Data Engineer

Kā rīkoties, ja user_id ir skew (nepareiza sadalīšana) pievienojoties transakciju un lietotāju tabulām? Kā izvēlēties optimālo sadalījuma atslēgu?

importē clickhouse_driver from airflow.hooks.base import * def get_clickhouse_client(): conn = BaseHook.get_connection("clickhouse_default") return clickhouse_driver.Client( host=conn.host, port=conn.port, user=conn.login, password=conn.password, database=conn.schema )

Kreisais apvienojums: tabula ar 10 ierakstiem Kreisais apvienojums: tabula ar 100 ierakstiem. Kāds ir minimālais un maksimālais rindu skaits, ko var iegūt?

Kas ir sadalīšana un izplatīšana (sharding)?

CREATE TABLE core.localUserMetadata ON CLUSTER cluster_4x2 ( UserId String, Country String, LastLoginDate DateTime ) ENGINE = ReplikatsiyuAylantiradiganMergeTree() ORDER BY (UserId); CREATE TABLE core.userMetadata ON CLUSTER cluster_4x2 ( UserId String, Country String, LastLoginDate DateTime ) ENGINE = Tarqatilgan('cluster_4x2', 'core', 'localUserMetadata', cityHash64(LastLoginDate));

no interview.utils importēt get_clickhouse_client no airflow importēt DAG no airflow.operators.python importēt PythonOperator no airflow.sensors.external_task importēt ExternalTaskSensor no datetime importēt datetime importēt pandas kā pd importēt clickhouse_driver importēt os CLICKHOUSE_CLIENT = get_clickhouse_client() default_args = { "start_date": datetime(2024, 1, 1) } ar DAG( dag_id="datamarts.daily_revenue_per_country", default_args=default_args, schedule_interval="@daily", catchup=False ) kā dag: transactions_sensor = S3KeySensor( task_id="transactions_sensor", bucket_key="data/transactions_{}.csv".format(datetime.now().strftime("%Y-%m-%d")), bucket_name="my-bucket", aws_conn_id="aws_default", timeout=600, poke_interval=30, mode="poke" ) def extract_from_s3(**kwargs): df = pd.read_csv("s3://my-bucket/data/transactions_{}.csv".format(datetime.now().strftime("%Y-%m-%d"))) kwargs["ti"].xcom_push(key="df", value=df.to_dict()) def load_to_raw_table(**kwargs): df = pd.DataFrame(kwargs["ti"].xcom_pull(task_ids="extract", key="df")) rows = [tuple(r) for r in df[["transaction_id", "user_id", "amount", "created_at"]].to_numpy()] CLICKHOUSE_CLIENT.execute( ... )

def extract_from_s3(**kwargs): df = pd.read_csv("s3://manobucket/data/transactions_{}.csv".format(datetime.now().strftime("%Y-%m-%d"))) kwargs["ti"].xcom_push(key="df", value=df.to_dict()) def load_to_raw_table(**kwargs): df = pd.DataFrame(kwargs["ti"].xcom_pull(task_ids="extract", key="df")) rows = [tuple(r) for r in df[["transaction_id", "user_id", "amount", "created_at"]].to_numpy()] CLICKHOUSE_CLIENT.execute( "INSERT INTO raw.transactions (transaction_id, user_id, amount, created_at) VALUES", rows ) def build_aggregate_view(): query = """ INSERT INTO datamarts.daily_revenue_per_country SELECT toDate(r.created_at) as event_date, u.country, sum(r.amount) as total_revenue FROM raw.transactions r LEFT JOIN core.userMetadata u ON r.user_id = u.UserId WHERE toDate(r.created_at) = '{}' GROUP BY event_date, u.country """.format(datetime.now().strftime("%Y-%m-%d")) CLICKHOUSE_CLIENT.execute(query) transactions_sensor = S3KeySensor( task_id="transactions_sensor", bucket_key="data/transactions_{}.csv".format(datetime.now().strftime("%Y-%m-%d")), bucket_name="my-bucket", aws_conn_id="aws_default", timeout=600, poke_interval=30, mode="poke" ) extract = PythonOperator( task_id="extract", python_callable=extract_from_s3, provide_context=True ) load = PythonOperator( task_id="load", python_callable=load_to_raw_table, provide_context=True ) aggregate = PythonOperator( task_id="aggregate", python_callable=build_aggregate_view ) transactions_sensor >> extract >> load >> aggregate

Kuru resursu visvairāk patērē Nested Loop Join?

IZVEIDOJOT raw.local_transactions TABULU KLASTERĒ cluster_4x2 ( transaction_id String, user_id String, amount Float64, created_at DateTime ) MOTORS = ReplicatedMergeTree() PARTĪCIONĒT PĒC amount KĀRTOT PĒC (transaction_id); Izveidojiet tabulu raw.transactions KLASTERĒ cluster_4x2 ( transaction_id String, user_id String, amount Float64, created_at DateTime ) MOTORS = Distributed('cluster_4x2', 'raw', 'local_transactions', cityHash64(user_id));

CREATE TABLE datamarts.daily_revenue_per_country ON CLUSTER cluster_4x2 ( event_date Date, country String, total_revenue Float64 ) ENGINE = MergeTree() PARTITION BY toYYYYMM(event_date) ORDER BY (event_date);

Kāda ir atšķirība starp ROWS BETWEEN un RANGE BETWEEN?

Logrāmatas funkcija SUM() OVER (PARTITION BY user_id) — vienā gadījumā pievienojam ORDER BY pirkuma datumu, citā — ne. Kāda ir atšķirība?

Kuri fiziskie JOIN algoritmi darbojas datu bāzēs?

NULL plus 5 — cik tas būs?