How do you handle schema evolution in data pipelines without breaking downstream consumers?

How to think about it

Schema evolution is one of the most common causes of silent pipeline breakage. A backend engineer adds a column, renames a field, or changes a type — downstream consumers that were not notified break quietly or start producing nulls.

Compatibility modes

Schema registries (Confluent Schema Registry, AWS Glue Schema Registry) enforce compatibility before a producer can publish a new schema version:

Mode	Allows
BACKWARD	New schema can read data written by old schema
FORWARD	Old schema can read data written by new schema
FULL	Both directions — strictest, safest for multi-team pipelines
NONE	No check — dangerous in production

BACKWARD-compatible changes (safe):

• Adding a column with a default value or null
• Adding an optional Avro/Protobuf field

Breaking changes:

• Removing a required column
• Renaming a column
• Changing a column type (e.g., INT to STRING)

Handling evolution in practice

Avro / Protobuf with schema registry:

# Producer registers schema before first publish
schema_registry_client.register_schema("orders-value", new_avro_schema)
# Registry checks compatibility; rejects if BACKWARD is violated

Parquet / Delta Lake — column addition:

# Delta Lake: adding a column is safe with mergeSchema
df.write.format("delta") \
    .option("mergeSchema", "true") \
    .mode("append") \
    .save("s3://my-bucket/orders/")

Iceberg — full schema evolution without rewriting data:

Iceberg tracks schema changes by column ID (not name), so renaming a column does not require rewriting Parquet files — the metadata layer maps the new name to the existing column ID.

-- Iceberg: rename column without touching data files
ALTER TABLE orders RENAME COLUMN amt TO amount_usd;

dbt contract enforcement

dbt 1.5+ supports model contracts: explicitly declare column names and types. A model that changes a column type or removes a declared column fails CI before it reaches production.

models:
  - name: orders_daily
    config:
      contract:
        enforced: true
    columns:
      - name: order_date
        data_type: date
      - name: revenue_usd
        data_type: numeric