Triggered vs Continuous & Run Behaviour
Databricks · Declarative Pipelines (Lakeflow)
Theory
A pipeline's Pipeline mode is either triggered or continuous. A triggered pipeline processes all data available when the update starts, refreshes every table once and then stops its compute; you run it manually, on a schedule or from a job. A continuous pipeline keeps running, processing new data as soon as it arrives for low latency, and keeps its compute up the whole time. Separately, how an update behaves depends on how it was started. Updates you start from the editor or the pipeline page use fast-start, debugging-friendly behaviour (this is what development mode used to mean): compute is reused for quick reruns and retries are off so you see errors immediately. Updates started from a job or the API, and all continuous updates, get production behaviour: automatic retries and restarts on recoverable errors, and compute shuts down after the run.
Why we use it
Latency and cost pull in opposite directions. Pipeline mode lets you choose fresh-every-few-seconds data or cheap scheduled refreshes, and the run behaviour keeps development fast while production runs stay resilient.
When to use it
Use triggered mode for most batch and incremental ETL, scheduled hourly or daily through a job. Use continuous only when the business needs data within seconds or minutes. While building, run from the editor; in production, run from a Lakeflow Job so updates get retries.
How it works
- Decide how fresh the tables must be
- Is seconds-to-minutes latency required? Yes: Set Pipeline mode to Continuous. No: Keep Triggered.
- Develop and test with Run pipeline from the editor (fast-start, no retries)
- Ready for production? Yes: Schedule it with a Lakeflow Job (retries enabled). No: Keep iterating in the editor.
- Monitor updates in Jobs & Pipelines
Syntax
UI: pipeline Settings → Pipeline mode: Triggered | Continuous
Run: Run pipeline | Run pipeline with full table refresh | Dry run
Schedule: Schedule button in the editor (creates a job) or a Pipeline task in Lakeflow Jobs
Settings JSON: "continuous": false, "development": falseExample
// Pipeline settings (JSON view) for a production, triggered pipeline
{
"name": "orders_pipeline",
"catalog": "main",
"schema": "sales",
"serverless": true,
"continuous": false,
"development": false,
"libraries": [{ "glob": { "include": "/Workspace/Users/asha@techtown.in/orders_pipeline/transformations/**" } }]
}
-- Then schedule it hourly with a Lakeflow Job (Pipeline task)
-- or with the Schedule button in the Lakeflow Pipelines Editor.Key points
- Triggered is the default: update once, then stop compute
- Continuous keeps compute running all the time, so it costs more
- Editor/UI runs: compute reused, no retries, so errors show fast
- Job/API runs and continuous pipelines: automatic retries, compute stops after the run
- Full refresh clears streaming tables and reprocesses everything, so use it rarely
- Run now with different settings lets you override behaviour for one update