[{"data":1,"prerenderedAt":1564},["ShallowReactive",2],{"active-banner":3,"navbar-featured-partner-blog":24,"navbar-pricing-featured":306,"blog-\u002Fblog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis":1086,"blog-authors-\u002Fblog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis":1531,"related-\u002Fblog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis":1545},{"id":4,"title":5,"date":6,"dismissible":7,"extension":8,"link":9,"link2":10,"linkText":11,"linkText2":12,"meta":13,"stem":21,"variant":22,"__hash__":23},"banners\u002Fbanners\u002Flakestream-ufk-launch.md","StreamNative Introduces Lakestream Architecture and Launches Native Kafka Service","2026-04-07",true,"md","\u002Fblog\u002Ffrom-streams-to-lakestreams","https:\u002F\u002Fconsole.streamnative.cloud\u002Fsignup?from=banner_lakestream-launch","Read Announcement","Sign Up Now",{"body":14},{"type":15,"value":16,"toc":17},"minimark",[],{"title":18,"searchDepth":19,"depth":19,"links":20},"",2,[],"banners\u002Flakestream-ufk-launch","default","zRueBGutATZB0ZnFFHwaEV7F0Di4tnZUHhgOiI4cu6k",{"id":25,"title":26,"authors":27,"body":29,"category":289,"createdAt":290,"date":291,"description":292,"extension":8,"featured":7,"image":293,"isDraft":294,"link":290,"meta":295,"navigation":7,"order":296,"path":297,"readingTime":298,"relatedResources":290,"seo":299,"stem":300,"tags":301,"__hash__":305},"blogs\u002Fblog\u002Fstreamnative-recognized-in-the-forrester-wave-streaming-data-platforms-2025.md","StreamNative Recognized as a Contender in The Forrester Wave™: Streaming Data Platforms, Q4 2025",[28],"David Kjerrumgaard",{"type":15,"value":30,"toc":276},[31,39,47,51,67,73,78,81,87,102,109,115,118,124,127,134,140,143,146,157,163,169,172,175,178,184,191,194,197,204,207,210,224,229,233,237,241,245,249,251,268,270],[32,33,35],"h3",{"id":34},"receives-highest-possible-scores-in-both-the-messaging-and-resource-optimization-criteria",[36,37,38],"em",{},"Receives Highest Possible Scores in BOTH the Messaging and Resource Optimization Criteria",[40,41,43],"h2",{"id":42},"introduction",[44,45,46],"strong",{},"Introduction",[48,49,50],"p",{},"Real-time data has become the backbone of modern innovation. As artificial intelligence (AI) and digital services demand instantaneous insights, organizations are realizing that streaming data is no longer optional – it's essential for delivering timely, context-rich experiences. StreamNative's data streaming platform is built precisely for this reality, ensuring data is immediate, reliable, and ready to power critical applications.",[48,52,53,54,63,64],{},"Today, we're excited to announce that Forrester Research has named StreamNative as a Contender in its evaluation, ",[55,56,58],"a",{"href":57},"\u002Freports\u002Frecognized-in-the-forrester-wave-tm-streaming-data-platforms-q4-2025",[36,59,60],{},[44,61,62],{},"The Forrester Wave™: Streaming Data Platforms, Q4 2025",". This report evaluated 15 top streaming data platform providers, and we're proud to share that ",[44,65,66],{},"StreamNative received the highest scores possible—5 out of 5—in both the Messaging and Resource Optimization criteria.",[48,68,69,70],{},"***Forrester's Take: ***",[36,71,72],{},"\"StreamNative is a good fit for enterprises that want an Apache Pulsar implementation that is also compatible with Kafka APIs.\"",[48,74,75],{},[36,76,77],{},"— The Forrester Wave™: Streaming Data Platforms, Q4 2025",[48,79,80],{},"Being recognized in the Forrester Wave is a proud milestone, and for us, it highlights how far StreamNative has come in enabling enterprises to unlock the power of real-time data. In the sections below, we'll dive into what we believe sets StreamNative apart—from our modern architecture and cloud-native design to our open-source foundation and real-time use cases—and how we see these strengths aligning with Forrester's findings.",[40,82,84],{"id":83},"trusted-by-industry-leaders",[44,85,86],{},"Trusted by Industry Leaders",[48,88,89,90,93,94,97,98,101],{},"Companies across industries are already leveraging StreamNative to drive real-time outcomes. Global enterprises like ",[44,91,92],{},"Cisco"," rely on StreamNative to handle massive IoT telemetry, supporting 245 million+ connected devices. Martech leaders such as ",[44,95,96],{},"Iterable"," process billions of events per day with StreamNative for hyper-personalized customer engagement. And in financial services, ",[44,99,100],{},"FICO"," trusts StreamNative to power its real-time fraud detection and analytics pipelines with a secure, scalable streaming backbone.",[48,103,104,105,108],{},"The Forrester report notes that, “",[36,106,107],{},"Customers appreciate the lower infrastructure costs that result from StreamNative’s cost-efficient, Kafka-compatible architecture. Customers note excellent support responsiveness…","”",[40,110,112],{"id":111},"modern-cloud-native-architecture-built-for-scale",[44,113,114],{},"Modern, Cloud-Native Architecture Built for Scale",[48,116,117],{},"From day one, StreamNative was designed with a modern architecture to meet the demanding scale and flexibility requirements of real-time data. Unlike legacy streaming systems that often rely on tightly coupled storage and compute, StreamNative's platform takes a cloud-native approach: it decouples these layers to enable elastic scalability and efficient resource utilization across any environment. The core is powered by Apache Pulsar—a distributed messaging and streaming engine—enhanced with multi-protocol support (including native Apache Kafka API compatibility) to unify diverse data streams under one roof. This means organizations can consolidate siloed messaging systems and handle both high-volume event streams and traditional message queues on a single platform, without sacrificing performance or reliability.",[48,119,120,121,108],{},"Forrester's evaluation described that “",[36,122,123],{},"StreamNative aims to provide a high-performance, multi-protocol streaming data platform: It uses Apache Pulsar with Kafka API compatibility to deliver cost-efficient, real-time applications for enterprises. It appeals to organizations that want a flexible, low-cost streaming solution, due to its focus on scalability and resource optimization, while its investments in Pulsar’s open-source ecosystem and performance optimization make it the primary platform for enterprises wishing to implement Pulsar.",[48,125,126],{},"Our cloud-first, leaderless architecture (with no single broker bottlenecks) and tiered storage model were built to maximize throughput and cost-efficiency for real-time workloads. By separating compute from storage and leveraging distributed object storage, StreamNative can retain huge volumes of event data indefinitely while keeping compute costs in check—effectively providing a flexible, low-cost streaming solution.",[48,128,129,130,133],{},"This modern design not only delivers high performance, but also ensures fault tolerance and geo-distribution out of the box, so enterprises can trust their streaming data is always available and durable. As Forrester’s evaluation noted, StreamNative ",[36,131,132],{},"\"excels at messaging and resource optimization\" and “Its platform supports use cases like real-time analytics and event-driven architectures with robust scalability.","” Our architecture provides the strong foundation that today's real-time applications demand, from ultra-fast data ingestion to seamless scale-out across hybrid and multi-cloud environments.",[40,135,137],{"id":136},"open-source-foundation-and-pulsar-expertise",[44,138,139],{},"Open Source Foundation and Pulsar Expertise",[48,141,142],{},"StreamNative's DNA is rooted in open source innovation. Our founders are the original creators of Apache Pulsar, and we've built our platform with the same open principles: freedom, flexibility, and community-driven innovation. For developers and data teams, this means adopting StreamNative comes with no proprietary lock-in—instead, you get a platform built on open standards and a thriving ecosystem. We offer broad API compatibility (Pulsar, Kafka, JMS, MQTT, and more) so that teams can work with familiar interfaces and integrate StreamNative into existing systems with ease.",[48,144,145],{},"StreamNative is the primary commercial contributor to the Apache Pulsar project and its surrounding ecosystem. We invest heavily in Pulsar's ongoing improvements our investments in Pulsar's open-source ecosystem and performance optimization bolster StreamNative's value. We also foster a vibrant community through initiatives like the Data Streaming Summit and free training resources.",[48,147,148,149,152,153,156],{},"Forrester's assessment noted that StreamNative’s “",[36,150,151],{},"events-driven agents, extensibility, and performance architecture are solid,","” and we're continuing to build on that foundation. ",[44,154,155],{},"We're actively investing in expanding our tooling for observability, governance, schema management, and developer productivity","—areas we recognize as critical for enterprise adoption and where we're committed to accelerating our roadmap.",[48,158,159,160],{},"Being open also means embracing an open ecosystem of technologies. StreamNative actively integrates with the tools and platforms that matter most to our users. We partner with industry leaders like Snowflake, Databricks, Google, and Ververica to ensure our streaming platform works seamlessly with data warehouses, lakehouse storage, and stream processing frameworks. Forrester’s evaluation observed that StreamNative’s ",[36,161,162],{},"\"investments in Pulsar’s open-source ecosystem and performance optimization make it the primary platform for enterprises wishing to implement Pulsar.\"",[40,164,166],{"id":165},"powering-real-time-use-cases-across-industries",[44,167,168],{},"Powering Real-Time Use Cases Across Industries",[48,170,171],{},"One of the greatest validations of StreamNative's approach is the success our customers are achieving with real-time data. StreamNative's platform is versatile and use-case agnostic—if an application demands high-volume, low-latency data movement, we can power it. This flexibility is why our customer base spans industries from finance and IoT to major automobile manufacturers and online gaming. The common thread is that these organizations need to process and react to data in milliseconds, and StreamNative is delivering the capabilities to make that possible.",[48,173,174],{},"Cisco uses StreamNative to underpin an IoT telemetry system of colossal scale, connecting hundreds of millions of devices and thousands of enterprise clients with real-time data streams. The platform's multi-tenant design and proven reliability allow Cisco to offer its customers a live feed of device data with unwavering confidence. In the financial sector, FICO has built streaming pipelines on StreamNative to detect fraud as transactions happen and to monitor systems in real time. With StreamNative's strong guarantees around message durability and ordering, FICO can catch anomalies or suspicious patterns within seconds. And in digital customer engagement, Iterable relies on StreamNative to process billions of events every day—clicks, views, purchases—so that marketers can trigger personalized campaigns instantly based on user behavior.",[48,176,177],{},"Our customers uniformly deal with mission-critical data streams, where downtime or delays are unacceptable. StreamNative's fault-tolerant, scalable infrastructure has proven equal to the task, handling scenarios like bursting to millions of events per second or seamlessly spanning multiple cloud regions. Forrester's report recognized StreamNative for supporting event-driven architectures with robust scalability—which for us is a reflection of our platform's ability to meet the most demanding enterprise requirements.",[40,179,181],{"id":180},"continuing-to-innovate-ursa-orca-and-the-road-ahead",[44,182,183],{},"Continuing to Innovate: Ursa, Orca, and the Road Ahead",[48,185,186,187,190],{},"While we are thrilled to be recognized in Forrester's Streaming Data Platforms Wave, we view this as just the beginning. StreamNative's vision has always been bold: to ",[44,188,189],{},"provide a unified platform that not only handles today's streaming needs but also anticipates the emerging requirements of tomorrow",".",[48,192,193],{},"One key area of focus is the convergence of streaming data with advanced analytics and AI. As Forrester points out in the report, technology leaders should look for platforms that natively integrate messaging, stream processing, and analytics to provide AI agents with real-time, contextualized information. We couldn't agree more. Our award-winning Ursa Engine and Orca Agent Engine are aimed at extending our platform up the stack—bridging the gap between data streams and data lakes, and between event streams and intelligent processing.",[48,195,196],{},"Our new Ursa Engine introduces a lakehouse-native approach to streaming: it can write events directly to table formats like Iceberg on cloud storage, eliminating entire classes of ETL jobs and making fresh data instantly available for analytics queries. By integrating streaming and lakehouse technologies, we help customers collapse data silos and accelerate their AI\u002FML pipelines.",[48,198,199,200,203],{},"Beyond analytics integration, we are also enhancing StreamNative with more out-of-the-box processing and governance capabilities. In the coming months, we plan to introduce new features for lightweight stream processing and transformation, making it easier to build reactive applications directly on the platform. We're also expanding our ecosystem of connectors and integrations, so that whether your data lands in Snowflake, Databricks, or an AI model, StreamNative will seamlessly feed it. ",[44,201,202],{},"We're investing significantly in enterprise features including security, schema registry, governance, and monitoring tooling","—capabilities that are essential for mission-critical deployments and where we're committed to continued improvement.",[48,205,206],{},"This recognition from Forrester energizes us to keep innovating at full speed. We're sharing this honor with our amazing customers, community, and partners who drive us forward every day. Your feedback and real-world challenges have helped shape StreamNative into what it is today, and together, we will shape the future of streaming data. Thank you for joining us on this journey—we're just getting started, and we can't wait to deliver even more value as we continue to evolve our platform. Onward to real-time everything!",[208,209],"hr",{},[32,211,213],{"id":212},"streamnative-in-the-forrester-wave-evaluation-findings",[44,214,215,216,223],{},"StreamNative in ",[44,217,218],{},[55,219,220],{"href":57},[44,221,222],{},"The Forrester Wave™",": Evaluation Findings",[225,226,228],"h5",{"id":227},"recognized-as-a-contender-among-15-streaming-data-platform-providers","• Recognized as a Contender among 15 streaming data platform providers",[225,230,232],{"id":231},"received-the-highest-scores-possible-50-in-both-the-messaging-and-resource-optimization-criteria","* Received the highest scores possible (5.0) in both the Messaging and Resource Optimization criteria",[225,234,236],{"id":235},"cited-as-the-primary-platform-for-enterprises-wishing-to-implement-pulsar","• Cited as the primary platform for enterprises wishing to implement Pulsar",[225,238,240],{"id":239},"noted-for-excelling-at-messaging-and-resource-optimization","• Noted for excelling at messaging and resource optimization",[225,242,244],{"id":243},"customers-cited-lower-infrastructure-costs-and-excellent-support-responsiveness","• Customers cited lower infrastructure costs and excellent support responsiveness",[225,246,248],{"id":247},"recognized-for-supporting-event-driven-architectures-with-robust-scalability","• Recognized for supporting event-driven architectures with robust scalability",[208,250],{},[252,253,255,256,259,260,190],"h6",{"id":254},"forrester-disclaimer-forrester-does-not-endorse-any-company-product-brand-or-service-included-in-its-research-publications-and-does-not-advise-any-person-to-select-the-products-or-services-of-any-company-or-brand-based-on-the-ratings-included-in-such-publications-information-is-based-on-the-best-available-resources-opinions-reflect-judgment-at-the-time-and-are-subject-to-change-for-more-information-read-about-forresters-objectivity-here","**Forrester Disclaimer: **",[36,257,258],{},"Forrester does not endorse any company, product, brand, or service included in its research publications and does not advise any person to select the products or services of any company or brand based on the ratings included in such publications. Information is based on the best available resources. Opinions reflect judgment at the time and are subject to change",". *For more information, read about Forrester’s objectivity *",[55,261,265],{"href":262,"rel":263},"https:\u002F\u002Fwww.forrester.com\u002Fabout-us\u002Fobjectivity\u002F",[264],"nofollow",[36,266,267],{},"here",[208,269],{},[252,271,273],{"id":272},"apache-apache-pulsar-apache-kafka-apache-flink-and-other-names-are-trademarks-of-the-apache-software-foundation-no-endorsement-by-apache-or-other-third-parties-is-implied",[36,274,275],{},"Apache®, Apache Pulsar®, Apache Kafka®, Apache Flink® and other names are trademarks of The Apache Software Foundation. No endorsement by Apache or other third parties is implied.",{"title":18,"searchDepth":19,"depth":19,"links":277},[278,280,281,282,283,284,285],{"id":34,"depth":279,"text":38},3,{"id":42,"depth":19,"text":46},{"id":83,"depth":19,"text":86},{"id":111,"depth":19,"text":114},{"id":136,"depth":19,"text":139},{"id":165,"depth":19,"text":168},{"id":180,"depth":19,"text":183,"children":286},[287],{"id":212,"depth":279,"text":288},"StreamNative in The Forrester Wave™: Evaluation Findings","Company",null,"2025-12-16","StreamNative is recognized in The Forrester Wave™: Streaming Data Platforms, Q4 2025. Discover why Forrester highlights StreamNative's high-performance messaging, efficient resource use, and cost-effective Kafka API compatibility for real-time innovation.","\u002Fimgs\u002Fblogs\u002F693bd36cf01b217dcb67278f_Streamnative_blog_thumbnail.png",false,{},0,"\u002Fblog\u002Fstreamnative-recognized-in-the-forrester-wave-streaming-data-platforms-2025","10 mins read",{"title":26,"description":292},"blog\u002Fstreamnative-recognized-in-the-forrester-wave-streaming-data-platforms-2025",[302,303,304],"Announcements","Real-Time","Forrester","sOeeJtEO3O-IIfTPJjY1AFOMawZ_rf8FOH8A98NEKgU",{"id":307,"title":308,"authors":309,"body":314,"category":1073,"createdAt":290,"date":1074,"description":1075,"extension":8,"featured":7,"image":1076,"isDraft":294,"link":290,"meta":1077,"navigation":7,"order":296,"path":1078,"readingTime":1079,"relatedResources":290,"seo":1080,"stem":1081,"tags":1082,"__hash__":1085},"blogs\u002Fblog\u002Fhow-we-run-a-5-gb-s-kafka-workload-for-just-50-per-hour.md","How We Run a 5 GB\u002Fs Kafka Workload for Just $50 per Hour",[310,311,312,313],"Matteo Meril","Neng Lu","Hang Chen","Penghui Li",{"type":15,"value":315,"toc":1043},[316,319,322,325,328,331,335,338,348,354,357,365,370,374,381,384,387,395,399,402,407,411,414,417,420,423,432,436,439,450,453,457,460,463,474,477,481,485,493,496,500,508,537,541,544,549,553,556,560,563,566,571,580,585,588,591,602,606,609,620,624,627,630,635,638,667,671,673,679,682,687,692,695,699,713,717,728,732,747,756,767,770,773,777,780,783,794,797,800,803,808,813,817,821,838,842,856,861,865,876,879,895,899,910,915,920,928,932,935,939,946,950,953,962,967,976,982,991,1000,1009,1018,1027,1035],[48,317,318],{},"The rise of DeepSeek has shaken the AI infrastructure market, forcing companies to confront the escalating costs of training and deploying AI models. But the real pressure point isn’t just compute—it’s data acquisition and ingestion costs.",[48,320,321],{},"As businesses rethink their AI cost-containment strategies, real-time data streaming is emerging as a critical enabler. The growing adoption of Kafka as a standard protocol has expanded cost-efficient options, allowing companies to optimize streaming analytics while keeping expenses in check.",[48,323,324],{},"Ursa, the data streaming engine powering StreamNative’s managed Kafka service, is built for this new reality. With its leaderless architecture and native lakehouse storage integration, Ursa eliminates costly inter-zone network traffic for data replication and client-to-broker communication while ensuring high availability at minimal operational cost.",[48,326,327],{},"In this blog post, we benchmarked the infrastructure cost and total cost of ownership (TCO) for running a 5GB\u002Fs Kafka workload across different Kafka vendors, including Redpanda, Confluent WarpStream, and AWS MSK. Our benchmark results show that Ursa can sustain 5GB\u002Fs Kafka workloads at just 5% of the cost of traditional streaming engines like Redpanda—making it the ideal solution for high-performance, cost-efficient ingestion and data streaming for data lakehouses and AI workloads.",[48,329,330],{},"Note: We also evaluated vanilla Kafka in our benchmark; however, for simplicity, we have focused our cost comparison on vendor solutions rather than self-managed deployments. That said, it is important to highlight that both Redpanda and vanilla Kafka use a leader-based data replication approach. In a data-intensive, network-bound workload like 5GB\u002Fs streaming, with the same machine type and replication factor, Redpanda and vanilla Kafka produced nearly identical cost profiles.",[40,332,334],{"id":333},"key-benchmark-findings","Key Benchmark Findings",[48,336,337],{},"Ursa delivered 5 GB\u002Fs of sustained throughput at an infrastructure cost of just $54 per hour. For comparison:",[339,340,341,345],"ul",{},[342,343,344],"li",{},"MSK: $303 per hour → 5.6x more expensive compared to Ursa",[342,346,347],{},"Redpanda: $988 per hour → 18x more expensive compared to Ursa",[48,349,350],{},[351,352],"img",{"alt":18,"src":353},"\u002Fimgs\u002Fblogs\u002F679c71b67d9046f26edc7977_AD_4nXfvTqyBNUBu2lObdkKAx-5UNkpNP8UYULLZyOcixE6z99VMZUUEsUqWjzexI7vjyNGRNSAUoM9smYvdTP55ctAhIbrs5lmQgcSVMWdaoigbWouCl95DVSQsxooY-qqfGcYqS4g4zA.png",[48,355,356],{},"Beyond infrastructure costs, when factoring in both storage pricing, vendor pricing and operational expenses, Ursa’s total cost of ownership (TCO) for a 5GB\u002Fs workload with a 7-day retention period is:",[339,358,359,362],{},[342,360,361],{},"50% cheaper than Confluent WarpStream",[342,363,364],{},"85% cheaper than MSK and Redpanda",[48,366,367],{},[351,368],{"alt":18,"src":369},"\u002Fimgs\u002Fblogs\u002F679c602d77e9c706de5343b8_AD_4nXeDv8rrv_C1CTCCiqYo1zpvlGYbdBk1r0VEqovAPu22iFMQZgh54Hfw9PBMLzM7jDFxKwAFDxbdG0np4XVk_tGsWhEKMloLRcmmea7lvueCx-0cFsyaE3Mya4Mxc1Dox95A6JEc.png",[40,371,373],{"id":372},"ursa-highly-cost-efficient-data-streaming-at-scale","Ursa: Highly Cost-Efficient Data Streaming at Scale",[48,375,376,380],{},[55,377,379],{"href":378},"\u002Fblog\u002Fursa-reimagine-apache-kafka-for-the-cost-conscious-data-streaming","Ursa"," is a next-generation data streaming engine designed to deliver high performance at a fraction of the cost of traditional disk-based solutions. It is fully compatible with Apache Kafka and Apache Pulsar APIs, while leveraging a leaderless, lakehouse-native architecture to maximize scalability, efficiency, and cost savings.",[48,382,383],{},"Ursa’s key innovation is separating storage from compute and decoupling metadata\u002Findex operations from data operations by utilizing cloud object storage (e.g., AWS S3) instead of costly inter-zone disk-based replication. It also employs open lakehouse formats (Iceberg and Delta Lake), enabling columnar compression to significantly reduce storage costs while maintaining durability and availability.",[48,385,386],{},"In contrast, traditional streaming systems—like Kafka and Redpanda—depend on leader-based architectures, which drive up inter-zone traffic costs due to replication and client communication. Ursa mitigates these costs by:",[339,388,389,392],{},[342,390,391],{},"Eliminating inter-zone traffic costs via a leaderless architecture.",[342,393,394],{},"Replacing costly inter-zone replication with direct writes to cloud storage using open lakehouse formats.",[40,396,398],{"id":397},"how-ursa-eliminates-inter-zone-traffic","How Ursa Eliminates Inter-Zone Traffic",[48,400,401],{},"Ursa minimizes inter-zone traffic by leveraging a leaderless architecture, which eliminates inter-zone communication between clients and brokers, and lakehouse-native storage, which removes the need for inter-zone data replication. This approach ensures high availability and scalability while avoiding unnecessary cross-zone data movement.",[48,403,404],{},[351,405],{"alt":18,"src":406},"\u002Fimgs\u002Fblogs\u002F679c602e21b3571bb7117dca_AD_4nXd7Oahc77NjRLNvA9clLt0tsyU6MrIqVibFYv5pW5giTIcCHPr3EA_yTGzfVEUIVO3VXK56qWK8zmBCp5lY0E_4nmlWIPFrHjtHylA5NhwELjn-UB0fLG2h_kbrxrc7Cs_edvveNA.png",[32,408,410],{"id":409},"leaderless-architecture","Leaderless architecture",[48,412,413],{},"Traditional streaming engines such as Kafka, Pulsar, or RedPanda rely on a leader-based model, where each partition is assigned to a single leader broker that handles all writes and reads.",[48,415,416],{},"Pros of Leader-Based Architectures:\n✔ Maintains message ordering via local sequence IDs\n✔ Delivers low latency and high performance through message caching",[48,418,419],{},"Cons of Leader-Based Architectures:\n✖ Throughput bottlenecked by a single broker per partition\n✖ Inter-zone traffic required for high availability in multi-AZ deployments",[48,421,422],{},"While Kafka and Pulsar offer partial solutions (e.g., reading from followers, shadow topics) to reduce read-related inter-zone traffic, producers still send data to a single leader.",[48,424,425,426,431],{},"Ursa removes the concept of topic ownership, allowing any broker in the cluster to handle reads or writes for any partition. The primary challenge—ensuring message ordering—is solved with ",[55,427,430],{"href":428,"rel":429},"https:\u002F\u002Fgithub.com\u002Fstreamnative\u002Foxia",[264],"Oxia",", a scalable metadata and index service created by StreamNative in 2022.",[32,433,435],{"id":434},"oxia-the-metadata-layer-enabling-leaderless-architecture","Oxia: The Metadata Layer Enabling Leaderless Architecture",[48,437,438],{},"Ensuring message ordering in a leaderless architecture is complex, but Ursa solves this with Oxia:",[339,440,441,444,447],{},[342,442,443],{},"Handles millions of metadata\u002Findex operations per second",[342,445,446],{},"Generates sequential IDs to maintain strict message ordering",[342,448,449],{},"Optimized for Kubernetes with horizontal scalability",[48,451,452],{},"Producers and consumers can connect to any broker within their local AZ, eliminating inter-zone traffic costs while maintaining performance through localized caching.",[32,454,456],{"id":455},"zero-interzone-data-replication","Zero interzone data replication",[48,458,459],{},"In most distributed systems, data replication from a leader (primary) to followers (replicas) is crucial for fault tolerance and availability. However, replication across zones can inflate infrastructure expenses substantially.",[48,461,462],{},"Ursa avoids these costs by writing data directly to cloud storage (e.g., AWS S3, Google GCS):",[339,464,465,468,471],{},[342,466,467],{},"Built-In Resilience: Cloud storage inherently offers high availability and fault tolerance without inter-zone traffic fees.",[342,469,470],{},"Tradeoff: Slightly higher latency (sub-second, with p99 at 500 milliseconds) compared to local disk\u002FEBS (single-digit to sub-100 milliseconds), in exchange for significantly lower costs (up to 10x lower).",[342,472,473],{},"Flexible Modes: Ursa is an addition to the classic BookKeeper-based engine, providing users with the flexibility to optimize for either cost or low latency based on their workload requirements.",[48,475,476],{},"By foregoing conventional replication, Ursa slashes inter-zone traffic costs and associated complexities—making it a compelling option for organizations seeking to balance high-performance data streaming with strict budget constraints.",[40,478,480],{"id":479},"how-we-ran-a-5-gbs-test-with-ursa","How We Ran a 5 GB\u002Fs Test with Ursa",[32,482,484],{"id":483},"ursa-cluster-deployment","Ursa Cluster Deployment",[339,486,487,490],{},[342,488,489],{},"9 brokers across 3 availability zones, each on m6i.8xlarge (Fixed 12.5 Gbps bandwidth, 32 vCPU cores, 128 GB memory).",[342,491,492],{},"Oxia cluster (metadata store) with 3 nodes of m6i.8xlarge, distributed across three availability zones (AZs).",[48,494,495],{},"During peak throughput (5 GB\u002Fs), each broker’s network usage was about 10 Gbps.",[32,497,499],{"id":498},"openmessaging-benchmark-workers-configuration","OpenMessaging Benchmark Workers & Configuration",[48,501,502,503,507],{},"The OpenMessaging Benchmark(OMB) Framework is a suite of tools that make it easy to benchmark distributed messaging systems in the cloud. Please check ",[55,504,505],{"href":505,"rel":506},"https:\u002F\u002Fopenmessaging.cloud\u002Fdocs\u002Fbenchmarks\u002F",[264]," for details.",[339,509,510,525,534],{},[342,511,512,513,518,519,524],{},"12 OMB workers: 6 for ",[55,514,517],{"href":515,"rel":516},"https:\u002F\u002Fgist.github.com\u002Fcodelipenghui\u002Fd1094122270775e4f1580947f80c5055",[264],"producers",", 6 for ",[55,520,523],{"href":521,"rel":522},"https:\u002F\u002Fgist.github.com\u002Fcodelipenghui\u002F06bada89381fb77a7862e1b4c1d8963d",[264],"consumers"," across 3 availability zones, on m6i.8xlarge instances. Each worker is configured with 12 CPU cores and 48 GB memory.",[342,526,527,528,533],{},"Sample YAML ",[55,529,532],{"href":530,"rel":531},"https:\u002F\u002Fgist.github.com\u002Fcodelipenghui\u002F204c1f26c4d44a218ae235bf2de99904",[264],"scripts"," provided for Kafka-compatible configuration and rate limits.",[342,535,536],{},"Achieved consistent 5 GB\u002Fs publish\u002Fsubscribe throughput.",[40,538,540],{"id":539},"ursa-benchmark-tests-results","Ursa Benchmark Tests & Results",[48,542,543],{},"The following diagram demonstrates that Ursa can consistently handle 5 GB\u002Fs of traffic, fully saturating the network across all broker nodes.",[48,545,546],{},[351,547],{"alt":18,"src":548},"\u002Fimgs\u002Fblogs\u002F679c602d7b261bac1113f7d6_AD_4nXdDPsRc3koXICiFF0bqSmGWbJt_RlUy4FE3ruuWOfbCfpcqZ1dejjqGbkaCJv2hQFL1nirRouBVRW2l5uMWBvY9naMqGB_wHcLI14dBM0f85TXhmdm3UxEv1yGX9Y4hf5FttSkZew.png",[40,550,552],{"id":551},"comparing-infrastructure-cost","Comparing Infrastructure Cost",[48,554,555],{},"This benchmark first evaluates infrastructure costs of running a 5 GB\u002Fs streaming workload (1:1 producer-to-consumer ratio) across different data streaming engines, including Ursa, Redpanda, and AWS MSK, with a focus on multi-AZ deployments to ensure a fair comparison.",[32,557,559],{"id":558},"test-setup-key-assumptions","Test Setup & Key Assumptions",[48,561,562],{},"All tests use multi-AZ configurations, with clusters and clients distributed across three AWS availability zones (AZs). Cluster size scales proportionally to the number of AZs, and rack-awareness is enabled for all engines to evenly distribute topic partitions and leaders.",[48,564,565],{},"To ensure a fair comparison, we selected the same machine type capable of fully utilizing both network and storage bandwidth for Ursa and Redpanda in this 5GB\u002Fs test:",[339,567,568],{},[342,569,570],{},"9 × m6i.8xlarge instances",[48,572,573,574,579],{},"However, MSK's storage bandwidth limits vary depending on the selected instance type, with the highest allowed limit capped at 1000 MiB\u002Fs per broker, according to",[55,575,578],{"href":576,"rel":577},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fmsk\u002Flatest\u002Fdeveloperguide\u002Fmsk-provision-throughput-management.html#throughput-bottlenecks",[264]," AWS documentation",". Given this constraint, achieving 5 GB\u002Fs throughput with a replication factor of 3 required the following setup:",[339,581,582],{},[342,583,584],{},"15 × kafka.m7g.8xlarge (32 vCPUs, 128 GB memory, 15 Gbps network, 4000 GiB EBS).",[48,586,587],{},"This configuration was necessary to work around MSK's storage bandwidth limitations, ensuring a comparable cost basis to other evaluated streaming engines.",[48,589,590],{},"Additional key assumptions include:",[339,592,593,596,599],{},[342,594,595],{},"Inter-AZ producer traffic: For leader-based engines, two-thirds of producer-to-broker traffic crosses AZs due to leader distribution.",[342,597,598],{},"Consumer optimizations: Follower fetch is enabled across all tests, eliminating inter-AZ consumer traffic.",[342,600,601],{},"Storage cost exclusions: This benchmark only evaluates streaming costs, assuming no long-term data retention.",[32,603,605],{"id":604},"inter-broker-replication-costs","Inter-Broker Replication Costs",[48,607,608],{},"Inter-broker (cross-AZ) replication is a major cost driver for data streaming engines:",[339,610,611,614,617],{},[342,612,613],{},"RedPanda: Inter-broker replication is not free, leading to substantial costs when data must be copied across multiple availability zones.",[342,615,616],{},"AWS MSK: Inter-broker replication is free, but MSK instance pricing is significantly higher (e.g., $3.264 per hour for kafka.m7g.8xlarge vs $1.306 per hour for an on-demand m7g.8xlarge). The storage price of MSK is $0.10 per GB-month which is significantly higher than st1, which costs $0.045 per GB-month. Even though replication is free, client-to-broker traffic still incurs inter-AZ charges.",[342,618,619],{},"Ursa: No inter-broker replication costs due to its leaderless architecture, eliminating inter-zone replication costs entirely.",[32,621,623],{"id":622},"zone-affinity-reducing-inter-az-costs","Zone Affinity: Reducing Inter-AZ Costs",[48,625,626],{},"We evaluated zone affinity mechanisms to further reduce inter-AZ data transfer costs.",[48,628,629],{},"Consumers:",[339,631,632],{},[342,633,634],{},"Follower fetch is enabled across all tests, ensuring consumers fetch data from replicas in their local AZ—eliminating inter-zone consumer traffic except for metadata lookups",[48,636,637],{},"Producers:",[339,639,640,649,658],{},[342,641,642,643,648],{},"Kafka protocol lacks an easy way to enforce producer AZ affinity (though ",[55,644,647],{"href":645,"rel":646},"https:\u002F\u002Fcwiki.apache.org\u002Fconfluence\u002Fdisplay\u002FKAFKA\u002FKIP-1123:+Rack-aware+partitioning+for+Kafka+Producer",[264],"KIP-1123"," aims to address this). And it only works with the default partitioner (i.e., when no record partition or record key is specified).",[342,650,651,652,657],{},"Redpanda recently introduced ",[55,653,656],{"href":654,"rel":655},"https:\u002F\u002Fdocs.redpanda.com\u002Fredpanda-cloud\u002Fdevelop\u002Fproduce-data\u002Fleader-pinning\u002F",[264],"leader pinning",", but this only benefits setups where producers are confined to a single AZ—not applicable to our multi-AZ benchmark.",[342,659,660,661,666],{},"Ursa is the only system in this test with ",[55,662,665],{"href":663,"rel":664},"https:\u002F\u002Fdocs.streamnative.io\u002Fdocs\u002Fconfig-kafka-client#eliminate-cross-az-networking-traffic",[264],"built-in zone affinity for both producers and consumers",". It achieves this by embedding producer AZ information in client.id, allowing metadata lookups to route clients to local-AZ brokers, eliminating inter-AZ producer traffic.",[32,668,670],{"id":669},"cost-comparison-results","Cost Comparison Results",[48,672,337],{},[339,674,675,677],{},[342,676,344],{},[342,678,347],{},[48,680,681],{},"Ursa’s leaderless architecture, zone affinity, and native cloud storage integration deliver unparalleled cost efficiency, making it the most cost-effective choice for high-throughput data streaming workloads.",[48,683,684],{},[351,685],{"alt":18,"src":686},"\u002Fimgs\u002Fblogs\u002F679c72208198ca36a352f228_AD_4nXeeZuM8T-xBlD4Vf3j67K618n08qh8wIDLLtiLJG0ssA1Wj1V26u7wIDTX9sqLrtw8mB2c299dwzarGen62CG0Vh7nWstn5qbPGFcBaKJYEepTsLr5fHWv1U8uqbg8Y0UOK6fJ7.png",[48,688,689],{},[351,690],{"alt":18,"src":691},"\u002Fimgs\u002Fblogs\u002F679c625978031f40229de484_AD_4nXdLkLLJ30KKr-_A_rN1j8akVwBYacAWIPzWHoOReJF421890kfByZoQQxkLczihVSmiw5Q9J51-V9I2SEKITbwsYnANDDTlAVL5nQ_jfaHNTe9VEWhSoa7DZooCnilDYL6l6msmJg.png",[48,693,694],{},"The detailed infrastructure cost calculations for each data streaming engine are listed below:",[32,696,698],{"id":697},"streamnative-ursa","StreamNative - Ursa",[339,700,701,704,707,710],{},[342,702,703],{},"Server EC2 costs: 9 * $1.536\u002Fhr = $14",[342,705,706],{},"Client EC2 costs: 9 * $1.536\u002Fhr =$14",[342,708,709],{},"S3 write requests costs: 1350 r\u002Fs * $0.005\u002F1000r * 3600s = $24",[342,711,712],{},"S3 read requests costs: 1350 r\u002Fs * $0.0004\u002F1000r * 3600s = $2",[32,714,716],{"id":715},"aws-msk","AWS MSK",[339,718,719,722,725],{},[342,720,721],{},"Server EC2 costs: 15 * $3.264\u002Fhr = $49",[342,723,724],{},"Client side EC2 costs: 9 * $1.536\u002Fhr =$14",[342,726,727],{},"Interzone traffic - producer to broker: 5GB\u002Fs * ⅔ * $0.02\u002FG(in+out) * 3600 = $240",[32,729,731],{"id":730},"redpanda","RedPanda",[339,733,734,736,738,741,744],{},[342,735,703],{},[342,737,706],{},[342,739,740],{},"Interzone traffic - producer to broker: 5GB\u002Fs * ⅔ * $0.02\u002FGB(in+out) * 3600 = $240",[342,742,743],{},"Interzone traffic - replication: 10GB\u002Fs * $0.02\u002FGB(in+out) * 3600 = $720",[342,745,746],{},"Interzone traffic - broker to consumer: $0 (fetch from local zone)",[48,748,749,750,755],{},"Please note that we were unable to test ",[55,751,754],{"href":752,"rel":753},"https:\u002F\u002Fwww.redpanda.com\u002Fblog\u002Fcloud-topics-streaming-data-object-storage",[264],"Redpanda with Cloud Topics",", as it remains an announced but unreleased feature and is not yet available for evaluation. Based on the limited information available, while Cloud Topics may help optimize inter-zone data replication costs, producers still need to traverse inter-availability zones to connect to the topic partition owners and incur inter-zone traffic costs of up to $240 per hour.",[339,757,758,764],{},[342,759,760,763],{},[55,761,647],{"href":645,"rel":762},[264]," (when implemented) will help mitigate producer-to-broker inter-zone traffic, but it is not yet available. And it only works with the default partitioner (no record partition or key is specified).",[342,765,766],{},"Redpanda’s leader pinning helps only when all producers for the pinned topic are confined to a single AZ. In multi-AZ environments (like our benchmark), inter-zone producer traffic remains unavoidable.",[48,768,769],{},"Additionally, Redpanda’s Cloud Topics architecture is not documented publicly. Their blog mentions \"leader placement rules to optimize produce latency and ingress cost,\" but it is unclear whether this represents a shift away from a leader-based architecture or if it uses techniques similar to Ursa’s zone-aware approach.",[48,771,772],{},"We may revisit this comparison as more details become available.",[40,774,776],{"id":775},"comparing-total-cost-of-ownership","Comparing Total Cost of Ownership",[48,778,779],{},"As highlighted earlier, with a BYOC Ursa setup, you can achieve 5 GB\u002Fs throughput at just 5% of the infrastructure cost of a traditional leader-based data streaming engine, such as Kafka or RedPanda, while managing the infrastructure yourself. This significant cost reduction is enabled by Ursa’s leaderless architecture and lakehouse-native storage design, which eliminate overhead costs such as inter-zone traffic and leader-based data replication. By leveraging a lakehouse-native, leaderless architecture, Ursa reduces resource requirements, enabling you to handle high data throughput efficiently and at a fraction of the cost of RedPanda.",[48,781,782],{},"Now, let’s examine the total cost comparison, evaluating Ursa alongside other vendors, including those that have adopted a leaderless architecture (e.g., Confluent WarpStream). This comparison is based on a 5GB\u002Fs workload with a 7-day retention period, factoring in both storage cost and vendor costs Here are the key findings:",[339,784,785,788,791],{},[342,786,787],{},"Ursa ($164,353\u002Fmonth) is: 50% cheaper than Confluent WarpStream ($337,068\u002Fmonth)",[342,789,790],{},"85% cheaper than AWS MSK ($1,115,251\u002Fmonth)",[342,792,793],{},"86% cheaper than Redpanda ($1,202,853\u002Fmonth)",[48,795,796],{},"In addition to Ursa’s architectural advantages—eliminating most inter-AZ traffic and leveraging lakehouse storage for cost-effective data retention—it also adopts a more fair and cost-efficient pricing model: Elastic Throughput-based pricing. This approach aligns costs with actual usage, avoiding unnecessary overhead.",[48,798,799],{},"Unlike WarpStream, which charges for both storage and throughput, Ursa ensures that customers only pay for the throughput they actively use. Ursa’s pricing is based on compressed data sent by clients, meaning the more data compressed on the client side, the lower the cost. In contrast, WarpStream prices are based on uncompressed data, unfairly inflating expenses and failing to incentivize customers to optimize their client applications.",[48,801,802],{},"This distinction is crucial, as compressed data reduces both storage and network costs, making Ursa’s pricing model not only more cost-effective but also more transparent and predictable.",[48,804,805],{},[351,806],{"alt":18,"src":807},"\u002Fimgs\u002Fblogs\u002F679c602d194800c9206d9d58_AD_4nXcFlf755xgyz7htxhMhBV5fGrsxy642mQNodt61DTok_z1dwkw5A6lkO5hatXVneCaB0anbZPAyvLI3MlIMuQEYLEACHHvQMOr5UfaB37dfzkdqewDEvcT-20VGd_zzvJsuA00zGA.png",[48,809,810],{},[351,811],{"alt":18,"src":812},"\u002Fimgs\u002Fblogs\u002F679c62594e9c2e629fae73aa_AD_4nXeU6cOgItnjLsEZCOf13TEvMY_SHWWIxYP2OYUj-B1GUPyWO78OG08K_v03hwYSVcg06f9dqDiGmdwy76vynjmiDGL5bluZ5_XF4nSU_r59oOZdfViXndXt6s11vVOY7qwfZN8v.png",[32,814,816],{"id":815},"cost-breakdown","Cost Breakdown",[818,819,820],"h4",{"id":697},"StreamNative – Ursa",[339,822,823,826,829,832,835],{},[342,824,825],{},"EC2 (Server): 9 × $1.536\u002Fhr × 24 hr × 30 days = $9,953.28",[342,827,828],{},"S3 Write Requests: 1,350 r\u002Fs × $0.005\u002F1,000 r × 3,600 s × 24 hr × 30 days = $17,496",[342,830,831],{},"S3 Read Requests: 1,350 r\u002Fs × $0.0004\u002F1,000 r × 3,600 s × 24 hr × 30 days = $1,400",[342,833,834],{},"S3 Storage Costs: 5 GB\u002Fs × $0.021\u002FGB × 3,600 s × 24 hr × 7 days = $63,504",[342,836,837],{},"Vendor Cost: 200 ETU × $0.50\u002Fhr × 24 hr × 30 days = $72,000",[818,839,841],{"id":840},"warpstream","WarpStream",[339,843,844,847],{},[342,845,846],{},"Based on WarpStream’s pricing calculator (as of January 29, 2025), we assume a 4:1 client data compression ratio, meaning 20 GB\u002Fs of uncompressed data translates to 5 GB\u002Fs of compressed data.",[342,848,849,850,855],{},"It's important to note that WarpStream’s pricing structure has fluctuated frequently throughout January. We observed the cost reported by their calculator changing from $409,644 per month to $337,068 per month. This variability has been previously highlighted in the blog post “",[55,851,854],{"href":852,"rel":853},"https:\u002F\u002Fbigdata.2minutestreaming.com\u002Fp\u002Fthe-brutal-truth-about-apache-kafka-cost-calculators",[264],"The Brutal Truth About Kafka Cost Calculators","”. To ensure transparency, we have documented the pricing as of January 29, 2025.",[48,857,858],{},[351,859],{"alt":18,"src":860},"\u002Fimgs\u002Fblogs\u002F679c602e42713e0028e9af5e_AD_4nXcu5_VWTLu9jRYs6zX1MBAOtLQEo5gyfNSWPcbpnQHXTa8qNCFAXezRR2E8daygzYTTwd4dhJjaLaLM8C6y_3OGbu2NS7pdvEv3a8-ptNKOg7AeKnYqPQCAYvQ5EuxzuI3JYIvY.png",[818,862,864],{"id":863},"msk","MSK",[339,866,867,870,873],{},[342,868,869],{},"EC2 (Server): 15 * $3.264\u002Fhr × 24 hr × 30 days = $35,251",[342,871,872],{},"Interzone Traffic (Client-Server): 5 GB\u002Fs × ⅔ × $0.02\u002FGB (in+out) × 3,600 s × 24 hr × 30 days = $172,800",[342,874,875],{},"Storage: 5 GB\u002Fs × $0.1\u002FGB-month × 3,600 s × 24 hr × 7 days * 3 replicas = $907,200",[818,877,731],{"id":878},"redpanda-1",[339,880,881,884,886,889,892],{},[342,882,883],{},"EC2 (Server): 9 × $1.536\u002Fhr × 24 hr × 30 days = $9953",[342,885,872],{},[342,887,888],{},"Interzone Traffic (Replication): 5 GB\u002Fs × 2 × $0.02\u002FGB (in+out) × 3,600 s × 24 hr × 30 days = $518,400",[342,890,891],{},"Storage: 5 GB\u002Fs × $0.045\u002FGB-month(st1) × 3,600 s × 24 hr × 7 days * 3 replicas = $408,240",[342,893,894],{},"Vendor Cost: $93,333 per month (based on limited information. See additional notes below).",[818,896,898],{"id":897},"additional-notes","Additional Notes",[339,900,901],{},[342,902,903,904,909],{},"Redpanda does not publicly disclose its BYOC pricing, making it difficult to accurately assess its total costs. We refer to information from the whitepaper “",[55,905,908],{"href":906,"rel":907},"https:\u002F\u002Fwww.redpanda.com\u002Fresources\u002Fredpanda-vs-confluent-performance-tco-benchmark-report#form",[264],"Redpanda vs. Confluent: A Performance and TCO Benchmark Report by McKnight Consulting Group.","” for estimation purposes. Based on the Tier-8 pricing model in the whitepaper,  the estimated cost to support a 5GB\u002Fs workload would be $1.12 million per year ($93,333 per month). However, since this calculation is based on an estimation, we will revisit and refine the cost assessment once Redpanda publishes its BYOC pricing.",[48,911,912],{},[351,913],{"alt":18,"src":914},"\u002Fimgs\u002Fblogs\u002F679c602dc8a9859eed89a0ef_AD_4nXdbcO8vsNNPy4GtkNLlmNKf22fjxRvzLzH7CtOna1L08sTbvnZx3HhufeFqc1w4K2gEF7lxO2IR5supotxebAiGnA07Qa8Yr3Rd1pVK2LYKK4WurlJGwgdwwucZIFoF-N_2oBjY.png",[48,916,917],{},[351,918],{"alt":18,"src":919},"\u002Fimgs\u002Fblogs\u002F679c602d6bc1c2287e012540_AD_4nXfcHZnLfjbjIr3ZAgoQXT9dwP3aQCOQPmGZZJUtpNZSwE6qY6M3yehIaBxCwxEIeu5PVdUPY0zhyjnow26YfgjdYgSG4GnV9ibxu0YWTIpwng6z_F6FUGJMpERMKtpsFESzXSN_Sw.png",[339,921,922,925],{},[342,923,924],{},"When estimating the storage costs for Kafka and Redpanda, we assume the use of HDD storage at $0.045\u002FGB, based on the premise that both systems can fully utilize disk bandwidth without incurring the higher costs associated with GP2 or GP3 volumes. However, in practice, many users opt for GP2 or GP3, significantly increasing the total storage cost for Kafka and Redpanda.",[342,926,927],{},"Unlike disk-based solutions, S3 storage does not require capacity preallocation—Ursa only incurs costs for the actual data stored. This contrasts with Kafka and Redpanda, where preallocating storage can drive up expenses. As a result, the real-world storage costs for Kafka and Redpanda are often 50% higher than the estimates above.",[40,929,931],{"id":930},"conclusion","Conclusion",[48,933,934],{},"Ursa represents a transformative shift in streaming data infrastructure, offering cost efficiency, scalability, and flexibility without compromising durability or reliability. By leveraging a leaderless architecture and eliminating inter-zone data replication, Ursa reduces total cost of ownership by over 90% compared to traditional leader-based streaming engines like Kafka and Redpanda. Its direct integration with cloud storage and scalable metadata & index management via Oxia ensure high availability and simplified infrastructure management.",[32,936,938],{"id":937},"balancing-latency-and-cost","Balancing Latency and Cost",[48,940,941,945],{},[55,942,944],{"href":943},"\u002Fblog\u002Fcap-theorem-for-data-streaming","Ursa trades off slightly higher latency for ultra low cost",", making it an ideal choice for the majority of streaming workloads, especially those that prioritize throughput and cost savings over ultra-low latency. Meanwhile, StreamNative’s BookKeeper-based engine remains the preferred solution for real-time, latency-sensitive applications. By combining these two approaches, StreamNative empowers customers with the flexibility to choose the right engine for their specific needs—whether it's maximizing cost savings or achieving ultra low-latency real-time performance.",[32,947,949],{"id":948},"the-future-of-streaming-infrastructure","The Future of Streaming Infrastructure",[48,951,952],{},"In an era where data fuels AI, analytics, and real-time decision-making, managing infrastructure costs is critical to sustaining innovation. Ursa is not just a cost-cutting alternative—it is a forward-thinking, lakehouse-native platform that redefines how modern data streaming infrastructure should be built and operated.",[48,954,955,956,961],{},"Whether your priority is reducing costs, improving flexibility, or ingesting massive data into lakehouses, Ursa delivers a future-proof solution for the evolving demands of real-time data streaming. ",[55,957,960],{"href":958,"rel":959},"https:\u002F\u002Fconsole.streamnative.cloud\u002F",[264],"Get started"," with StreamNative Ursa today!",[963,964,966],"h1",{"id":965},"references","References",[48,968,969,972,973],{},[970,971,430],"span",{}," ",[55,974,975],{"href":975},"\u002Fblog\u002Fintroducing-oxia-scalable-metadata-and-coordination",[48,977,978,972,980],{},[970,979,379],{},[55,981,378],{"href":378},[48,983,984,972,987],{},[970,985,986],{},"StreamNative pricing",[55,988,989],{"href":989,"rel":990},"https:\u002F\u002Fdocs.streamnative.io\u002Fdocs\u002Fbilling-overview",[264],[48,992,993,972,996],{},[970,994,995],{},"WarpStream pricing",[55,997,998],{"href":998,"rel":999},"https:\u002F\u002Fwww.warpstream.com\u002Fpricing#pricingfaqs",[264],[48,1001,1002,972,1005],{},[970,1003,1004],{},"AWS S3 pricing",[55,1006,1007],{"href":1007,"rel":1008},"https:\u002F\u002Faws.amazon.com\u002Fs3\u002Fpricing\u002F",[264],[48,1010,1011,972,1014],{},[970,1012,1013],{},"AWS EBS pricing",[55,1015,1016],{"href":1016,"rel":1017},"https:\u002F\u002Faws.amazon.com\u002Febs\u002Fpricing\u002F",[264],[48,1019,1020,972,1023],{},[970,1021,1022],{},"AWS MSK pricing",[55,1024,1025],{"href":1025,"rel":1026},"https:\u002F\u002Faws.amazon.com\u002Fmsk\u002Fpricing\u002F",[264],[48,1028,1029,972,1032],{},[970,1030,1031],{},"The Brutal Truth about Kafka Cost Calculators",[55,1033,852],{"href":852,"rel":1034},[264],[48,1036,1037,972,1040],{},[970,1038,1039],{},"Redpanda vs. Confluent: A Performance and TCO Benchmark Report by McKnight Consulting Group",[55,1041,906],{"href":906,"rel":1042},[264],{"title":18,"searchDepth":19,"depth":19,"links":1044},[1045,1046,1047,1052,1056,1057,1066,1069],{"id":333,"depth":19,"text":334},{"id":372,"depth":19,"text":373},{"id":397,"depth":19,"text":398,"children":1048},[1049,1050,1051],{"id":409,"depth":279,"text":410},{"id":434,"depth":279,"text":435},{"id":455,"depth":279,"text":456},{"id":479,"depth":19,"text":480,"children":1053},[1054,1055],{"id":483,"depth":279,"text":484},{"id":498,"depth":279,"text":499},{"id":539,"depth":19,"text":540},{"id":551,"depth":19,"text":552,"children":1058},[1059,1060,1061,1062,1063,1064,1065],{"id":558,"depth":279,"text":559},{"id":604,"depth":279,"text":605},{"id":622,"depth":279,"text":623},{"id":669,"depth":279,"text":670},{"id":697,"depth":279,"text":698},{"id":715,"depth":279,"text":716},{"id":730,"depth":279,"text":731},{"id":775,"depth":19,"text":776,"children":1067},[1068],{"id":815,"depth":279,"text":816},{"id":930,"depth":19,"text":931,"children":1070},[1071,1072],{"id":937,"depth":279,"text":938},{"id":948,"depth":279,"text":949},"StreamNative Cloud","2025-01-31","Discover how Ursa achieves 5GB\u002Fs Kafka workloads at just 5% of the cost of traditional streaming engines like Redpanda and AWS MSK. See our benchmark results comparing infrastructure costs, total cost of ownership (TCO), and performance across leading Kafka vendors.","\u002Fimgs\u002Fblogs\u002F679c6593d25099b1cdcec4ca_image-31.png",{},"\u002Fblog\u002Fhow-we-run-a-5-gb-s-kafka-workload-for-just-50-per-hour","30 min",{"title":308,"description":1075},"blog\u002Fhow-we-run-a-5-gb-s-kafka-workload-for-just-50-per-hour",[1083,1084,303],"TCO","Apache Kafka","A0o_2xdJiLI6rf6xj4RKsxJNo_A6QN2fYzCp6gaLrFw",{"id":1087,"title":1088,"authors":1089,"body":1091,"category":1519,"createdAt":290,"date":1520,"description":1521,"extension":8,"featured":294,"image":1522,"isDraft":294,"link":290,"meta":1523,"navigation":7,"order":296,"path":1524,"readingTime":1525,"relatedResources":290,"seo":1526,"stem":1527,"tags":1528,"__hash__":1530},"blogs\u002Fblog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis.md","Inside Apache Pulsar’s Millisecond Write Path: A Deep Performance Analysis",[1090],"Renyi Wang",{"type":15,"value":1092,"toc":1505},[1093,1100,1103,1106,1109,1113,1116,1125,1130,1133,1137,1140,1145,1148,1153,1156,1170,1173,1176,1179,1183,1186,1197,1202,1206,1209,1229,1234,1238,1241,1252,1255,1269,1272,1275,1278,1281,1284,1287,1291,1294,1305,1310,1313,1317,1320,1323,1326,1330,1338,1341,1345,1351,1361,1368,1372,1375,1378,1381,1384,1387,1390,1393,1396,1399,1402,1406,1409,1412,1414,1417,1420,1423,1425,1428,1431,1435,1438,1444,1447,1465,1469,1472,1475,1481,1483,1488,1491,1494,1497,1499,1502],[48,1094,1095,1096],{},"‍Original author: Renyi Wang (Software Engineer, 360 Cloud Platform Messaging Middleware Team). The blog post was originally published at ",[55,1097,1098],{"href":1098,"rel":1099},"https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002FQa2uzvO0oiBD9caDi763xw",[264],[48,1101,1102],{},"Apache Pulsar is an excellent distributed messaging system, purpose-built for modern, real-time data needs. Its compute–storage separation architecture offers significant advantages over many other open-source messaging queues, enabling both scalability and operational flexibility. Pulsar also comes with powerful features out of the box, such as delayed message delivery and cross-cluster geo-replication, making it resilient in mission-critical deployments.",[48,1104,1105],{},"What truly sets Pulsar apart, however, is its ability to combine enterprise-grade durability with millisecond latency. In optimized environments, Pulsar can persist messages across replicas in as little as 0.3 milliseconds, while achieving throughput of over 1.5 million messages per second on a single producer thread—all while preserving strict message ordering. This blog post takes a deep dive into how Pulsar achieves these remarkable performance characteristics and what makes its design uniquely capable of delivering both speed and reliability.",[48,1107,1108],{},"Testing environment (context & limits). All latency numbers in this blog post were measured in a low-latency setup: a producer client, two Pulsar brokers, and three bookies deployed in the same data center and connected through a single switch. Storage was backed by NVMe SSDs, and replication was configured as ensemble=2, write quorum=2, ack quorum=2. Under these conditions, one-way broker↔bookie network transit is approximately 50 µs, which is essential to achieving ~0.3 ms end-to-end durable writes. Traversing additional switches, crossing racks\u002FAZs\u002Fregions, enabling heavier inline processing (e.g., TLS inspection), or using slower disks will increase observed latencies. A fuller breakdown of hardware and settings appears later in this post.",[40,1110,1112],{"id":1111},"pulsar-write-latency-breakdown","Pulsar Write Latency Breakdown",[48,1114,1115],{},"A Pulsar producer sending a message incurs latency in two main stages:",[1117,1118,1119,1122],"ol",{},[342,1120,1121],{},"Client to Broker: The time for the client to send the message to the Pulsar broker.",[342,1123,1124],{},"Broker to Bookies: The time for the broker to persist the message by writing it to multiple bookie storage nodes in parallel (replicas in Apache BookKeeper).",[48,1126,1127],{},[351,1128],{"alt":18,"src":1129},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703d6_98e8a10d.png",[48,1131,1132],{},"In other words, the end-to-end publish latency includes the network hop from client to broker, plus the broker’s internal processing and the storage write to BookKeeper (which itself replicates to multiple bookies). We will analyze each part in detail in the next sections.",[40,1134,1136],{"id":1135},"broker-side-latency-analysis","Broker-Side Latency Analysis",[48,1138,1139],{},"First, we will look into the broker latency. Pulsar provides a metric pulsar_broker_publish_latency (illustrated in Figure 2.) which provides insight into the total time a message spends on the broker side, encompassing the period from its initial receipt to its successful write to BookKeeper and the subsequent client callback completion.",[48,1141,1142],{},[351,1143],{"alt":18,"src":1144},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703e5_1852d47c.png",[48,1146,1147],{},"To understand where the broker spends time handling incoming writes, we used Alibaba Arthas to capture a thread-level flame graph. The flame graph revealed that most of the broker latency is spent in a few network I\u002FO operations (sending data to bookies and waiting for responses). Figure 3 illustrates the broker-side flame graph, showing how the workload is distributed across threads.",[48,1149,1150],{},[351,1151],{"alt":18,"src":1152},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703dc_d218e99d.png",[48,1154,1155],{},"From this flame graph, we identified four main threads on the broker that contribute to publish latency. The breakdown of their time is as follows:",[339,1157,1158,1161,1164,1167],{},[342,1159,1160],{},"BookKeeperClientWorker-1 – accounts for roughly 10% of the end-to-end latency. Within this thread, about half of the time is spent dequeuing write requests, and the other half executing the bookie callback and enqueuing the next task.",[342,1162,1163],{},"BookKeeperClientWorker-2 – about 25% of latency. Approximately 40% of its time is spent taking entries from the bookie queue, 40% sending the write requests to bookies over the network, and ~20% running the broker-side callback when a bookie write succeeds.",[342,1165,1166],{},"pulsar-io-1 – about 25% of latency. It spends ~30% of the time executing tasks to write to the bookie’s network queue, ~30% reading the bookie’s network response and putting data into a queue, and ~30% waiting (idle or blocking on I\u002FO).",[342,1168,1169],{},"pulsar-io-2 – about 30% of latency. Roughly 40% is spent reading the client’s incoming write request and adding it to the bookie request queue, 40% on sending the ACK response back to the client, and ~20% waiting (idle).",[48,1171,1172],{},"These four threads—BookKeeperClientWorker and pulsar-io—collaborate to manage data flow and client responses. Specifically, BookKeeperClientWorker threads are responsible for sending data to BookKeeper, ensuring durable storage. Concurrently, pulsar-io threads handle responding to the client, confirming data receipt and processing. This distributed handling means that on the broker, there isn't a single, dominant bottleneck; instead, latency is diffused across various networking and callback processing stages, leading to a more robust and efficient write path.",[48,1174,1175],{},"Bookie-Side Latency Analysis",[48,1177,1178],{},"After understanding publish latency at the broker side, we know that a significant amount of time is spent writing data to bookies. Let's now examine bookie write latency. Once the broker forwards a message to BookKeeper, the BookKeeper bookies (storage nodes) take over. Pulsar’s storage layer is backed by BookKeeper, so understanding BookKeeper’s write path is crucial for optimizing latency. Below, we examine Pulsar’s data storage model, the internals of a bookie’s write operation, and how we tuned it for maximum performance.",[32,1180,1182],{"id":1181},"data-storage-model-in-pulsarbookkeeper","Data Storage Model in Pulsar\u002FBookKeeper",[48,1184,1185],{},"Pulsar’s data model is designed to handle millions of topics with high throughput. Each topic in Pulsar is backed by a BookKeeper ledger (an append-only log stored on bookies). Multiple ledgers are aggregated and written to ledger files in bookies (similar to a commit log in a database or RocksDB), which is optimized for sequential writes. This is illustrated in Figure 4.",[339,1187,1188,1191,1194],{},[342,1189,1190],{},"Bookies batch and sort writes to optimize disk access. Within a single topic, data is written in order, which means when that topic is consumed, the data is mostly sequential on disk. This improves read efficiency by reducing random disk seeks (fewer RocksDB lookups and disk reads).",[342,1192,1193],{},"Bookies employ a multi-tier caching mechanism. When bookies write messages, the data is first written to a Write Cache (in-memory) as well as to the Write-Ahead Log (WAL), and later to the main ledger storage. Consumers read from the Read Cache if possible; if the data is not in cache, the consumer will query RocksDB (which indexes the ledger files) to find the data location on disk. That data is then fetched and also put into the Read Cache for future reads.",[342,1195,1196],{},"At any given time, on each bookie disk, only one ledger file is open for writes. Each bookie also uses a RocksDB instance (for indexing entry locations). This design maximizes sequential write throughput by appending to a single file per disk at a time.",[48,1198,1199],{},[351,1200],{"alt":18,"src":1201},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703df_e78030cd.png",[32,1203,1205],{"id":1204},"write-process-in-a-bookie","Write Process in a Bookie",[48,1207,1208],{},"When a Pulsar broker publishes a message to a topic, the message is internally forwarded to BookKeeper clients and then to the bookie storage nodes. The end-to-end write process on a bookie is illustrated in Figure 5 and described as follows (solid arrows indicate synchronous\u002Fblocking steps in the flow):",[1117,1210,1211,1214,1217,1220,1223,1226],{},[342,1212,1213],{},"Broker to bookies: The Pulsar broker’s BookKeeper client selects a set of bookies (the ensemble for that ledger, e.g. 2 or 3 replicas) and sends the write request to all bookies in parallel.",[342,1215,1216],{},"Write to cache: Each bookie receives the write request and immediately writes the entry to its in-memory write cache. (By default, the write cache is sized to 1\u002F4 of the bookie’s heap and is off-heap memory.) The write is acknowledged in memory and will be flushed to disk asynchronously (allowing batching).",[342,1218,1219],{},"Write to WAL (Journal): If journaling is enabled (it is by default), the bookie also appends the entry to a journal file (WAL) on disk. This is done to ensure durability. The journal write is buffered and triggers a flush based on the journal’s flush policy (detailed below).",[342,1221,1222],{},"Journal thread flush: The bookie’s Journal thread pulls pending entries from a queue and writes them into an in-memory buffer. When this buffer is full or a flush condition is met, the data is written to the OS page cache (accumulating data to eventually be written to the physical disk).",[342,1224,1225],{},"Force write to disk: A separate ForceWrite thread is responsible for ensuring durability. It takes data that has been written to the page cache and issues an fsync (flush) to force the data to persist to the physical disk (this is often the slowest step, as it involves actual disk I\u002FO).",[342,1227,1228],{},"Acknowledge back to broker: Once the data is safely written (WAL fsynced) on the bookie, it sends a write acknowledgment back to the broker. After the broker receives write acknowledgments from a quorum of bookies, the broker then knows this entry is durably stored and can trigger the client’s callback to signal a successful publish.",[48,1230,1231],{},[351,1232],{"alt":18,"src":1233},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703e2_df1c90ff.png",[32,1235,1237],{"id":1236},"journal-flush-policy-tuning","Journal Flush Policy Tuning",[48,1239,1240],{},"Our previous analysis showed that journal flushing contributes to bookie write latency. BookKeeper flushes the journal (WAL) to disk when any of the following conditions is met (whichever comes first):",[339,1242,1243,1246,1249],{},[342,1244,1245],{},"Max wait time: 1 ms by default. The journal thread will flush the accumulated writes if 1 millisecond has passed since the last flush, even if there is little data (ensuring low latency).",[342,1247,1248],{},"Max batch size: 512 KB of data by default. If the buffered writes reach 512KB, it will flush immediately (to optimize throughput by writing larger sequential chunks).",[342,1250,1251],{},"Flush when idle: Disabled by default. If this is enabled, the journal will also flush as soon as the write queue becomes empty (i.e., no more writes to batch). This avoids waiting when there's a lull in traffic. We enabled this in our test to reduce latency for sporadic writes.",[48,1253,1254],{},"For our extreme latency tuning, we adjusted the bookie configuration as follows:",[339,1256,1257,1260,1263,1266],{},[342,1258,1259],{},"Enabled force flush of journal data from page cache to disk on each flush (journalSyncData=true) to ensure data is actually on disk before acknowledging.",[342,1261,1262],{},"Ensured the journal is actually written (journalWriteData=true, which is usually true by default).",[342,1264,1265],{},"Enabled flushing even when the queue is not full (journalFlushWhenQueueEmpty=true), which is safe for high-IOPS SSDs. This makes even single entries get flushed without delay (useful under light load; under heavy load, the batch triggers will naturally dominate).",[342,1267,1268],{},"Aligned the journal writes to disk sector boundaries for efficiency: we set journalAlignmentSize=4096 and readBufferSizeBytes=4096 (4 KB) to match the SSD’s physical sector size. (Note: The alignment settings require journalFormatVersionToWrite=5 or higher to take effect.)",[48,1270,1271],{},"These settings optimize the bookie to flush to NVMe disks as quickly as possible, trading off some CPU\u002FIO overhead for the lowest possible latency.",[48,1273,1274],{},"journalSyncData=true",[48,1276,1277],{},"journalWriteData=true",[48,1279,1280],{},"journalFlushWhenQueueEmpty=true",[48,1282,1283],{},"journalAlignmentSize=4096",[48,1285,1286],{},"readBufferSizeBytes=4096",[32,1288,1290],{"id":1289},"bookie-side-flame-graph-analysis","Bookie-Side Flame Graph Analysis",[48,1292,1293],{},"After applying the above optimizations, we profiled the bookie’s performance. The thread-level flame graph on the bookie side (Figure 5) showed that there were no abnormal blocking delays in the write path—each thread is doing its part efficiently. The time breakdown across key bookie threads was:",[339,1295,1296,1299,1302],{},[342,1297,1298],{},"Journal thread – ~17% of the total request latency. It spends roughly 30% of its time reading entries from the journal queue, 30% writing data into the OS page cache, and ~25% enqueueing data into the force-write queue (handing off to the ForceWrite thread).",[342,1300,1301],{},"ForceWrite thread – ~48% of latency (this is where the heavy disk I\u002FO happens). About 10% of its time is spent dequeuing data from the force-write queue, ~80% on forcing the data from page cache to disk (fsync calls), and ~10% handling the completion (notifying and queuing the response back to the network thread).",[342,1303,1304],{},"Bookie I\u002FO thread – ~28% of latency. This thread handles network I\u002FO. Around 30% of the time goes to parsing incoming write requests and adding them to the journal queue, ~30% executing tasks in the network queue (sending the acknowledgment back to the broker), and ~30% waiting (idle or blocking on network waits).",[48,1306,1307],{},[351,1308],{"alt":18,"src":1309},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703d9_4528a36f.png",[48,1311,1312],{},"With these optimizations, the bookie effectively pipelines the work: writing to the journal and flushing to disk happens concurrently with network communication. No single thread is stalling the process significantly, and the overall bookie write path is highly efficient.",[40,1314,1316],{"id":1315},"performance-testing","Performance Testing",[48,1318,1319],{},"Having optimized brokers and bookies for low latency, we conducted end-to-end throughput and latency tests to determine Pulsar's performance limits. To focus on single-thread performance and preserve message ordering, we utilized a single topic (partition) and producer. We tested both synchronous and asynchronous publishing modes with varying message sizes. The test environment and results are detailed below.",[48,1321,1322],{},"Test Environment: Brokers were deployed on 2 nodes (each 4 CPU cores, 16 GB RAM, 25 Gb Ethernet). Bookies were on 3 nodes (each with four 4 TB NVMe SSDs, and 25 Gb network). All nodes are deployed under the same network switch. The topic was configured with 2 bookie replicas (ensemble size=2, write quorum=2, ack quorum=2), meaning each message is written to 2 bookies and acknowledged when both succeed (for strong durability).",[48,1324,1325],{},"Before testing, we created a partitioned topic with 1 partition and set the persistence to 2 replicas and ack quorum 2:",[963,1327,1329],{"id":1328},"create-a-single-partition-topic","Create a single-partition topic",[48,1331,1332,1333,1337],{},"bin\u002Fpulsar-admin --admin-url ",[55,1334,1335],{"href":1335,"rel":1336},"http:\u002F\u002F192.0.0.1:8080",[264]," topics create-partitioned-topic persistent:\u002F\u002Fpublic\u002Fdefault\u002Ftest_qps -p 1",[48,1339,1340],{},"‍",[963,1342,1344],{"id":1343},"set-the-namespace-persistence-bookie-ensemble-size-2-write-quorum-2-ack-quorum-2","Set the namespace persistence: bookie ensemble size = 2, write quorum = 2, ack quorum = 2",[48,1346,1332,1347,1350],{},[55,1348,1335],{"href":1335,"rel":1349},[264]," namespaces set-persistence public\u002Fdefault \\",[1352,1353,1358],"pre",{"className":1354,"code":1356,"language":1357},[1355],"language-text","--bookkeeper-ensemble 2 \\\n\n--bookkeeper-write-quorum 2 \\\n\n--bookkeeper-ack-quorum 2\n","text",[1359,1360,1356],"code",{"__ignoreMap":18},[48,1362,1363,1364,1367],{},"For synchronous publishing (each send waits for acknowledgment before sending the next), with pulsar-perf tool we used ",[1359,1365,1366],{},"--batch-max-messages 1"," and we used a single producer thread with no batching (to observe latency per message):",[963,1369,1371],{"id":1370},"synchronous-publish-single-thread-measuring-latency","Synchronous publish, single thread, measuring latency",[48,1373,1374],{},"bin\u002Fpulsar-perf produce persistent:\u002F\u002Fpublic\u002Fdefault\u002Ftest_qps \\",[48,1376,1377],{},"-u pulsar:\u002F\u002F192.0.0.1:6650 \\",[48,1379,1380],{},"--disable-batching \\",[48,1382,1383],{},"--batch-max-messages 1 \\",[48,1385,1386],{},"--max-outstanding 1 \\",[48,1388,1389],{},"--rate 500000 \\",[48,1391,1392],{},"--test-duration 120 \\",[48,1394,1395],{},"--busy-wait \\",[48,1397,1398],{},"--size 1024 > sync_1024.log &",[48,1400,1401],{},"For asynchronous publishing with batching, we allowed a large batch and higher outstanding messages to maximize throughput (while preserving message order on a single thread). We also enabled compression (LZ4) to improve throughput for larger messages:",[963,1403,1405],{"id":1404},"asynchronous-publish-batching-and-compression-enabled-measuring-throughput","Asynchronous publish, batching and compression enabled, measuring throughput",[48,1407,1408],{},"export OPTS=\"-Xms10g -Xmx10g -XX:MaxDirectMemorySize=10g\"",[48,1410,1411],{},"bin\u002Fpulsar-perf produce persistent:\u002F\u002Fpublic\u002Fdefault\u002Ftest_qps_async \\",[48,1413,1377],{},[48,1415,1416],{},"--batch-max-messages 10000 \\",[48,1418,1419],{},"--memory-limit 2G \\",[48,1421,1422],{},"--rate 2000000 \\",[48,1424,1395],{},[48,1426,1427],{},"--compression LZ4 \\",[48,1429,1430],{},"--size 1024 > async_1024.log &",[32,1432,1434],{"id":1433},"throughput-and-latency-results","Throughput and Latency Results",[48,1436,1437],{},"After running the tests for long enough duration in each scenario, we gathered the maximum sustainable throughput (QPS) and the average latency observed, for various message sizes. The results are summarized below:",[48,1439,1440],{},[351,1441],{"alt":1442,"src":1443},"__wf_reserved_inherit","\u002Fimgs\u002Fblogs\u002F68c2d9ea6f9c7d211893391f_iShot_2025-09-11_22.16.55.png",[48,1445,1446],{},"Key Takeaways:",[339,1448,1449,1452,1455,1462],{},[342,1450,1451],{},"In synchronous mode, a single producer could send ~3200–3400 messages per second for small messages (up to 16KB), limited by the one-at-a-time round-trip to the broker. The average end-to-end latency for each message (client send -> stored on 2 bookies -> ack received) was only about 0.3 milliseconds! This is incredibly low and mainly consists of network propagation and context switching time. Even a 512 KB message was acknowledged in ~1.4 ms on average (throughput ~2.75 Gb\u002Fs), showing Pulsar’s ability to handle large messages with low latency.",[342,1453,1454],{},"In asynchronous mode with batching, Pulsar achieved over 1 million writes per second on a single producer thread to a single topic. With 1 KB messages, we saw about 1.06 million msgs\u002Fs (~8.3 Gb\u002Fs). With compression enabled (LZ4), the throughput increased to about 1.5 million msgs\u002Fs for 128-byte messages (since compression reduces the data size, effectively pushing more messages through per second). The trade-off was a higher average latency of ~5–6 ms (because batches of messages are sent and flushed together).",[342,1456,1457,1458,1461],{},"At very high QPS with small messages (128 B and 1 KB), throughput is constrained primarily by per-message CPU overhead on the broker\u002Fbookie (Netty, callbacks), callback scheduling and GC, plus the journal\u002Fforce-write (fsync) pipeline on bookies — the link is ",[44,1459,1460],{},"not"," NIC-limited in these cases. As message size grows (e.g., 16 KB), the bottleneck shifts toward NIC and disk throughput, while GC remains a secondary factor. In such tests, the 25 Gb\u002Fs network was nearly saturated (e.g., ~9.5 Gb\u002Fs per bookie for 16 KB messages, which is ~19 Gb\u002Fs total for 2 bookies).",[342,1463,1464],{},"Importantly, even in asynchronous mode, Pulsar maintains message order. The Pulsar client library and broker ensure that callbacks are executed in order for a given producer, so batching does not reorder messages. Also, using multiple threads did not improve throughput for a single topic\u002Fpartition because Pulsar uses a single IO thread per partition to preserve ordering (all messages for one partition go through the same channel and IO thread).",[32,1466,1468],{"id":1467},"disk-io-microbenchmark","Disk I\u002FO Microbenchmark",[48,1470,1471],{},"To better understand the lower bound of latency, we also measured the raw disk performance for fsync on the NVMe drives. Using fio, we simulated a single-thread writing 1KB to the page cache and immediately fsyncing (forced flush to disk):",[48,1473,1474],{},"fio --name=fsync_test --filename=\u002Fdata2\u002Ftestfile --bs=1k --size=1k --rw=write \\",[1352,1476,1479],{"className":1477,"code":1478,"language":1357},[1355],"--ioengine=sync --fsync=1 --numjobs=1 --iodepth=1 --direct=0 \\\n\n--group_reporting --runtime=60 --time_based\n",[1359,1480,1478],{"__ignoreMap":18},[48,1482,1340],{},[48,1484,1485],{},[351,1486],{"alt":18,"src":1487},"\u002Fimgs\u002Fblogs\u002F68c2de286f9c7d21189703e8_4c5e2de6.png",[48,1489,1490],{},"The result showed that an NVMe disk can handle a single-threaded sequential write+fsync in roughly 44 microseconds on average (about 18 µs to write to the page cache and 26 µs to flush to disk). In our Pulsar bookie tests, a single message fsync (journal write) took on the order of ~100 µs. The slight increase is due to additional overhead in the bookie (thread context switches, queue synchronization, etc., as seen in the flame graph breakdown).",[48,1492,1493],{},"Another factor in end-to-end latency is network propagation. Within the same availability zone (low network latency environment), we observed a one-way network transit time of roughly 0.05 ms (50 µs) between broker and bookie. Since our test used two bookies and required both to acknowledge, the client’s message experienced two network hops (to two bookies) plus the return hop from the broker.",[48,1495,1496],{},"Combining these factors: ~100 µs to durably write to an NVMe on each bookie, plus ~50 µs network each way, plus some processing overhead, it matches our observed ~0.3 ms end-to-end latency for a synchronous write with 2 replicas. This confirms that Pulsar’s architecture, when running on high-performance hardware (NVMe SSDs, 25GbE network), can indeed achieve sub-millisecond durable message writes.",[40,1498,931],{"id":930},[48,1500,1501],{},"This deep dive into Apache Pulsar’s performance demonstrates its ability to achieve ultra-low latency and high throughput with the right tuning and hardware. By leveraging a tiered architecture (separating compute and storage), optimizing write paths, and batching intelligently, Pulsar was able to reliably persist messages to multiple NVMe-backed replicas and acknowledge the client in about 0.3 milliseconds on average. In asynchronous mode, a single producer on one topic achieved on the order of 1 million messages per second, and up to 1.5 million msgs\u002Fs with compression, all while preserving message ordering.",[48,1503,1504],{},"Such performance is impressive for a distributed messaging system with strong durability guarantees. It showcases that Apache Pulsar’s design – with its write-ahead logs, caches, and efficient BookKeeper storage – can push the boundaries of messaging speed. For developers with demanding low-latency, high-throughput messaging needs, Pulsar’s architecture offers a compelling solution that can deliver lightning-fast data streaming without sacrificing reliability.",{"title":18,"searchDepth":19,"depth":19,"links":1506},[1507,1508,1514,1518],{"id":1111,"depth":19,"text":1112},{"id":1135,"depth":19,"text":1136,"children":1509},[1510,1511,1512,1513],{"id":1181,"depth":279,"text":1182},{"id":1204,"depth":279,"text":1205},{"id":1236,"depth":279,"text":1237},{"id":1289,"depth":279,"text":1290},{"id":1315,"depth":19,"text":1316,"children":1515},[1516,1517],{"id":1433,"depth":279,"text":1434},{"id":1467,"depth":279,"text":1468},{"id":930,"depth":19,"text":931},"Apache Pulsar","2025-09-11","Discover how Apache Pulsar achieves sub-millisecond durable writes and 1M+ msgs\u002Fsec throughput. A deep dive into its high-performance write path design.","\u002Fimgs\u002Fblogs\u002F68c2e16ccf31a040e2f87e3e_Inside-Apache-Pulsar’s-Millisecond-Write-Path-no-logo.png",{},"\u002Fblog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis","8 min read",{"title":1088,"description":1521},"blog\u002Finside-apache-pulsars-millisecond-write-path-a-deep-performance-analysis",[1519,1529],"Intro","-kXZWqQ36BCf0Ky-Gs33qYeEiJS9BPoc8x3EGB8FMZA",[1532],{"id":1533,"title":1090,"bioSummary":1534,"email":290,"extension":8,"image":1535,"linkedinUrl":290,"meta":1536,"position":290,"stem":1543,"twitterUrl":290,"__hash__":1544},"authors\u002Fauthors\u002Frenyi-wang.md","Renyi Wang is a software engineer on the 360 Cloud Platform Messaging Middleware team, specializing in Kafka and Pulsar. During his tenure at 360, he resolved the cascading Kafka cluster failures frequently triggered by short-connection DDoS attacks and built a serverless message-queuing platform based on Apache Pulsar, delivering tens-fold cost savings. He is an Apache Pulsar contributor.","\u002Fimgs\u002Fauthors\u002Frenyi-wang.jpg",{"body":1537},{"type":15,"value":1538,"toc":1541},[1539],[48,1540,1534],{},{"title":18,"searchDepth":19,"depth":19,"links":1542},[],"authors\u002Frenyi-wang","_o1hZgZh-buH4e2_JDKYk_gkg_XHrtnKSR46T-175IQ",[1546,1554,1559],{"path":1547,"title":1548,"date":1549,"image":1550,"link":-1,"collection":1551,"resourceType":1552,"score":1553,"id":1547},"\u002Fblog\u002Fapache-pulsar-seven-years-on-what-we-built-what-we-learned-whats-next","Apache Pulsar, Seven Years On: What We Built, What We Learned, What’s Next","2025-09-25","\u002Fimgs\u002Fblogs\u002F68d4ddd72eeca005c8fc8334_Pulsar-7-years.png","blogs","Blog",1,{"path":1555,"title":1556,"date":1557,"image":1558,"link":-1,"collection":1551,"resourceType":1552,"score":1553,"id":1555},"\u002Fblog\u002Fcompliance-and-data-governance-with-apache-pulsar-and-streamnative","Enhance Your Compliance and Data Governance with Apache Pulsar and StreamNative","2023-07-24","\u002Fimgs\u002Fblogs\u002F64bed6a5fbd7456e630fde0e_compliance.png",{"path":1560,"title":1561,"date":1562,"image":1563,"link":-1,"collection":1551,"resourceType":1552,"score":1553,"id":1560},"\u002Fblog\u002Fsharpen-your-apache-pulsar-skills-with-streamnatives-hands-on-self-paced-courses","Sharpen Your Apache Pulsar Skills with StreamNative’s Hands-On Self-Paced Courses","2023-04-13","\u002Fimgs\u002Fblogs\u002F643896e22543f3724800be2d_SNAcademyBlog.jpg",1775615042013]