Streaming Data & Apache Flink Blog

VERA-X: Introducing the First Native Vectorized Apache Flink® Engine

Written by Ben Gamble | 15 October 2025

Redefining Speed and Efficiency in Apache Flink with VERA-X

Apache Flink® is the standard for real-time stream processing, powering mission-critical applications worldwide. As the original creators of Apache Flink, Ververica pushes the boundaries of what stream processing can achieve. This includes the creation of VERA, the original enhanced engine that improves latency, stability, and efficiency for most streaming use cases.

Today, Ververica is proud to announce a new era of streaming technology: VERA-X. A native vectorized Apache Flink engine integrated into the Ververica Platform for selected prospects and customers. VERA-X delivers groundbreaking improvements in performance and efficiency, without requiring any changes to existing Flink applications. This marks a major step forward in our mission to empower our customers with the future of stream processing, accounting even for the most demanding use cases in the industry, where every millisecond matters.

VERA-X: A New Era of Stream Processing Performance

Figure 1: Key Features, Benefits, and Results of VERA-X

Modern enterprises demand ever higher throughput, lower latency, and greater cost-efficiency from their data streaming infrastructure. The biggest challenge is achieving order-of-magnitude performance gains while maintaining 100% compatibility with the Apache Flink APIs that users know and love. Incremental optimizations to Flink’s engine yield improvements, but these eventually hit a ceiling. Ververica recognizes the need for a new approach and delivers a solution that dramatically accelerates Flink workloads by fully leveraging today’s hardware, yet seamlessly integrates with the rich Flink ecosystem. 

This results in VERA-X, the next-generation engine that combines Apache Flink’s proven framework with a powerful native execution core. Just as other data platforms have embraced native vectorized processing to boost performance, Flink now gains its own high-performance runtime for streaming.

This is particularly important for extremely latency-sensitive industries, where every millisecond matters. Many use cases, including financial trading, fraud detection, ad-tech bidding systems, and IoT control loops all require sub-millisecond responsiveness. VERA-X is designed to deliver just that: ultra-low latency combined with an average of 52% lower resource usage compared to the standard Flink runtime.

This is only the beginning: VERA-X currently does not support disaggregated state storage. When introduced, users can expect costs to drop even further, making high-performance streaming more affordable than ever.

VERA-X Architecture

VERA-X introduces a re-architected Flink runtime that is fully compatible with the standard Flink APIs, but internally adds several groundbreaking components to push the limits of speed and efficiency. The diagram below illustrates how VERA-X augments Flink’s architecture:

Figure 2: How the VERA-X Augments Apache Flink Architecture

Key Innovations of VERA-X Include:

  • Native Vectorized Execution: The engine processes data in highly efficient batches, taking full advantage of modern hardware. This results in dramatically higher throughput and lower latency compared to Flink’s traditional runtime.
  • High-Performance State Management (ForStDB): State-heavy streaming jobs run smoothly with a new state store that handles everything from in-memory operations to large-scale workloads. It keeps performance consistent even under heavy demand.
  • Seamless Flink Integration (Leno Layer): VERA-X works transparently with all Flink APIs. Jobs migrate with zero code changes, maintaining full compatibility while unlocking the benefits of the new runtime.
  • Broad Operator & UDF Coverage: From day one, most commonly used operators and functions are supported natively. User-defined functions (UDFs) also benefit, so custom business logic runs faster without modification.

Performance Gains in Action

With VERA-X, we’re entering a new era of performance for Apache Flink. Whether running continuous streaming jobs or large-scale batch queries, the results are clear: higher throughput, lower latency, and significantly better resource efficiency. Benchmarks and production use cases alike show improvements of up to an order of magnitude compared to open-source Apache Flink.

Streaming Workloads

Streaming jobs are the backbone of Flink, from continuous analytics to real-time recommendations. In this domain, VERA-X shines the brightest:

  • 5–10× throughput improvement compared to open-source Flink on SQL streaming benchmarks.
  • Latency improvements measured in single-digit milliseconds, even under heavy stateful workloads.
  • 50% lower infrastructure costs due to better CPU utilization and vectorized state access.

How the benchmarks were conducted:

Streaming benchmarks are based on the Nexmark suite, an industry-standard benchmark for event stream processing. The tests compare open-source Apache Flink 1.19 against the new native engine under identical cluster configurations. Both engines execute the same SQL queries on workloads of 100 million and 200 million events, covering join-heavy, aggregation-heavy, and windowed queries. The hardware environment and cluster size are fixed to ensure a fair comparison, with results measured in execution time per query and throughput (events processed per second).

Figures 3a & 3b: Streaming Benchmark Results for 100m and 200m Records  (Lower is Better)

Batch Workloads

Although VERA-X is built for streaming first, its architectural innovations also carry over to batch queries. Thanks to the same vectorized execution and optimized state access, batch analytics tasks benefit from:

  • 2–3x faster execution times compared to Flink’s Java runtime.
  • 3.1x faster execution times compared to Apache Spark 3.4.3.
  • Lower CPU overhead translates into reduced cluster costs.
  • The ability to unify stream and batch pipelines on a single platform simplifies architecture.

How the benchmarks were conducted:

Batch benchmarks use a collection of SQL analytical queries (including TPC-H–style queries) executed through Flink SQL. The focus is on evaluating scan, filter, aggregation, and join performance over large datasets. Tests are run with identical resource configurations across engines, comparing end-to-end query completion times and cluster utilization metrics. The improvements highlight how vectorized execution in VERA-X accelerates both streaming and batch jobs consistently.

Figure 4: Batch Workload Execution Time Results VERA-X vs. Apache Spark (Lower is Better)

Ververica: Unlocking What’s Next in Real-Time Data

In closing, VERA-X is about the future of Apache Flink and stream processing at large, even for the most demanding use cases. By combining Flink’s robust, battle-tested framework with a state-of-the-art execution engine, Ververica is delivering the best of both worlds: the familiarity and reliability of Apache Flink, and the blazing performance of a modern, vectorized runtime. This innovation empowers our users to tackle use cases that were previously impractical or costly, from high-frequency analytics to real-time AI and beyond, with greater ease and confidence.

Ververica remains committed to Apache Flink’s open standards and community. VERA-X is built in harmony with Flink, a continuation of our decade-long journey with stream processing technology, and we will continue to contribute back and collaborate on pushing Flink forward. We believe this next-gen engine represents a transformative step for the streaming ecosystem, and we’re incredibly excited to bring our customers along on this journey. Join us in embracing this future, and let’s unlock the full potential of real-time data together!

More Resources

  • [Use Cases] Learn how to detect and prevent fraud, and build AI systems that learn, adapt, and act in real-time, and more.
  • [Video] Learn more about the VERA engine.