Faster, Smarter, Autonomous: Snowflake SQL Performance Improvements Year in Review

For years, we’ve steered Snowflake under a simple principle: Performance should improve on its own. As workloads grow and evolve, engineers shouldn’t need to reconfigure storage, manage compute or rewrite queries just to keep up with increased demand. That’s what legacy platforms may require. A modern platform should automatically observe, learn and optimize.

Now more than ever, performance has to be autonomous. Our vision is that it should be easy to run optimally and impossible to run inefficiently on Snowflake. A modern workflow can generate thousands of unique query shapes in minutes. Filters shift, joins change and access patterns drift. Add AI agents into the mix, and predictability disappears entirely. Optimization becomes a moving target that’s impossible to chase with manual tuning.

In 2025, we pushed autonomous intelligence deeper into the core of Snowflake. These were not isolated feature improvements. They represent additional pillars for a self-optimizing platform designed to thrive under today’s BI dashboards, data engineering pipelines, application backends, and increasingly agentic and autonomous systems. This gives customers more headroom to innovate, derive deeper insights from petabytes of data they manage in Snowflake, and efficiently scale workloads from experimentation to production.

Some of the most striking improvements came from intelligent optimizations we applied to at-scale workloads. Customers didn’t have to rearchitect or rewrite code. The platform simply became more efficient.

Below, we detail the sweeping improvements we delivered last year to further strengthen Snowflake as a leading foundation for your data and AI success.

Snowflake Optima and the autonomous data layer

Snowflake Optima (generally available, as of August 2025) represents a major step forward for our autonomous vision. It is a continuous optimization engine that monitors workload patterns in real time to improve performance. You can think of Optima as a “virtual database administrator” who’s always on the lookout for optimization opportunities across your Snowflake deployment.

Because the best query plan is useless if the underlying data is fragmented, our first goal for Optima is to help you optimize your data layer, starting with Optima Indexing and Optima Metadata. Instead of asking you to figure out how to tune your tables, Snowflake Optima identifies repeated access patterns and applies hidden indexes and custom metadata in the background, helping your data layer evolve in lockstep with your workloads.

The results of this "virtual database administrator" are stark. Snowflake Optima Indexing has automatically accelerated workloads by up to 54x for hundreds of customers using Gen2 warehouses, seamlessly pruning billions of micro-partitions per day for point-lookup queries. In one recent production example, an automotive customer saw an immediate acceleration of frequent point-lookup queries, with latencies dropping from 17 seconds to just over 1 second. Optima Indexing automatically generated a hidden search index that increased micropartition pruning from 30% to 96%, eliminating the vast majority of I/O without a single manual ALTER TABLE command. Similarly, a Fortune 500 healthcare enterprise saw significant gains in its threat detection workload from Optima Metadata, which automatically creates custom metadata to accelerate queries with common predicate patterns. They got faster threat detection, with security rules executing 10% faster with 74% less data scanned, a 4x I/O reduction from over 20 TB/day down to about 5 TB/day.

We believe this zero-maintenance philosophy shouldn't be restricted to proprietary formats. In 2025, we erased the operational distinction between "internal" and "external" storage by bringing our native background maintenance engines directly to Apache Iceberg™. All of our pruning features, including Search Optimization and Snowflake Optima, work seamlessly for both native Snowflake and Iceberg tables. And we launched the new Table Optimization Service groups compaction and Automatic Clustering into a single background process, treating your Iceberg tables with the exact same rigor as native Snowflake storage. Coupled with other new optimizations detailed below, these managed capabilities allowed us to reduce the execution time on our internal TPC-DS 10TB benchmark on Iceberg format by 35%.

Compute: Runtime execution engine that adapts

Efficiency isn’t just about raw compute and infrastructure: It’s about how the engine consumes it. We are on a journey to dismantle the static compute model in two phases: radically optimizing the runtime, then revolutionizing compute consumption.

Gen2 warehouses

Gen2 represents a fundamental evolution of the Snowflake query runtime, delivering massive gains “out of the box” when comparing February 2025 to February 2026:

• Up to 5.5x faster DML operations: This is the most critical improvement for data engineering. We optimized execution paths of DELETE, UPDATE, and MERGE to massively reduce write amplification. This shrinks execution runtimes for heavy change data capture (CDC) pipelines and high-churn tables.
• Up to 1.8x faster core analytics: The new runtime accelerates complex aggregations and massive table scans, allowing dashboards and interactive queries to run more than twice as fast.

As you can see, Gen2 significantly increases how efficiently work is executed. Next, we will be doubling down on the evolution of our compute model and will continue with innovations like the upcoming Adaptive Compute that takes away the undifferentiated heavy lifting of making infrastructure choices.

Query engine: Intelligent and adaptive planning and execution

Modern analytics and data engineering patterns introduce skew, correlation and unpredictability that challenge even well-optimized systems. Every database engineer knows that compile-time estimates alone are often insufficient to help the engine deliver optimal results.

Last year, we released more than 50 workload-intelligent performance improvements under the hood to make the core query engine run even more efficiently, even in the face of changing or incomplete statistics. Notably, we introduced:

• Smarter metadata: For large, skewed or highly correlated workloads — common in migrations, log analytics and security use cases — traditional statistics might misestimate the ultimate selectivity of a filter. Our optimizer now gathers additional targeted optimization data about predicates to optimize join orders and access paths. The result is fewer join explosions and long‑tail regressions.
• Smarter planning: We upgraded the optimizer’s join ordering, selectivity modeling, cardinality estimation and search-space pruning to enable more stable performance on complex analytics. For example, by transparently removing unnecessary group-by keys, we improved execution times for hundreds of thousands of aggregation queries powering critical dashboards.
• Enhanced pruning coverage and efficiency: We significantly expanded the scope of runtime pruning, enabling the engine to more effectively skip unnecessary data during query execution (including scenarios involving scalar subqueries and joins with expressions). Across our platform, this has resulted in significant gains — on the order of hundreds of billions of micro-partitions per day across millions of queries. In production workloads, we have observed substantial performance gains for certain query patterns. Query duration improved by between 10% and 20% on average, with the best observed improvements reaching over 300x improvement compared to prior behavior. Additionally, we’ve extended pruning optimizations to Iceberg tables, Top-K queries, LIKE predicates, and Geometry data types.
• Adaptive execution: We enhanced the execution engine to dynamically rebalance work at runtime, eliminating bottlenecks caused by data skew. When an imbalance is detected, the system automatically redistributes data and adjusts execution strategies in real time. This optimization has accelerated more than 3.4 million queries without user intervention, with the largest observed speedups reaching 180x.
• Automated query simplification: We automatically simplify queries prior to execution by removing redundant logic, deduplicating subplans and factoring complex expressions. For example, the engine automatically builds bloom filters when a CTE is joined with other tables. This is especially powerful for CTE-heavy analytic workloads that previously had to materialize large intermediate result sets — with execution times for some queries improving by up to 400x.
• Blazing fast analytics: Snowflake Interactive offers a modern replacement for legacy real-time databases and external caches. This innovation achieves subsecond query latency and consistent tail latencies, even under heavy data ingestion, by pre-warming data on local SSDs for faster access, and by eliminating compilation and compiler overhead. Benchmarking demonstrates that Interactive Analytics delivers a 9x increase in throughput compared to Standard Warehouses. For more details, refer to the engineering blog post here.
• Faster AISQL queries: We made AISQL both faster and more accurate at scale. Results from six public benchmarks show that AI‑aware query optimization generated 2x–8x faster execution plans; adaptive model cascading achieved 2x-6x quicker inference, while preserving roughly 90%–95% of large‑model accuracy; and rewriting semantic joins as multilabel classification cut execution time and cost by 15x-70x.

The year ahead

In 2026, the stakes have changed. We are entering an era where machines and AI agents will access data at a scale, frequency and complexity that far outpaces human oversight. In this environment, relying on a human to manually optimize infrastructure is a critical liability.

Our vision and engineering focus remains unchanged and bears repeating: It should be easy to run optimally and impossible to run inefficiently on Snowflake. The architectural answer for the data stack of the future is autonomy. And the future is already here.

¹ Based on high-concurrency/low-latency Interactive Analytics Benchmark. (Jan. 2026)

² Based on testing against six publicly available benchmarks, described in further detail here. (Nov. 2025)

³ Based on observed query performance in production environments. (Oct. 2025)