Data Engineer

How to handle skew in user_id when JOINing transaction and user tables? How to choose the optimal sharding key?

import 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 )

LEFT JOIN: a table with 10 records LEFT JOIN a table with 100 records. What is the minimum and maximum number of rows that can be obtained?

What is partitioning and sharding (distribution)?

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

from interview.utils import get_clickhouse_client from airflow import DAG from airflow.operators.python import PythonOperator from airflow.sensors.external_task import ExternalTaskSensor from datetime import datetime import pandas as pd import clickhouse_driver import os CLICKHOUSE_CLIENT = get_clickhouse_client() default_args = { "start_date": datetime(2024, 1, 1) } with DAG( dag_id="datamarts.daily_revenue_per_country", default_args=default_args, schedule_interval="@daily", catchup=False ) as 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://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( "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

In Nested Loop Join, which resource is consumed the most?

CREATE TABLE raw.local_transactions ON CLUSTER cluster_4x2 ( transaction_id String, user_id String, amount Float64, created_at DateTime ) ENGINE = ReplicatedMergeTree() PARTITION BY amount ORDER BY (transaction_id); CREATE TABLE raw.transactions ON CLUSTER cluster_4x2 ( transaction_id String, user_id String, amount Float64, created_at DateTime ) ENGINE = 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);

What is the difference between ROWS BETWEEN and RANGE BETWEEN?

The window function SUM() OVER (PARTITION BY user_id) — in one case adding ORDER BY purchase date, in another not. What's the difference?

What physical algorithms of JOIN work in databases?

NULL plus 5 — what will it be?