# Hybrid Tables Now Up to 8x Faster, Automatically > Source: > Published: 2026-11-10 20:58:00+00:00 In June, we [previewed a new generation of Snowflake Hybrid Tables performance](https://www.snowflake.com/en/blog/hybrid-tables-performance-improvements/). **Today, those improvements are generally available and on by default.** [Hybrid Tables](https://docs.snowflake.com/en/user-guide/tables-hybrid) now **deliver up to 8x higher throughput, 3x lower p50 latency and 2x lower p99 latency.**¹ For the thousands of customers already running Hybrid Tables, their existing workloads will get faster and cheaper automatically, with room to handle more volume on the same compute. ## Transactional data without a separate database Pipelines, AI agents and internal apps all depend on fast transactional reads and writes, whether to track job state, store agent context or record what users do in enterprise apps built with [Streamlit](https://docs.snowflake.com/en/developer-guide/streamlit/about-streamlit) or [Snowflake App Runtime](https://docs.snowflake.com/en/developer-guide/snowflake-app-runtime/about-snowflake-app-runtime). Too often, that has meant standing up a separate operational database, a pipeline to keep it in sync with analytics and a second set of governance controls. Hybrid Tables bring fast, ACID-compliant reads and writes into the Snowflake database instead. Transactional and analytical data live together under the same governance framework, joinable in a single query. And with this release, existing apps and workflows get better price performance with nothing to rearchitect. ## What's generally available today with Hybrid Tables ### Up to 8x throughput improvement Single-statement query execution for repetitive point operations now delivers up to 8x higher throughput.¹ That means, for certain workloads, what used to take eight XS warehouses now takes one, which can cut compute costs by more than 80%.2 ### Lower latency at the median and the tail **With the new Hybrid Tables engine,** benchmark results show **p50 latency is 3x lower and p99 latency is 2x lower.**3 A faster median means quicker responses on every request. A tighter tail means fewer slow requests at peak load, which is what users and downstream systems feel most. ### Faster, more cost-effective bulk loads Bulk inserts are optimized at the storage engine level, with benchmark results demonstrating **up to** **10x faster batch writes at 10x lower cost,** even on tables with existing data.3 Backfills and large syncs that once required careful orchestration now work with a simple INSERT. ### Simpler pricing Hybrid Tables follow the same compute-and-storage model Snowflake customers already know, but we’ve removed one of the dimensions so you’re billed on fewer things. **Now, customers have seen costs reduced by 15% on average and 40% or more for high-throughput workloads.**4 | **Hybrid Tables evolution** | **Results** | |---|---| | Higher throughput | Up to 8x more throughput, applied automatically | | Faster batch performance | Up to 10x faster, 10x cheaper bulk loads with no workflow changes | | Cost reduction | 40% or more cost savings | ## Where customers use Hybrid Tables Whether you're building an internal app, running a pipeline or orchestrating AI agents, Hybrid Tables workloads typically follow the three patterns discussed below. Many apps combine more than one. ### Workflow state Pipelines, apps and agents all need a place to track state, including job status, checkpoints, sessions and intermediate results. Teams have often run an external database just for this. Hybrid Tables keep workflow state transactionally consistent, in the same Snowflake database as the analytical data those workflows produce, so teams can query both together. For example, [MarketWise](https://www.snowflake.com/en/customers/all-customers/case-study/marketwise/) replaced MySQL and DynamoDB to track the workflow state of its data engineering pipelines in real time, **cutting infrastructure costs by 35% and eliminating unnecessary data movement.** [Elementum](https://medium.com/snowflake/how-elementum-builds-the-next-generation-of-ai-driven-workflows-with-unistores-hybrid-tables-1f2cf4f9d67c) orchestrates AI agents, business rules and human steps in multistep processes on Hybrid Tables, with enforced primary key and foreign key constraints. In preview, we’ve seen customers’ workflow state workloads run up to 5x faster at p50 and 2.8x faster at p99. One customer saw p90 latency drop from 21 milliseconds to 5 milliseconds while running 55% more jobs.⁵ ### Application data serving When an app or agent needs to answer a question instantly, such as whether to approve a return, which product to recommend or what state a machine is in, Hybrid Tables turn Snowflake into the serving tier. Data-serving workloads see the biggest gains from this release. In preview, they saw up to 6.7x lower p50 latency and 6.3x lower p99 latency.⁵ Now, teams serve far more concurrent lookups on far less compute. For example, [Grailed](https://medium.com/snowflake/grailed-revolutionizes-style-discovery-with-snowflake-unistore-e7c5061e5795) serves subsecond, personalized style recommendations to millions of users from precomputed ML results in Hybrid Tables, with no separate serving database. The same pattern fits demanding operational environments such as manufacturing, logistics and supply chain. Operators need millisecond lookups on the current state of a job, a machine or a shipment, while planners run analytics on the same data. Hybrid Tables serve the lookups, and standard tables handle large-scale joins, both on one platform without per-table tuning. The p99 improvements matter most here, where a slow tail can stall a line or a schedule. ### Reliable, high-concurrency writes Some workloads write constantly. Apps and services capture every user action, API call, model invocation or transaction as a record that has to land reliably. Standard tables are built for analytics and can hit lock contention under heavy concurrent writes. Hybrid Tables use row-level locking designed for exactly this pattern. In preview, event log and record capture workloads ran up to 3.4x faster at p50 and 3.7x faster at p99. One customer's operational-log workload saw p90 insert latency drop by 66% while handling more than 3x the job volume.⁵ [Writer](https://medium.com/snowflake/how-writer-keeps-its-ai-infrastructure-running-with-snowflake-hybrid-tables-80f55dd652fa), an enterprise AI platform, writes roughly 1.3 million LLM trace records a day to Snowflake to power model observability and QA. On a standard table, p50 insert latency soared, and up to 45% of writes failed. **Writer switched the destination to a Hybrid Table, essentially a table swap, and p50 insert latency dropped roughly 15x. Failures fell to near zero.** Hybrid Tables absorb the writes, while a standard table clustered by timestamp powers analytics, all in the same database. ### Appriss Retail stops fraud using fresh data [Appriss Retail helps stop fraud](https://apprissretail.com/) related to returns and claims for more than 60 of the top 100 U.S. retailers at 99.99% accuracy. Every return and claim is scored in real time. Each shopper's purchases, returns and claims are matched into a single profile. Risk signals are computed from that history at the moment of return to approve, warn or deny. Before Hybrid Tables, that scoring depended on a separate SQL Server database fed by a custom ETL pipeline. Engineering time went into maintaining the pipeline instead of improving detection. Moving data between systems meant fraud decisions ran on delayed data. And loading data slowed down the same database serving live returns and claims. Now, retailer returns and claims apps run lookups directly against Hybrid Tables, while data science teams query standard tables on the same platform. Lookups complete quickly, and analytics scan terabytes of data, all in Snowflake. The custom pipeline is gone, and decisions run on data the moment it's generated. "Every return and claim our biggest retailer processes is now checked against Hybrid Tables. That’s more than a million lookups a day, with no data exported out of Snowflake to do it." ##### Kenny Vu VP of Engineering, Appriss Retail ## How to get started with Hybrid Tables If you already use Hybrid Tables, there's nothing you need to do. The improvements are on by default. New to Hybrid Tables? Creating one takes a single statement, and you can join it with your existing tables right away. **Create a hybrid table to track pipeline job state** ``` CREATE HYBRID TABLE job_state ( job_id INT PRIMARY KEY, status VARCHAR, updated_at TIMESTAMP_NTZ, INDEX idx_status (status) ); ``` **Fast point writes and lookups** ``` INSERT INTO job_state VALUES (1001, 'RUNNING', CURRENT_TIMESTAMP()); SELECT status FROM job_state WHERE job_id = 1001; ``` **Join transactional state with analytical data in one query** ``` SELECT j.job_id, j.status, r.rows_processed FROM job_state j JOIN job_run_metrics r ON j.job_id = r.job_id WHERE j.status = 'FAILED'; ``` From there, walk through a full example with the [quickstart guide](https://docs.snowflake.com/en/user-guide/tutorials/getting-started-with-hybrid-tables-tutorial), and explore the documentation for [Hybrid Tables best practices.](https://docs.snowflake.com/en/user-guide/tables-hybrid-best-practices) ¹ Data based on benchmarks run using the Yahoo Cloud Serving Benchmark with a 100% read workload on a Gen2 XS warehouse. 2 Derived from 8x throughput on a single XS warehouse versus eight XS warehouses at equivalent load. 3 Data based on internal benchmarks for loading 50 GB into Hybrid Tables using a 2XL warehouse. 4 Data based on Snowflake estimates of average cost savings across customers, measured on real-world production credit consumption in February 2026. ⁵ Data based on Snowflake observations of p50, p90 and p99 query latency for customer production workloads before and after enabling the new Hybrid Tables engine during public preview.