Depends on your data and use case. Kafka provides way better throughput and latency compared to Casandra. Casandra is a proper NOSQL store where as Kafka is just a pub-sub based Message Queue. If you want the data to persist and run queries while it is in transition use Casandra. Otherwise, Kafka is enough.

Apache Kafka is much more than a simple pub/sub message queue!! You can use Kafka to realize a Kappa architecture instead of a Lambda architecture. You can even use Kafka as your only data store, even so I would not recommend using it that way (no real query language except KSQL, might get into partitioning problems for differing compute node architectures/disk sizes).

Kafka is used in Big Data solutions for data ingestion (Kafka Connectors) and ELT (not ETL!!) because of its simple interface (pub/sub), its functionality for data aggregation/ transformation/streaming, its capability to store huge amounts of data and its very high throughput without backpressing.

Consequently, it is often used as the source for Lambda architectures, so e.g. to feed HDFS or Cassandra realizing master data stores to permanently persist your data, and feeding e.g. Spark for the batch and stream processing parts.

The results from batch and streaming processing are then stored in fast databases like Cassandra, to realize fast access and query of analysis results.

So: it heavily depends on your data and your use cases (what do you want to do with your exchange rates?), i.e. you might use Kafka or Cassandra standalone, or a combination of both (in the later case it is advised to not run Kafka and other tools like Spark on the same machine because the system caches on which Kafka relies on might get compromised, reducing the performance/throughput of Kafka).