When would you use Parquet versus Avro, and why?

Parquet is columnar, so it compresses well and lets a query read only the columns it needs, which makes it right for analytical reads. Avro is row oriented with a compact binary encoding and strong schema evolution support, which makes it right for write heavy paths and message payloads where you consume whole records. In short, Avro for transport and landing, Parquet for the analytical tables you query.

The difference is physical layout. Parquet stores values column by column, so all the timestamps sit together and compress hard, and a query selecting three of forty columns genuinely reads three columns' worth of bytes. It also keeps min and max statistics per row group, so predicate pushdown can skip chunks entirely. Avro is row oriented, so a record is contiguous, which is what you want when you are appending events or reading whole messages off a topic. On the last pipeline I built, the Kafka payloads were Avro with a schema registry, we landed them as Avro in the raw zone, then compacted into Parquet in the curated layer. That gave us cheap writes at the edge and cheap analytical reads downstream. Both handle schema evolution, but Avro's rules around adding a field with a default are more forgiving for producers and consumers that deploy at different times, which is why it tends to win on the transport side.