Snowflake says cross-query feedback cut a 19-join query to 30 seconds

Snowflake says its Optima Planning optimizer used cross-query feedback to improve a 19-join analytics query from nearly two hours to 30 seconds in a production workload. The company identifies the customer only as a large U.S.-based healthcare technology company and says the improvement required no manual configuration. The change extends the system beyond recognizing a repeated query: it can draw on experience from a matching plan fragment or individual operator embedded in a different query.

After compilation, Optima Planning calculates hashes for fragments and individual operators using the optimizer's output for those plan components, rather than the query's SQL text. If a later, otherwise new query contains a piece with the same hash, feedback from the earlier query's execution can inform the plan for that shared piece on the later query's first run. That matters because Snowflake describes its optimizer as cost-based: it estimates operator row counts to select an expected low-cost plan, and mistaken estimates can lead to poor join choices and billions of unnecessary intermediate rows.

The matching process is guarded: a shared fragment or operator is treated as a possible match, and the system assesses whether applying its feedback in a new context would overgeneralize. Snowflake says it withholds the feedback when that transfer appears too aggressive, rather than assuming that a useful estimate in one query remains valid everywhere. The cross-query method is designed to complement, not replace, per-query feedback; it can help a new query on its first execution, after which its own repeats are handled by the existing mechanism.

Snowflake gives examples including adding a previously unused join to an established view, reusing a familiar customers-and-orders join in a new report, and queries whose table order differs between tools. In the last case, the exact queries may be treated separately, but matching at the plan level can still identify a shared join and reuse its feedback. For teams building analytics queries around established views or recurring joins, the reported consequence is that execution evidence may improve planning before that exact query has run before.