ETL vs ELT: what actually changes when you move the T

ETL   source ──▶ [Spark / Informatica / Python] ──▶ warehouse (clean only)
                     transform BEFORE loading

ELT   source ──▶ warehouse (raw) ──▶ [SQL / dbt in-warehouse] ──▶ marts
                                        transform AFTER loading

In ETL, the warehouse only ever sees finished data. In ELT, the warehouse holds everything and becomes the transformation engine.

ETL — clean in PySpark, land only the result:

orders = spark.read.json("s3://raw/orders/")
clean = (orders
    .dropDuplicates(["order_id"])
    .withColumn("amount_usd", col("amount_cents") / 100)
    .filter(col("status").isNotNull())
    .drop("credit_card_number"))          # PII never reaches the warehouse

clean.write.mode("append").saveAsTable("warehouse.orders")

ELT — land everything, clean with SQL in the warehouse:

-- models/staging/stg_orders.sql  (dbt)
select distinct on (order_id)
  order_id,
  customer_id,
  amount_cents / 100.0 as amount_usd,
  status,
  ordered_at
from {{ source('raw', 'orders') }}
where status is not null
order by order_id, ingested_at desc

Note the difference in blast radius: the PySpark job dropped the credit card column permanently; the dbt model just doesn't select it, and the raw column still exists in the warehouse. That's a governance decision, not a style preference.

• Regulated PII must be masked or tokenized before it lands anywhere.
• You're filtering 10 TB down to 50 GB — paying to store the rest is waste.
• The transformation isn't expressible in SQL (image parsing, ML inference, complex parsing).
• The destination isn't a warehouse (a search index, an API, a key-value store).

• You want to reprocess history when business logic changes — and you will.
• Analysts should own transformations, not file tickets for them.
• Your warehouse (Snowflake, BigQuery, Databricks) scales compute independently.
• You want lineage, tests and docs generated automatically from the code.

A hybrid. Sensitive or huge sources get an ETL pre-step (mask, sample, pre-aggregate), everything else lands raw and gets ELT'd with dbt into a medallion layout. If you can explain why a specific source sits on either side, you're already answering this better than most candidates.