitude_2):
diff_longitude = F.radians(longitude_2 - longitude_1)
r_latitude_1 = F.radians(latitude_1)
r_longitude_1 = F.radians(longitude_1)
r_latitude_2 = F.radians(latitude_2)
r_longitude_2 = F.radians(longitude_2)
y = F.sin(diff_longitude) * F.cos(r_longitude_2)
x = (F.cos(r_latitude_1) * F.sin(r_latitude_2) -
F.sin(r_latitude_1) * F.cos(r_latitude_2) * F.cos(diff_longitude))
return F.degrees(F.atan2(x, y))
def main():
# Parameters for the algorithm
roam_dist = 100 # meters
min_stay = 10 # minutes
# Parameters for the paths
input_path = '/path_to_parquet' # parquet file
output_path = '/path_to_parquet_out' # parquet file
spark = SparkSession \
.builder \
.appName("Stop locations") \
.getOrCreate()
# Read data
source_df = spark.read.parquet(input_path)
source_df = source_df.select(
'user_id', 'timestamp',
F.radians('latitude').alias('lat'),
F.radians('longitude').alias("lon")).orderBy('timestamp')
source_df.cache()
# Filter out all the data that is not necessary (e.g. positions equals to others in a time-distance less than min_stay
w = Window.partitionBy(['user_id']).orderBy('timestamp')
source_df = source_df.select("user_id", "timestamp", "lat", "lon",
F.lead("lat", 1).over(w).alias("next_lat"),
F.lead("lon", 1).over(w).alias("next_lon"))
dist_df = source_df.withColumn(
"distance_next", EARTH_RADIUS * 2 * F.asin(
F.sqrt(
F.pow(F.sin((col("next_lat") - col("lat")) / 2.0), 2) +
F.cos("lat") * F.cos("next_lat") *
F.pow(F.sin((col("next_lon") - col("lon")) / 2.0), 2))))
dist_df = dist_df.withColumn("distance_prev",
F.lag("distance_next").over(w))
exclude_df = dist_df.where((
(col("distance_next") < 5) & (col("distance_prev") < 5))
| ((col("distance_next") > roam_dist)
& (col("distance_prev") > roam_dist)))
df = source_df.join(exclude_df, ['user_id', 'timestamp'],
"left_anti").select("user_id", "timestamp", "lat",
"lon")
# Transform to RDD, in order to apply the function get_stop_location
# RDD that contains: (user_id, [timestamp, lat, lon, lat_degrees, lon_degrees])
df_rdd = df.orderBy(['user_id', 'timestamp']).rdd.map(tuple)
df_rdd = df_rdd.map(lambda x: (x[0], [[x[1], x[2], x[3]]]))
# RDD that contains: (user_id, [[timestamp, lat, lon], ..., [timestamp, lat, lon]]), sorted by timestamp
grouped_rdd = df_rdd.reduceByKey(lambda x, y: x + y)
stop_locations_rdd = grouped_rdd.map(lambda x: (
x[0],
get_stop_location(
x[1], min_stay_duration=min_stay, roaming_distance=roam_dist)))
stop_locations_rdd = stop_locations_rdd.flatMapValues(lambda x: x).map(
lambda x: (x[0], x[1][0], x[1][1], x[1][2], x[1][3]))
# Output schema
schema = StructType([
StructField('user_id', StringType(), False),
StructField('lat', DoubleType(), False),
StructField('lon', DoubleType(), False),
StructField('from', TimestampType(), False),
StructField('to', TimestampType(), False)
])
result_df = spark.createDataFrame(stop_locations_rdd, schema)
result_df = result_df.withColumn('lat', F.degrees('lat'))
result_df = result_df.withColumn('lon', F.degrees('lon'))
result_df.write.save(output_path)
def main():
"""Main function"""
# Get args
args = get_args()
# Azure credentials
sas_token = args.sas
storage_account_name = args.storage
container_in = args.container_in
container_out = args.container_out
azure_accounts = list()
azure_accounts.append({
"storage": storage_account_name,
"sas": sas_token,
"container": container_in
})
azure_accounts.append({
"storage": storage_account_name,
"sas": sas_token,
"container": container_out
})
# VM
cores = args.vm_cores
ram = args.vm_ram
shuffle_partitions = args.shuffle_partitions
# Geohash file path
geohash_path = args.geohashpath
# Date, country, prefix
country = args.country
date_string = args.date
prefix = args.prefix
# Set date variables
day_time = datetime.strptime(date_string, "%Y-%m-%d")
year = day_time.year
month = day_time.month
day = day_time.day
# stop config
seconds = 60
accuracy = args.accuracy
roam_dist = args.roam_dist
min_stay = args.min_stay
overlap_hours = args.overlap_hours
# Path in - path out
blob_in = f"wasbs://{container_in}@{storage_account_name}.blob.core.windows.net/preprocessed/{country}/"
path_out = f"stoplocation-v{VERSION}_r{roam_dist}-s{min_stay}-a{accuracy}-h{overlap_hours}/{country}"
if prefix:
path_out = f"stoplocation-v{VERSION}_prefix_r{roam_dist}-s{min_stay}-a{accuracy}-h{overlap_hours}/{country}"
# config spark
conf = getSparkConfig(cores, ram, shuffle_partitions, azure_accounts)
# Create spark session
sc = SparkContext(conf=conf).getOrCreate()
sqlContext = SQLContext(sc)
spark = sqlContext.sparkSession
# Init azure client
blob_service_client = BlobServiceClient.from_connection_string(
CONN_STRING.format(storage_account_name, sas_token))
# build keys, date is mandatory, prefix opt
partition_key = "year={}/month={}/day={}".format(year, month, day)
if prefix:
partition_key = "year={}/month={}/day={}/prefix={}".format(
year, month, day, prefix)
blob_base = "{}/{}".format(path_out, partition_key)
#
# check for skip
# TODO
#
skip = False
print("process " + partition_key + " to " + blob_base)
start_time = time.time()
local_dir = LOCAL_PATH + partition_key
print("write temp to " + local_dir)
# cleanup local if exists
if (os.path.isdir(local_dir)):
map(os.unlink,
(os.path.join(local_dir, f) for f in os.listdir(local_dir)))
# TODO cleanup remote if exists
# Output schema
schema = ArrayType(
StructType([
#StructField('device_type', IntegerType(), False),
StructField('serial', IntegerType(), False),
StructField('latitude', DoubleType(), False),
StructField('longitude', DoubleType(), False),
StructField('begin', TimestampType(), False),
StructField('end', TimestampType(), False),
StructField('personal_area', BooleanType(), False),
StructField('distance', DoubleType(), False),
StructField('geohash6', StringType(), False),
StructField('after_stop_distance', DoubleType(), False)
]))
spark_get_stop_location = udf(
lambda z: get_stop_location(z, roam_dist, min_stay), schema)
# Geohash file
print("read geohash parquet")
csv_time = time.time()
dfs_us_states = spark.read.format("parquet").load(geohash_path)
# states = [s.STUSPS for s in dfs_us_states.select(
# 'STUSPS').distinct().collect()]
dfs_us_states = dfs_us_states.select(
col('STUSPS').alias('state'),
col('geohash').alias('geohash5'))
dfs_us_states = dfs_us_states.drop_duplicates(subset=['geohash5'])
# Input dataset
print("read dataset table")
read_time = time.time()
# dfs = spark.read.format("parquet").load(blob_in)
# # apply partition filter
# dfs_partition = dfs.where(
# f"(year = {year} AND month = {month} AND day = {day} AND prefix = '{prefix}')")
# read only partition to reduce browse time
dfs_cur_partition = spark.read.format("parquet").load(
f"{blob_in}/{partition_key}")
# lit partition filters as data
dfs_cur_partition = dfs_cur_partition.withColumn('year', F.lit(year))
dfs_cur_partition = dfs_cur_partition.withColumn('month', F.lit(month))
dfs_cur_partition = dfs_cur_partition.withColumn('day', F.lit(day))
if prefix:
dfs_cur_partition = dfs_cur_partition.withColumn(
'prefix', F.lit(prefix))
# read next day for overlap
next_day = day_time + timedelta(days=1)
next_partition_key = "year={}/month={}/day={}".format(
next_day.year, next_day.month, next_day.day)
if prefix:
next_partition_key = "year={}/month={}/day={}/prefix={}".format(
next_day.year, next_day.month, next_day.day, prefix)
dfs_next_partition = spark.read.format("parquet").load(
f"{blob_in}/{next_partition_key}")
dfs_next_partition = dfs_next_partition.where(
F.hour("timestamp") <= (overlap_hours - 1))
# lit partition filters as data
dfs_next_partition = dfs_next_partition.withColumn('year',
F.lit(next_day.year))
dfs_next_partition = dfs_next_partition.withColumn('month',
F.lit(next_day.month))
dfs_next_partition = dfs_next_partition.withColumn('day',
F.lit(next_day.day))
if prefix:
dfs_next_partition = dfs_next_partition.withColumn(
'prefix', F.lit(prefix))
# union with overlap
dfs_partition = dfs_cur_partition.unionAll(dfs_next_partition)
print("process with spark")
spark_time = time.time()
# select columns
dfs_partition = dfs_partition.select(
'prefix', 'userID', 'timestamp', 'latitude', 'longitude',
(F.when(col('opt1') == 'PERSONAL_AREA',
True).otherwise(False)).alias('personal_area'), 'accuracy')
# keep only data with required accuracy
dfs_partition = dfs_partition.where((col('accuracy') <= accuracy)
& (col('accuracy') >= 0))
# stats - enable only for debug!
# num_inputs = dfs_partition.count()
# print(f"read {num_inputs} rows from "+partition_key)
# Lowering the granularity to 1 minutes
# explicitely convert to timestamp
#dfs_partition = dfs_partition.withColumn('timestamp', col('timestamp').cast('timestamp'))
seconds_window = F.unix_timestamp(
'timestamp') - F.unix_timestamp('timestamp') % seconds
w = Window().partitionBy('userID', seconds_window).orderBy('accuracy')
dfs_partition = dfs_partition.withColumn(
'rn',
F.row_number().over(w).cast('int')).where(col('rn') == 1).drop('rn')
# Radians lat/lon
dfs_partition = dfs_partition.withColumn('latitude',
F.radians('latitude')).withColumn(
'longitude',
F.radians('longitude'))
# Groups GPS locations into chucks. A chunk is formed by groups of points that are distant no more than roam_dist
w = Window.partitionBy(['prefix', 'userID']).orderBy('timestamp')
dfs_partition = dfs_partition.withColumn('next_lat',
F.lead('latitude', 1).over(w))
dfs_partition = dfs_partition.withColumn('next_lon',
F.lead('longitude', 1).over(w))
# Haversine distance
dfs_partition = dfs_partition.withColumn(
'distance_next', EARTH_RADIUS * 2 * F.asin(
F.sqrt(
F.pow(F.sin((col('next_lat') - col('latitude')) / 2.0), 2) +
F.cos('latitude') * F.cos('next_lat') *
F.pow(F.sin((col('next_lon') - col('longitude')) / 2.0), 2))))
dfs_partition = dfs_partition.withColumn(
'distance_prev',
F.lag('distance_next', default=0).over(w))
# Chunks
dfs_partition = dfs_partition.withColumn(
'chunk',
F.when(col('distance_prev') > roam_dist, 1).otherwise(0))
windowval = (Window.partitionBy(
'prefix',
'userID').orderBy('timestamp').rangeBetween(Window.unboundedPreceding,
0))
dfs_partition = dfs_partition.withColumn(
'chunk',
F.sum('chunk').over(windowval).cast('int'))
# Remove chunks of the next day
w = Window.partitionBy(['prefix', 'userID', 'chunk'])
dfs_partition = dfs_partition.withColumn(
'min_timestamp', F.dayofmonth(F.min('timestamp').over(w)))
dfs_partition = dfs_partition.where(
col('min_timestamp') == day).drop('min_timestamp')
# Get the stops
result_df = dfs_partition.groupBy('prefix', 'userID', 'chunk').agg(
F.array_sort(
F.collect_list(
F.struct('timestamp', 'latitude', 'longitude', 'distance_prev',
'personal_area'))).alias('gpsdata'),
F.sum('distance_prev').alias('dist_sum'))
result_df = result_df.withColumn('gpsdata',
spark_get_stop_location('gpsdata'))
result_df = result_df.select('userID', 'chunk',
F.explode_outer('gpsdata').alias('e'),
'dist_sum')
result_df = result_df.select(
'userID', 'chunk',
col('e.latitude').alias('latitude'),
col('e.longitude').alias('longitude'),
col('e.begin').alias('begin'),
col('e.end').alias('end'),
col('e.personal_area').alias('personal_area'),
col('e.geohash6').alias('geohash6'),
col('e.serial').alias('serial'),
col('e.distance').alias('stop_distance'),
col('e.after_stop_distance').alias('after_stop_distance'), 'dist_sum')
result_df = result_df.fillna(0, subset=['after_stop_distance'])
# Remove all those stop that start the next day
result_df = result_df.where((col('begin').isNull())
| (F.dayofmonth('begin') != next_day.day))
result_df = result_df.withColumn(
'isStop',
F.when(col('serial').isNotNull(), 1).otherwise(0))
result_df = result_df.withColumn(
'dist_sum',
F.when(col('isStop') == 1,
col('stop_distance')).otherwise(col('dist_sum')))
windowval = (Window.partitionBy('userId').orderBy(
'chunk', 'serial').rowsBetween(Window.currentRow,
Window.unboundedFollowing))
result_df = result_df.withColumn('isStop_cum',
F.sum('isStop').over(windowval))
result_df = result_df.groupBy('userId', 'isStop_cum').agg(
F.first('latitude', ignorenulls=True).alias('latitude'),
F.first('longitude', ignorenulls=True).alias('longitude'),
F.first('begin', ignorenulls=True).alias('begin'),
F.first('end', ignorenulls=True).alias('end'),
F.first('personal_area', ignorenulls=True).alias('personal_area'),
F.first('geohash6', ignorenulls=True).alias('geohash6'),
F.sum('dist_sum').alias('prev_travelled_distance'),
F.sum('after_stop_distance').alias('after_stop_distance'))
# compute next distance, which is null if it's the last
windowval = Window.partitionBy('userId').orderBy(F.desc('isStop_cum'))
result_df = result_df.withColumn(
'next_travelled_distance',
F.lead('prev_travelled_distance').over(windowval))
result_df = result_df.withColumn(
'next_travelled_distance',
F.when((col('next_travelled_distance').isNull()) &
(col('after_stop_distance') > 0),
col('after_stop_distance')).otherwise(
col('next_travelled_distance')))
# Drop nulls
result_df = result_df.dropna(subset=['latitude']).drop('isStop_cum')
# Transform latitude and longitude back to degrees
result_df = result_df.withColumn('latitude', F.degrees('latitude'))
result_df = result_df.withColumn('longitude', F.degrees('longitude'))
# US states
result_df = result_df.withColumn(
"geohash5", F.expr("substring(geohash6, 1, length(geohash6)-1)"))
result_df = result_df.join(F.broadcast(dfs_us_states),
on="geohash5",
how="inner").drop('geohash5')
# lit partition data - enable only if added to partitionBy
# result_df = result_df.withColumn('year', F.lit(year))
# result_df = result_df.withColumn('month', F.lit(month))
# result_df = result_df.withColumn('day', F.lit(day))
# write
out_partitions = len(US_STATES)
result_df.repartition(out_partitions, "state").write.partitionBy(
"state").format('parquet').mode("overwrite").save(local_dir + "/")
# stats - enable only for debug!
# num_records = result_df.count()
# print(f"written {num_records} rows to "+local_dir)
# if num_records == 0:
# raise Exception("Zero rows output")
print("upload local data to azure")
upload_time = time.time()
# upload parts over states
for state in US_STATES:
print(f"upload files for {state}")
state_dir = local_dir + "/state=" + state
state_key = f"{partition_key}/state={state}/"
if (os.path.isdir(state_dir)):
files = [
filename for filename in os.listdir(state_dir)
if filename.startswith("part-")
]
if len(files) > 0:
for file_local in files:
file_path = state_dir + "/" + file_local
part_num = int(file_local.split('-')[1])
part_key = '{:05d}'.format(part_num)
# fix name as static hash to be reproducible
filename_hash = hashlib.sha1(
str.encode(state_key + part_key)).hexdigest()
blob_key = "{}/state={}/part-{}-{}.snappy.parquet".format(
blob_base, state, part_key, filename_hash)
print("upload " + file_path + " to " + container_out +
":" + blob_key)
blob_client = blob_service_client.get_blob_client(
container_out, blob_key)
with open(file_path, "rb") as data:
blob_client.upload_blob(data, overwrite=True)
# cleanup
os.remove(file_path)
else:
print(f"no files to upload for {state}")
else:
print(f"missing partition for {state}")
print("--- {} seconds elapsed ---".format(int(time.time() - start_time)))
print()
stop_time = time.time()
spark.stop()
end_time = time.time()
print("Done in {} seconds (csv:{} read:{} spark:{} upload:{} stop:{})".
format(int(end_time - start_time), int(read_time - csv_time),
int(spark_time - read_time), int(upload_time - spark_time),
int(stop_time - upload_time), int(end_time - stop_time)))
print('Done.')
from df_tools import *
from histfile import *
from tools import *
from pyspark.sql import functions as F
import sys
nside = int(sys.argv[1])
spark = SparkSession.builder.getOrCreate()
df = spark.read.parquet("nside{}.parquet".format(nside))
df = df.withColumn(
"dx",
F.degrees(
F.sin((df["theta"] + df["theta_c"]) / 2) * (df["phi"] - df["phi_c"])) *
60)
df = df.withColumn("dy", F.degrees(df["theta"] - df["theta_c"]) * 60)
#df=df.withColumn("r",F.hypot(df["dx"],df["dy"]))
df = df.withColumn("x",
F.sin(df["theta"]) * F.cos(df["phi"])).withColumn(
"y",
F.sin(df["theta"]) * F.sin(df["phi"])).withColumn(
"z", F.cos(df["theta"])).drop("theta", "phi")
df = df.withColumn("xc",
F.sin(df["theta_c"]) * F.cos(df["phi_c"])).withColumn(
"yc",
F.sin(df["theta_c"]) * F.sin(df["phi_c"])).withColumn(
"zc",
