Release Notes
Last updated
Last updated
AI model distribution is a process of deploying and managing trained machine learning models across various environments on-premise or on-cloud to enable model inference. The efficiency of this process is critical for businesses to rapidly utilize newly trained models.
In this release, Alluxio boosts the model distribution workloads, providing up to 3x faster throughput.
By equipping client side memory pools, the Alluxio FUSE client can utilize the client node's memory capacity to cache models and reuse the cached models to quickly fulfill application’s model loading requests. As the benchmark shows, within a 100Gbps network environment, Alluxio normally can offer up to 9.3 GiB/s reading throughput when distributing a single model across multiple GPUs on the same node. With this enhancement enabled, Alluxio can offer up to 30GiB/s read throughput, which is 3x faster than before, exceeding network limits.
Checkpoints in model training are critical for saving progress to prevent data loss from interruptions and allow resuming when validating models. A slow checkpointing step can dramatically increase the overall time of the training process.
In the previous release, checkpoint writing through Alluxio was limited by under storage performance.
In this release, by introducing the new ASYNC write mode, Alluxio provides up to 8GiB/s write throughput on a 100Gbps network environment to drastically shorten the model training checkpoint process. It is also able to persist checkpoints to the under storage asynchronously.
By acting as a high-performance caching and acceleration layer, Alluxio achieves sub-millisecond Time-to-First-Byte (TTFB) performance, comparable to AWS S3 Express One Zone, without requiring specialized hardware, data format changes, or data migration.
Key Use Case:
Accelerating queries on petabyte to exabyte scale data lakes stored on cloud object storage (e.g., AWS S3)
Improving performance for latency-sensitive AI applications like Retrieval-Augmented Generation (RAG) and AI feature store workloads.
Key Benefits:
1000x Performance Improvement: Achieves ultra-low-latency point queries on Parquet files, reducing query latencies to hundreds of microseconds per query and 3,000 queries per second on a single thread.
Linear Scalability: Throughput scales linearly with cluster size, enabling a 50-node deployment to achieve one million queries per second, surpassing S3 Express throughput by 50x without latency degradation.
Sub-Millisecond Latency: Delivers sub-millisecond TTFB performance, matching AWS S3 Express One Zone. Employs optimizations like eager data caching with zero-copy transfer, predicate and projection pushdown, single RPC execution path, and metadata caching.
Efficient Parquet File Handling: Offloads partial Parquet read operations from query engines, reducing overhead associated with file parsing and index lookups.
Cost-Effective Solution: Provides high performance without the need for expensive solutions like S3 Express, offering significant cost savings.
Seamless Integration: Works with existing applications without requiring changes to data formats or query logic.
Deployment in multiple Availability Zones (Multi-AZ) within the same region is a High Availability strategy provided by cloud vendors to utilize redundant architecture to mitigate risks from single AZ failures.
In this release, Alluxio naturally integrates with Cloud Multi-AZ strategy to offer high availability and high performance of data access and ensure stronger data access resilience at all times. To make this happen, per user definitions, Alluxio can distribute the data replications across multiple AZ. When an application is accessing data, Alluxio client can intelligently detect and direct the access to proper data replications by following the priority of same AZ’s replications and then other AZ’s replications. As long as there is one data replication alive across all AZs, Alluxio can offer resilient and high performance data access. Furthermore, Alluxio’s passive caching feature can automatically recover the lost replications so that Alluxio can always keep a good number of data replications for the high availability data access.
Please refer to Multiple Availability Zones (AZ) for more details
Alluxio now supports flexible multi-tenancy through a new Gateway component integrated with Open Policy Agent (OPA). This enables fine-grained, team-based access control over the management console and Alluxio REST APIs, tailored to your organization’s structure.
Key Capabilities:
Integrates with OPA to enforce custom role-based access control (RBAC) policies.
Supports arbitrary role definitions, with full control over path-based authorization rules.
Works seamlessly across Alluxio’s Management Console and Alluxio APIs.
Example Roles Template:
Super Admin: Full visibility and control over all resources.
Team Admin: Can view the entire namespace; access rights enforced per path.
User: Access restricted; Management Console visibility completely disabled.
This model provides maximum flexibility for enterprise-scale deployments with diverse user groups and security boundaries. A default policy template is included to help you get started.
We’ve released a new Alluxio Management Console designed to improve operational efficiency and support production-scale deployments. This web based new management console introduces a modular navigation structure aligned with key workflows: Storage mounting, Cache lifecycle management (ex. preloading and eviction), and fine-grained system configuration (e.g., TTL, quota, eviction priority).
Built with scalability and extensibility in mind, the Management Console sets the foundation for UI-driven cluster operations, automation workflows, and future enterprise control surfaces.
Alluxio provides an ability to allow users to define virtual paths to access data. This could be useful to abstract or hide the underlying data paths.
In prior versions, the Alluxio java source code was compiled to bytecode compatible with Java 8 or higher. This is updated to be compatible with Java 11 or higher. For bare metal deployments that provide their own Java environments, Alluxio 3.6 will no longer be able to run with a Java 8 JVM and must run with Java 11 JVM or newer. Containerized deployments, such as those deployed on Kubernetes, are not affected.
Major third party libraries are upgraded to their latest versions to mitigate potential security vulnerabilities surfaced by CVE scans. Some major upgrades include:
As a result, all CVEs of high or critical severity levels are resolved.
Please refer to for more details.
Please refer to for more details.
Please refer to for Management Console for more details.
Please refer to for more details.
is upgraded from 1.60.1 to 1.68.1
is upgraded from 3.3.6 to 3.4.1
is upgraded from 2.13.5 to 2.15.0
is upgraded from 0.7.5 to 0.8.5
is upgraded from 4.1.108.Final to 4.1.118.Final
is upgraded from 3.19.6 to 3.25.5