Fluid is a CNCF open-source platform that gives big data and AI applications elastic data abstraction and acceleration on Kubernetes. Training and serving jobs often stall waiting on data: Fluid introduces a Dataset abstraction that co-locates cached data with the workloads that use it, orchestrating cache engines such as Alluxio, JuiceFS and Vineyard so datasets behave like a native, high-throughput layer of the cluster. The project has around 2,000 GitHub stars and is actively developed, with code pushed in 2026.
Instead of every training job pulling terabytes from remote object storage, Fluid lets platform teams declare a dataset once; Fluid provisions and manages the cache tier, prefetches data, and exposes it to pods through a POSIX or CSI interface with data-aware scheduling. It handles the full lifecycle: cache scaling, tiered storage across memory and SSD, data prefetching and migration, and observability of cache hit rates. For LLM-era workloads it is commonly used to accelerate model and dataset loading for training clusters.
Free. Fluid is open source under the Apache-2.0 license and runs on your own Kubernetes clusters.
Platform and MLOps engineers running training or inference on Kubernetes who lose GPU time to slow data loading and need a managed caching layer for datasets and models.
Fluid is a CNCF sandbox-turned-incubating project originally created by researchers and engineers working on large-scale training platforms in China, and it is used in production to serve dataset and model distribution for sizeable GPU clusters. Its controller watches Dataset custom resources and reconciles cache capacity automatically, and it exposes Prometheus metrics so platform teams can see whether slow training jobs are actually data-bound before buying more storage bandwidth.