Stateful vs stateless processing

Stateless processing looks only at the current event. Map, filter, and simple projections are usually stateless. Restart is easy: no memory of the past to restore.

Stateful processing remembers something across events: running counts, last seen value, session windows, joins between streams. That memory is state, often keyed by user_id / order_id.

Stateless:  event -> filter(amount > 100) -> sink
            (no memory)

Stateful:   event -> update count[user] += 1 -> emit if count > 5
            needs keyed state store + checkpointing

Tiny example

• Stateless: tag each click with is_mobile
• Stateful: count clicks per user in the last 5 minutes (needs window + state)

Why it matters

State must be checkpointed for recovery. Skewed keys blow up memory. Exactly-once sinks care about state + offsets together.

Interview tip: If the answer depends on "what happened before," it is stateful. Then mention keys, checkpoints, and TTL.