Streaming execution
Event time, watermarks, windows, state, and long-lived query lifecycles.
Streaming execution processes a sequence of Arrow batches without requiring the logical input to be finite. A stream job owns a longer lifecycle than one batch query and may hold operator state between inputs.
Event time and watermarks
An event-time column assigns records to logical time. A watermark describes the runtime's progress through that time and lets window operators decide when a result can close. The configured lateness bound determines how late records are handled.
Windows
Krishiv exposes tumbling, sliding, and session window paths. Window choice affects state size, output cadence, and late-data behavior.
- Tumbling windows do not overlap.
- Sliding windows overlap at a configured slide interval.
- Session windows group activity separated by an inactivity gap.
Stateful operators
Keyed aggregation, joins, timers, and process functions store data across input batches. State may be in memory or backed by a durable store, depending on build features and runtime configuration.
Checkpoints are a protocol
A checkpoint is useful only when sources, operators, state storage, and sinks agree on offsets and commit behavior. Krishiv exposes checkpoint primitives, but does not claim global exactly-once behavior for every connector combination.
Process lifetime matters
In-process CLI stream jobs disappear when the CLI exits. Use a long-lived session or coordinator-backed lifecycle when the job must survive a client.