Architecture history — Lambda → Kappa → lakehouse → streamhouse

1. "What problem did Lambda architecture solve, and what did it cost?"

It gave both low-latency results and complete, reprocessable results. No single engine could do both. A batch layer recomputes the correct answer. A speed layer serves the recent window. A serving layer merges them. The cost is that you maintain the same business logic twice in two code bases. The two copies drift, and people must reconcile them. This is expensive and a correctness risk.

2. "How is Kappa different from Lambda, and what is its weakness?"

Kappa deletes the batch layer. It has one streaming pipeline, with the log (Kafka) as the system of record. You replay the log to reprocess history. There is one code path, so nothing drifts. Its weakness has three parts. Long-term, queryable history in Kafka is expensive. Ad-hoc OLAP over the log is awkward. Replaying months of events to fix a bug is slow.

3. "A lakehouse already unifies storage. Why would anyone add a streamhouse?"

A lakehouse commits at batch/micro-batch latency. Snapshots land every few minutes at best. So it cannot answer sub-minute streaming questions. Frequent commits also cause a small-files problem. A streamhouse puts stream-native storage (Paimon/Fluss) on the lake. Then one copy serves both a streaming/changelog read and a batch/OLAP read. This adds the sub-minute tier that the lakehouse lacks.

4. "How do you classify an unknown platform without trusting the vendor label?"

Look at the storage and pipeline shape.

• Two parallel pipelines merged at serving → Lambda.
• One streaming pipeline with an append-only log as the system of record → Kappa.
• One transactional lake table format at batch latency → lakehouse.
• A stream-native store that serves both streaming and batch reads from one copy → streamhouse.

5. "When is a streamhouse the wrong choice?"

There are three cases.

• Hourly or daily freshness is enough. A batch lakehouse is simpler and cheaper.
• Every consumer only replays a log, and nobody runs OLAP. Plain Kappa is less to operate.
• You genuinely need two different computations, for example an exact regulatory recompute plus a separate real-time estimate. That is a legitimate Lambda. Do not unify it away.

6. "Who coined 'streamhouse' and what does the term signal?"

Ververica coined it (~2024). It is now a cross-vendor working group (Aiven, Confluent, Redpanda, StreamNative, Ververica). The term signals that streaming storage and the lakehouse are converging. Stream-native storage sits on the lake, usually with Flink. The stream gets a cheap, queryable history. The lakehouse gets a sub-minute tier.