AlloyDB’s AI tuner can execute your SQL before advising you—treat that as a production boundary

Google Cloud’s new ai.tune_sql preview is more than a prompt wrapped around PostgreSQL documentation. The AlloyDB function can inspect a query’s schema and plan, use runtime statistics, and ask a registered Gemini endpoint for recommendations covering rewrites, indexes, statistics, configuration flags, ANALYZE and VACUUM. It requires PostgreSQL 18 or later and version 1.6.2 or later of the google_ml_integration extension. (Google Cloud documentation; release notes)

The default call executes the query

The operational detail that matters most is the default: explain_analyze is TRUE. AlloyDB therefore runs EXPLAIN ANALYZE, executing the statement to collect an actual plan before the LLM produces advice. Operators can set the parameter to FALSE to obtain a non-executing EXPLAIN, but Google says the runtime path produces more useful recommendations. When execution is enabled, the preview only accepts a single read-only SELECT; writes, SELECT INTO and SELECT FOR UPDATE return an error. (Google Cloud documentation)

That makes the function a workload action, not merely an offline reviewer. A slow analytical query can still consume database resources while being diagnosed. The safer production pattern is to start with explain_analyze => FALSE, reproduce representative literals and data in a controlled environment, then opt into runtime collection under normal query-governance controls. This is an operational inference from the documented execution behavior, not a Google-prescribed rollout procedure.

Advice is not an approved migration

Google’s own output includes a warning that LLM-assisted advice requires independent verification. The example recommendation rewrites a correlated NOT IN, proposes a composite index and increases work_mem; the documentation separately warns that suggested indexes consider the submitted query rather than the system’s overall workload. The function can take two to four minutes to return recommendations. (Google Cloud documentation)

Teams should therefore preserve three gates: semantic equivalence tests for rewrites, workload-level evaluation for indexes and configuration changes, and a normal reviewed deployment path. The function generates candidates; it does not validate application behavior, regression risk or fleet-wide resource effects.

Prepared and parameterized statements add another boundary. ai.tune_sql omits rewrite recommendations for them unless placeholders are replaced with representative literals, though it can still return other tuning advice. That substitution should be recorded so reviewers know which values shaped the plan and recommendation. (Google Cloud documentation)

The useful mental model is an AI-assisted database advisor with optional runtime evidence—not an autonomous optimizer. Its value is narrowing the investigation. Production ownership still begins where the generated recommendation ends.