Scatter query performance degradation in VTGate

Query Routing and Resource Exhaustion

In a Vitess deployment, VTGate is responsible for translating logical SQL queries into shard-specific requests. When a query is executed without a sharding key in the WHERE clause, the system cannot determine a specific destination via the Vindex. This triggers a scatter query, where VTGate broadcasts the request to every shard within the keyspace.

As the number of physical shards increases, the overhead of aggregating results from multiple Tablet servers can lead to significant latency and resource exhaustion on the proxy layer.

Technical Constraints

• The keyspace definition relies on Vindexes to map logical databases to physical shards.
• Cross-shard aggregation occurs at the VTGate level after the Tablet servers return local results.

What are the specific configuration limits within VTGate to prevent scatter queries from exhausting system resources? How does Vitess handle the timeout behavior when a subset of shards in a scatter operation fails to respond?