JuiceFS - IT Press Tour #68 June 2026

Transcript

1. The Distributed File System for AI Built on Object Storage Joe Zhou | DevRel @JuiceFS
2. Agenda ➔ Object Storage | The New Backend ➔ File System on Object Storage | But Why? ➔ S3 Files | A Decent Approach ➔ JuiceFS | The Distributed File System for AI
3. Object Storage | The New Backend
4. Object Storage | The New Backend ➔ “A giant digital attic where you toss everything.” ➔ “A place to dump data that doesn’t belong to a database.”
5. Object Storage | Advantages ➔ Simple access protocol, “basic” but powerful primitives ➔ Flat namespace ➔ Unlimited scalability ➔ Extreme durability ➔ Multi-region availability ➔ Immutability and versioning ➔ Ideal for unstructured data ➔ Cost efficiency (in most cases)
6. Object Storage | Diskless Architecture
7. Object Storage | CAS Example
8. Object Storage | Disadvantages ➔ Lack of rich API, operations are too basic ➔ No in-place updates ➔ Metadata support is available, but batch operations are costly ➔ No traditional file hierarchy ➔ Higher latency ➔ No direct execution ➔ Not ideal for structured data
9. File System on Object Storage
10. File System on Object Storage | Why? ➔ Lack of rich API, operations are too basic ➔ No in-place updates ➔ Metadata support is available, but batch operations are costly ➔ No traditional file hierarchy ➔ Higher latency ➔ No direct execution ➔ Not ideal for structured data
11. File System on Object Storage | Why? ➔ ✅ Rich API (POSIX) ➔ ✅ In-place updates (`sed`) ➔ ✅ Metadata & batch operations (`mv`) ➔ ✅ File hierarchy ➔ ✅ Low latency with multi-tier cache ➔ ✅ Application execution ➔ ✅ Operate directly on structured data
12. S3 Files | A Decent Approach
13. S3 Files | What It Is ➔ POSIX-compatible distributed file system on top of EFS & S3 ➔ Mount your S3 bucket as a volume ➔ S3 Files maintains a 1-to-1 mapping between files and objects ➔ Key idea: ◆ EFS -> the high-performance caching layer/tier for S3 ◆ S3 -> the source of truth backing storage
14. Reference: docs.aws.amazon.com/AmazonS3/latest/userguide/s3-files.html
15. S3 Files | How It Works ➔ Data Synchronization ◆ `prefix` - entire bucket or prefix ◆ `trigger` - when to sync files ◆ `sizeLessThan` - default to 128 KiB ➔ Read Path ➔ Write Path
16. S3 Files | Pros ➔ Minimal Data Migration ◆ Easy to get started with (if you are already using AWS) ➔ Standard POSIX API ◆ NFS protocol, file locks, reads/writes ➔ Unified Storage & Concurrent Access ◆ S3 is the source of truth ◆ Unlike EBS, concurrent access is allowed ➔ Based on Mature and Reliable Services
17. S3 Files | Cons ➔ Optimized for Small Files ➔ Write Amplification ➔ Costly Metadata Operations ➔ Batching Delays ➔ Conflict Resolution ➔ Not a Multi-Cloud Solution ➔ Complex Pricing ◆ Storage tier, reads, writes, syncs
18. S3 Files | Cons - Write Amplification
19. S3 Files | Cons - Write Amplification
20. S3 Files | Cons - Costly Metadata Operations $> mv /foo /bar File System Object Storage
21. S3 Files | Cons - Complex Pricing ➔ Writes land in the EFS high-performance tier first ◆ Writes are charged ◆ Additional EFS storage costs ➔ Then writes are aggregated and synced back to S3 ◆ Reads from EFS are charged again ◆ Eventual S3 storage and data access costs
22. S3 Files | Does It Solve the Problems? ➔ ✅ Rich API (POSIX) ➔ ❌ In-place updates (depending on file size) ➔ ❌ Metadata & batch operations (massive object rewrites) ➔ ✅ File hierarchy ➔ 🆗 Low latency with multi-tier cache ➔ ❌ Application execution (file size, object rewrites) ➔ ❌ Operate directly on structured data (file size, object rewrites)
23. JuiceFS | Distributed File System for AI
24. JuiceFS Architecture | Community Edition
25. JuiceFS Architecture | Enterprise Edition
26. JuiceFS | Data Chunking
27. JuiceFS | Data-Metadata Separation $> mv /foo /bar File System Object Storage
28. JuiceFS CE | Cost Efficiency $> juicefs config {METADATA-ENGINE} --tier-id 1 --tier-sc STANDARD_IA -y $> juicefs config {METADATA-ENGINE} --tier-id 2 --tier-sc INTELLIGENT_TIERING -y $> juicefs config {METADATA-ENGINE} --tier-id 3 --tier-sc GLACIER_IR -y
29. JuiceFS EE | 500 Billion Files
30. JuiceFS EE | Mirror File System (Cache-Only)
31. JuiceFS EE | Mirror File System (Full Mirror)
32. ➔ MiniMax is a leading general AI technology company. ➔ MiniMax employs a hybrid-cloud architecture. ➔ Why hybrid-cloud? ◆ Flexibility and cost efficiency. ◆ Critical workloads (i.e., GPUs) are deployed in IDC. ◆ Other components were deployed in the cloud. JuiceFS | Who Uses Mirror File Systems?
33. MiniMax | Cache-Only Mirrors
34. MiniMax | Considering Full Mirrors Already
35. JuiceFS vs. S3 Files
36. JuiceFS | Does It Solve the Problems? ➔ ✅ Rich API (POSIX) ➔ ✅ In-place updates (`sed`) ➔ ✅ Metadata & batch operations (`mv`) ➔ ✅ File hierarchy ➔ ✅ Low latency with multi-tier cache ➔ ✅ Application execution ➔ ✅ Operate directly on structured data
37. JuiceFS Adopters & Enterprise Customers PixVerse
38. Thanks Website-> https://juicefs.com GitHub -> https://github.com/juicedata/juicefs