Tech news in 3 minutes
IT Press Tour 68: Paradigm4
During the 68th IT Press Tour in Boston, MA, Joe Zhou, DevRel at JuiceFS (Juicedata), presented the company's file storage approach, comparing it with AWS S3 Files, released in April 2026. Zhou positioned object storage as the de facto backend for modern data infrastructure, citing its simple primitives (PUT, GET, Compare-And-Swap), flat namespace, unlimited scalability, 11 nines durability, multi-region availability, immutability, and low cost (~$0.02/GB/month). However, object stores lack in-place updates, real file hierarchy, efficient batch metadata operations, and are poor for structured data, driving a new generation of databases and file systems (e.g., JuiceFS, Neon) to rebuild richer APIs on top of immutable object buckets. Zhou described S3 Files as "a decent approach" but limited. It mounts S3 via NFS using EFS as a cache and metadata tier, with 1-to-1 file-object mapping. Defaults optimize for small files (sizeLessThan: 128 KiB), writes land in EFS and offload to S3 after ~60 seconds. Trade-offs include write amplification (appending to large files requires full object rewrite), costly metadata ops (e.g., mv over millions of keys triggers background rewrite), batching delays, S3-wins conflict resolution, no multi-cloud, and complex layered pricing. S3 Files delivers POSIX and hierarchy but fails on in-place updates, batch metadata, execution, and structured data. JuiceFS addresses these via strict data/metadata separation and file chunking. Community Edition (Apache 2.0) pairs external metadata engines (Redis, TiKV, MySQL, etc.) with any object store, exposed via FUSE, Java SDK, Python SDK, CSI Driver, and S3 Gateway. Files are chunked into 4 MB immutable blocks, so appends rewrite only the last block, and renames are single metadata ops. It supports configurable storage-class tiering. Enterprise Edition adds a Raft-based horizontally scalable metadata engine (3-copy redundancy), distributed data cache, cross-region replication, multi-cloud mirroring, and a soft limit of 500 billion files per volume (one EE deployment: 1.47 PiB, 404 billion inodes). Adopters include MiniMax, HeyGen, GMI, PixVerse, Xiaomi, NAVER, Trip.com, and Fly.io. EE pricing is by source-region capacity only, no per-client fees or JuiceFS-imposed data transfer charges.