Release Notes

Alluxio Enterprise AI 3.8

What is New

High-Performance Write Caching for S3 API

Experimental

This release introduces High-Performance Write Caching for S3 Put APIs in Alluxio, enabling accelerated object write performance for S3-compatible storage. Previously, Alluxio delivered significant read performance gains, but write operations were limited by underlying object storage. With this feature, Alluxio now provides write acceleration for S3 workloads, delivering lower latency, higher throughput, and scalable performance.

Use Case

This feature is designed for workloads that perform high-volume object writes to S3-compatible storage, such as:

• AI/ML training pipelines writing intermediate datasets

• Media and content platforms ingesting large volumes of objects

• High-frequency data ingestion and streaming pipelines

Users can deploy Alluxio as a write cache layer to dramatically improve write performance while maintaining S3 compatibility.

Key Benefits

• 5-10x Write Latency Improvement: Reduces S3 Put latency from ~50 ms to <10 ms.

• Write Throughput Improvement: Achieves up to 6 GB/s per worker, with linear scalability across workers

Limitations

• Multipart Upload (MPU) APIs write caching are not supported in this release.

Refer to S3 Write Cache on how to enable and configure these capabilities.

Optimized .safetensors Model Loading

This release introduces safetensor-aware model loading optimization, significantly reducing cold-start times for large-scale model inference workloads. By understanding safetensor metadata and access patterns, Alluxio merges thousands of small random reads into large sequential reads, delivering near–local-disk performance for model loading. This enables faster model initialization, quicker deployment cycles, and more responsive inference systems.

Use Case

This feature is designed for AI and ML inference platforms that frequently load large models in safetensors format, including:

• Model-as-a-Service (MaaS) platforms deploying and scaling models dynamically

• Enterprises running self-managed model serving pipelines

• A/B testing and model iteration workflows that require frequent model reloads

• Large-scale inference clusters where cold-start latency impacts user experience and SLA

By accelerating safetensor model loading, teams can reduce startup delays and improve operational agility.

Key Benefits

• Dramatically Faster Cold Starts: Reduces model loading time from minutes to seconds.

• Near Local-Disk Performance: Achieves within ~10% of NVMe local disk speed for large models.

• Massive Performance Gains vs Network Storage: Up to 18× faster than AWS FSx Lustre.

Limitations

• Supported only through the Alluxio FUSE interface.

Refer to Optimizing AI Model Loading on how to enable and configure these capabilities.

Job Service (Coordinator) High-Availability

This feature is designed for production environments that rely on Alluxio Job Service to manage data loading, eviction, and lifecycle workflows, including:

• AI and ML pipelines that pre-load or free datasets dynamically

• Enterprises deployments requiring high availability and zero-downtime operations

Key Benefits

• Eliminates Single Point of Failure: Multiple (N) coordinators ensure the Job Service remains available even if N-1 coordinators fail.

• High Availability & Resilience: Continuous job submission and scheduling during failures and maintenance windows.

• Scalable Job Throughput: Supports 100+ job submissions per second for hours.

• Real-Time Job Visibility: Provides real-time job status monitoring for operational transparency.

• Enterprise-Ready Reliability: Designed for production-scale deployments with strict uptime requirements.

Refer to Managing Coordinators on how to enable and configure these capabilities.