Snowflake Extends Bulk INSERT to Populated Hybrid Tables

Snowflake has extended optimized Bulk INSERT to Hybrid Tables that already contain data, giving large INSERT operations an object-storage path rather than sending every row directly through the primary row store. The company says the capability has been generally available since July 2026 and is selected automatically when its query engine identifies a sufficiently large write.

For a qualifying operation, Snowflake stages the write in object storage and propagates the data back to the row store. Hybrid Tables use the row store as primary storage and asynchronously copy data to object storage for large scans and workload isolation. Snowflake says the expanded architecture lets the system choose a write medium based on its estimate of the INSERT’s size, without requiring a separate user command.

The bulk path continues to enforce primary-key, UNIQUE and FOREIGN KEY constraints. Snowflake says it uses batched row-level locking during the load, so users do not have to disable those constraints. Snowflake’s Hybrid Table documentation identifies the product as generally available for accounts in AWS and Microsoft Azure commercial regions.

Snowflake reported faster loading in its own tests, but the supplied research contains no independent benchmark. In a warmed-warehouse test, the company loaded 100 GB, or 10 million rows, from a standard table using a Large warehouse. It reported that an INSERT into a nonempty Hybrid Table took 3 hours and 9 minutes through the row-store path and 19 minutes through Bulk INSERT. Snowflake described its largest improvement across tested cases as up to 10 times faster.

Snowflake also tested the effect on a concurrent transactional workload. In its internal 5 GB YCSB setup, 50 clients ran under a fixed ceiling of 2,200 queries per second on an X-Small warehouse while another client continuously inserted data using a Large warehouse. Snowflake measured 2,255 QPS at baseline, 814 QPS during a large row-store INSERT and 2,231 QPS during Bulk INSERT. It reported 90th-percentile latency at 28.5 times baseline in the row-store case and 1.2 times baseline with Bulk INSERT. Those figures reflect Snowflake’s disclosed configuration, not an independently reproduced production workload.

Snowflake’s own materials conflict over which INSERT forms qualify. The engineering post excludes INSERT ALL, while its current Hybrid Table loading guidance says all INSERT variants, including INSERT ALL, may use optimized bulk loading on populated tables when load-size and row-count conditions are met. The supplied sources do not reconcile the two statements.

The new path builds on optimized COPY and INSERT loading for empty Hybrid Tables, which became generally available on January 15, 2025. Snowflake also said its September 2026 internal telemetry showed that a few hundred bulk-load queries among tens of millions of daily INSERT queries accounted for more than 80% of data loaded into Hybrid Tables.

The database feature is separate from Snowflake’s recent Cortex AI model rollout, which concerned model access rather than Hybrid Table loading.