Download Latest Version cubefs-3.5.3-linux-amd64.tar.gz (104.5 MB)
Email in envelope

Get an email when there's a new version of CubeFS

Home / v3.5.3
Name Modified Size InfoDownloads / Week
Parent folder
cubefs-3.5.3-linux-amd64.tar.gz.asc 2025-12-23 490 Bytes
0
cubefs-3.5.3-linux-amd64.tar.gz.sha256sum 2025-12-23 98 Bytes
0
cubefs-3.5.3-linux-amd64.tar.gz 2025-12-23 104.5 MB
0
README.md 2025-12-22 3.5 kB
0
Release v3.5.3 - 2025_12_23 source code.tar.gz 2025-12-22 52.3 MB
0
Release v3.5.3 - 2025_12_23 source code.zip 2025-12-22 55.6 MB
0
Totals: 6 Items   212.5 MB 0

UPGRAGDE NOTICE

If you are using a CubeFS version earlier than v3.5.0, please refer to the UPGRADE NOTICE in version v3.5.0 for detailed upgrade steps and upgrade to v3.5.0 first.

Upgrade nodes in this order: flshnode → master → datanode → metanode → objectnode → lcnode → cli → client.

Upgrade lcnode and flashnode when needed.

Deploy flashgroupmanager when needed

Clients should use versions later than 3.2.0. Older versions need to be upgraded promptly; otherwise, there will be a risk of compromising stability.

Main Feature

High-throughput LLM/MLLM training (Tolerable storage-compute disaggregation latency: 8ms)

  • client: Support asynchronous flush for extent handler to improve write performance. Write speed exceeds 1.2 GB/s; on a high-spec H20 training node, a single client can achieve 10+ GB/s aggregate throughput with 10 concurrent large-file writes.(#3973,@bboyCH4 @leonrayang @Victor1319)
  • client: Optimize the client read-ahead mechanism and memory footprint; single file(10GB) read speeds exceed 2 GB/s. (#3982,@bboyCH4 @Victor1319 )
  • client: Metadata cache acceleration for small-file prewarm (#3995,@Victor1319)

Note: Refer to the latest community documentation for enabling and tuning.

Distributed cache can run as an independent service

  • flashgroupmanager: Introduce flashgroupmanager node and topology to support flashnode cluster management. (@bboyCH4)
  • flashnode: Support block-level data read and write operations. (#3977, @clinx)
  • flashnode: Support Master issuing cache warmup tasks to FlashNode for directory warmup. (#3997, @clinx)
  • tools: Add rctest (benchmark) and rcconfig (config) tools for remote cache system. (#3981,@bboyCH4,@clinx)
  • client: Provide SDK for FlashNode object storage data block upload/download service （#3985,@bboyCH4,@clinx）
  • client: Implement NearRead strategy to prioritize reading from the nearest replica to reduce latency. (#3976,@zhumingze1108)

Enhance

  • client: Fuse library supports parallel processing of FUSE requests to improve concurrency. (#3974,@Victor1319)
  • client: Optimize metadata cache performance. (#3974,@Victor1319)
  • client: Add tcpAliveTime parameter for better TCP connection management. (#3974,@Victor1319)
  • master: Support DP decommission status evolution history query. (#3987,@shuqiang-zheng)
  • master: Add TryDecommissionRunningDiskIgnoreDps to support differentiated strategies for disk decommission based on different reasons. (#3975,@shuqiang-zheng)
  • master: Add audit logs for migrateMetaPartition and record reasons for DP migration/rollback. (#3975,@shuqiang-zheng)
  • datanode: Support reason passthrough for DP migration. (#3975,@shuqiang-zheng)
  • flashnode: Optimize cache operation opcodes and processing logic. (#3988,@clinx,@bboyCH4)

Bugfix

  • master: Decommission token consumed twice on restart during two-replica DP decommissioning(#3978,@shuqiang-zheng)
  • client: Fix ltp iogen01 test failure when pre-reading (ahead read) is enabled. (#3980,@clinx)
  • client: Offset calculation error during client readahead with partial hits(#3979,@bboyCH4)
  • master: Some DPs remain in decommission queue when disk offline marking fails, affecting subsequent decommissions (#3983,@shuqiang-zheng)
  • master: Incorrect disk/node decommission progress display for 2-replica DPs due to leader change(#3984,@Victor1319)
Source: README.md, updated 2025-12-22