Modelando dados com PySpark e SQL

Transformação de Dados com Spark SQL no Databricks

Disha Mukherjee

Lead Data Engineer

Modelando seus dados

path = "/Volumes/.../default/transactions.csv"
df = spark.read.csv(path, header=True, inferSchema=True)
df.show(5, truncate=False)

Visualizando saídas no Databricks

select() → escolher colunas
filter() → escolher linhas

Selecionando colunas

df.select("ID", "Customer_ID", "Transaction_Amount", "Category").show(5)

+---+-----------+------------------+-----------+
| ID|Customer_ID|Transaction_Amount|   Category|
+---+-----------+------------------+-----------+
|  1|    CUST001|          27337.49|Electronics|
|  2|    CUST002|           97716.6|Electronics|
|  3|    CUST003|           5752.36|Electronics|
|  4|    CUST004|          93443.22|    Savings|
|  5|    CUST005|          15109.98|Electronics|
+---+-----------+------------------+-----------+

Filtrando linhas

from pyspark.sql import functions as F


filtered_df = df.filter(F.col("Transaction_Status") == "Completed")
filtered_df.show(5)

+---+-------------------+-----------+------------------+------------------+
| ID|               Date|Customer_ID|Transaction_Amount|Transaction_Status|
+---+-------------------+-----------+------------------+------------------+
|  3|2023-01-03 00:00:00|    CUST003|           5752.36|         Completed|
|  9|2023-01-09 00:00:00|    CUST009|          28959.12|         Completed|
| 10|2023-01-10 00:00:00|    CUST010|          72098.18|         Completed|
| 11|2023-01-11 00:00:00|    CUST011|          49771.05|         Completed|
| 20|2023-01-20 00:00:00|    CUST020|          69825.14|         Completed|
+---+-------------------+-----------+------------------+------------------+

Combinando condições de filtro

high_value_df = df.filter(

    (F.col("Transaction_Amount") > 20000) &

    (F.col("Transaction_Status") == "Completed")

)
high_value_df.show(3)

+---+-------------------+-----------+------------------+------------------+
| ID|               Date|Customer_ID|Transaction_Amount|Transaction_Status|
+---+-------------------+-----------+------------------+------------------+
|  9|2023-01-09 00:00:00|    CUST009|          28959.12|         Completed|
| 10|2023-01-10 00:00:00|    CUST010|          72098.18|         Completed|
| 11|2023-01-11 00:00:00|    CUST011|          49771.05|         Completed|
+---+-------------------+-----------+------------------+------------------+

Métodos Python ou SQL

Métodos Python: select(), filter()
- Construa transformações passo a passo

Consultas SQL: sintaxe SQL padrão
- Mais legível para agregações e agrupamentos

Python vs SQL

Criando uma temporary view

df.createOrReplaceTempView("transactions")

Temporary view = nome que aponta para o DataFrame
Dura só na sessão atual do Spark

Consultando com SQL

%sql
SELECT Category, COUNT(*) AS transaction_count

FROM transactions

WHERE Transaction_Status = 'Completed'
GROUP BY Category
ORDER BY transaction_count DESC

+-----------+-----------------+
|   Category|transaction_count|
+-----------+-----------------+
|   Clothing|             6764|
|     Dining|             6732|
|Electronics|             6640|
|    Savings|             6629|
|  Groceries|             6557|
+-----------+-----------------+

Validando com display()

display(filtered_df)

GIF mostrando tabela interativa com ordenação e busca

Ordenar e paginar resultados
Identificar nulos e outliers

Inspecionando planos de consulta

filtered_df.explain()

== Physical Plan ==
*(1) Filter (Transaction_Status = Completed)
+- FileScan csv [ID, Date, Customer_ID, ...]

Também funciona com SQL: EXPLAIN SELECT ...
Útil para resolver problemas de desempenho

Lendo erros do Spark

df.select("Account").show()

[UNRESOLVED_COLUMN.WITH_SUGGESTION] A column with name `Account` 
cannot be resolved. Did you mean one of the following? 
[`Date`, `Location`, `Customer_ID`, `Transaction_Amount`]

Vamos praticar!

Transformação de Dados com Spark SQL no Databricks