Reverse ETL is the practice of syncing modeled data out of a cloud data warehouse and into operational SaaS tools. The shape inverts traditional ETL: instead of pulling data FROM source systems INTO the warehouse, the warehouse pushes modeled tables OUT to the places that act on them — Klaviyo, Meta Custom Audiences, Google Ads, Iterable, Braze, Attentive, Zendesk, the brand’s CRM. The substrate is first-party data that has already been cleaned and modeled inside Snowflake, BigQuery, Databricks, or Redshift.
The pattern earns its place because the pre-reverse-ETL workflow was duplicated audience definitions. A dbt model produced the high-LTV cohort for the analytics dashboard; a marketer rebuilt the “same” cohort by hand inside Klaviyo or Meta; the two drifted within a quarter. Reverse ETL makes the warehouse model the single source of truth and pushes downstream from there. For teams that already have the warehouse and the modeling discipline in place, the lag between “the analyst built a new segment” and “the marketer can target it” collapses from weeks of CSV exports and list uploads to hours.
Mechanically, syncs run on a schedule (cadences range from a few minutes on premium tiers to once daily) and use idempotent upserts keyed on a stable identifier — customer ID, email, hashed PII — so re-running a sync doesn’t duplicate rows in the destination. The tool maps a warehouse column to the destination’s audience, list, or contact object and handles the API quirks each destination demands. The category leaders are Hightouch, Census, Polytomic, and RudderStack’s reverse-ETL surface. Snowflake Data Sharing and BigQuery Data Sharing get mentioned in the same breath but cover a narrower slice — they share read access between warehouse accounts; they don’t push into SaaS destinations.
Reverse ETL is not a CDP. A CDP also handles identity resolution and event streaming; most reverse-ETL tools do neither. The composable-CDP pattern pairs reverse ETL with the warehouse’s own identity model to approximate CDP behavior, but that’s an assembled stack, not a drop-in product. Reverse ETL is also not real-time. Batch sync cadences fit attribute-rich audiences well — LTV bands (see LTV), predicted-churn scores, cohort-defined lifecycle stages — and fit event triggers poorly. A cart abandoned ninety seconds ago belongs in an event-streaming tool, not a reverse-ETL sync that runs every fifteen minutes.
The precondition is the part operators underestimate. Reverse ETL needs a working warehouse and modeled data, typically dbt. Without the models, the only thing syncing out is raw source data, and destinations break on schema and quality the first time the sync runs. The tool inherits the warehouse’s modeling discipline; if the modeling isn’t there, the destinations surface the gap immediately.