Definition
Streaming systems (Kafka, Pub/Sub, Kinesis) move events as they happen. Consumers update dashboards, triggers, or online features with low latency.
Exactly-once processing, late events, and windowing are core streaming challenges.
In simple terms
Batch is checking your mailbox once a day. Streaming is standing at the door as letters arrive.
Where you see it
- Fraud alerts within seconds of a transaction.
- Live recommendation features from recent clicks.
How it works
1.Produce events
Apps write to a log or topic.
2.Consume
Workers process windows of events.
3.Update state
Sinks write to stores, features, or alerts.
Why it matters
- Many ML products need fresh signals — streaming keeps features from going stale.
Often confused
Streaming always replaces batch.
Batch remains simpler and cheaper for many analytics and training jobs; hybrid is common.