--- title: USDB: Scaling Storage for Next-Generation Recommender... description: ByteDance built USDB, a unified sequence storage system for recommender systems handling both streaming and batch workloads. It uses CascadeLSM time-based... canonical: https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance | daily.dev og:description: ByteDance built USDB, a unified sequence storage system for recommender systems handling both streaming and batch workloads. It uses CascadeLSM time-based... og:url: https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc og:image: https://api.daily.dev/og/posts/KMxCJKqQC.png og:image:alt: USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance **[Hacker News](https://daily.dev/sources/hn)** · 2 min read · 0 upvotes · 0 comments ## Summary ByteDance built USDB, a unified sequence storage system for recommender systems handling both streaming and batch workloads. It uses CascadeLSM time-based compaction, differentiated replicas sharing persistent files, partial dumps for inference input reconstruction, and a hybrid columnar format for faster projected reads. Production migration data shows 85.8% lower mean write latency, while normalized deployment accounting shows 42.1% lower infrastructure cost. Compared to an optimized RocksDB sequence baseline, USDB delivers faster long-history reads, and aggregation pushdown sustains 8.5 times the throughput of ordinary reads under a constrained receive budget. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much does ByteDance's USDB storage system reduce write latency compared to its previous system? Production migration records show an 85.8% reduction in mean write latency after moving to USDB. USDB is ByteDance's unified sequence storage system for recommender systems, using CascadeLSM time-based compaction, differentiated replicas sharing persistent files, and a hybrid columnar format for faster reads, alongside a 42.1% drop in infrastructure cost. _Teams scaling recommender storage can track infrastructure case studies like this one on daily.dev._ ### How does USDB's read performance compare to a RocksDB-based sequence storage baseline for recommendation data? Controlled comparisons against an optimized RocksDB sequence baseline show USDB achieving faster long-history reads across different projections and client loads. Additionally, aggregation pushdown in USDB sustains 8.5 times the throughput of ordinary reads when operating under a constrained receive budget. _Developers evaluating storage backends for large-scale read workloads can follow benchmarks like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 8 comments across hackernews (as of 2026-11-04). **TL;DR:** Discussion barely engages with the storage system itself, instead drifting into tangents about generative/LLM-based recommenders and the practical difficulty of adopting such approaches outside large companies. **Sentiment:** 10% positive · 60% mixed · 30% skeptical **The pushback** - Generative recommender approaches face real-world adoption barriers like cold starts, explainability, and training/prediction cost. - Expertise in generative recommenders is largely siloed within major social media companies, leaving little public tutorial material. - Smaller companies lack the resources to build custom storage systems like this, relying instead on workarounds such as Redis sorted sets. **By community** - hackernews (mixed): Commenters mostly sidestep the storage architecture to debate what 'agent-based' recommendation even means and lament the resource gap between big tech and smaller companies. **Hottest debate:** Whether describing feed algorithms as an 'agent' paging through data is a meaningful distinction or just rebranding of existing recommendation techniques. **Open questions** - Is there an accessible tutorial or pipeline for adopting generative recommenders outside large in-house teams? - Can smaller companies realistically bridge the gap from classical discriminative recommenders to generative/LRM approaches? **Highlights** > So bytedance has the luxury of building custom storage for this, meanwhile we're over here trying to shoehorn user sequences into redis sorted sets > — [bashhq\_r on hackernews](https://news.ycombinator.com/item?id=50356674) > Setting up a classical DLRM at work, show of hands: who has managed to bridge the gap to generative recommenders or LRMs? Our survey identified quite a substantial list of challenges: product cardinality (Semantic IDs?), low performance vs discriminative recommenders (our user intents are heavily biases towards repetitive products), cold starts, entity features, missing explainability suite, training cost, prediction cost, etc. Has anyone done it easily? Where is the tutorial for generative recommenders ? > — [BenoitP on hackernews · 1 comments](https://news.ycombinator.com/item?id=50355558) > What does this actually mean? The term "agent" is designed to evoke an image of a person riffling through their files on you but can't the way you just described it describe any feed algorithm, including chronological? > — [direwolf20 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50359913) > GR expertise is essentially only in-house at major social media companies or people hired from these companies. Closest thing to a tutorial for LRM is the Kuaishou onerec papers. > — [whimsicalism on hackernews · 1 comments](https://news.ycombinator.com/item?id=50356158) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50354250) · 36 points · 8 comments ## Similar posts on daily.dev - [When your data model is the bottleneck: lessons from Medium’s feature store](https://daily.dev/posts/when-your-data-model-is-the-bottleneck-lessons-from-medium-s-feature-store-19cd3kzv5) · The New Stack · 1 upvotes · 0 comments - [Uber Moves from Static Limits to Priority-Aware Load Control for Distributed Storage](https://daily.dev/posts/uber-moves-from-static-limits-to-priority-aware-load-control-for-distributed-storage-238nm1f0q) · InfoQ · 0 upvotes · 0 comments - [HPTS 2026 \(Day 2 and Day 3\)](https://daily.dev/posts/hpts-2026-day-2-and-day-3--xashwu5jv) · Metadata · 1 upvotes · 0 comments - [DumboDB: 0.8 Performance Update](https://daily.dev/posts/dumbodb-0-8-performance-update-6mbu4bxjj) · DoltHub Blog · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning) [View this post on daily.dev](https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance","url":"https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc"},"datePublished":"2026-11-03T20:53:43.887Z","dateModified":"2026-11-04T04:33:41.918Z","description":"ByteDance built USDB, a unified sequence storage system for recommender systems handling both streaming and batch workloads. It uses CascadeLSM time-based...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1fed0de190485cbbb37a86fdf656d71b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/1fed0de190485cbbb37a86fdf656d71b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/usdb-scaling-storage-for-next-generation-recommender-systems-at-bytedance-kmxcjkqqc#faq","mainEntity":[{"@type":"Question","name":"How much does ByteDance's USDB storage system reduce write latency compared to its previous system?","acceptedAnswer":{"@type":"Answer","text":"Production migration records show an 85.8% reduction in mean write latency after moving to USDB. USDB is ByteDance's unified sequence storage system for recommender systems, using CascadeLSM time-based compaction, differentiated replicas sharing persistent files, and a hybrid columnar format for faster reads, alongside a 42.1% drop in infrastructure cost. Teams scaling recommender storage can track infrastructure case studies like this one on daily.dev."}},{"@type":"Question","name":"How does USDB's read performance compare to a RocksDB-based sequence storage baseline for recommendation data?","acceptedAnswer":{"@type":"Answer","text":"Controlled comparisons against an optimized RocksDB sequence baseline show USDB achieving faster long-history reads across different projections and client loads. Additionally, aggregation pushdown in USDB sustains 8.5 times the throughput of ordinary reads when operating under a constrained receive budget. Developers evaluating storage backends for large-scale read workloads can follow benchmarks like this on daily.dev."}}]} ```