Skip to content

ROADMAP.md

Latest commit

 

History

History
78 lines (47 loc) · 3.6 KB

ROADMAP.md

File metadata and controls

78 lines (47 loc) · 3.6 KB

Fluid Roadmap

Fluid 2026 Roadmap

1. Data Anyway

Objective: Enable fluid data access regardless of infrastructure constraints (e.g., storage types, runtime environments) without developing controller code.

Generic Cache Runtime

  • Pluggable Architecture: Standardized Cache Runtime Interface for rapid integration of new engines (CubeFS, Dragonfly, Vineyard) with minimal boilerplate.
  • Orchestration Based on AdvancedStatefulSet: Migrate from StatefulSet to AdvancedStatefulSet for fine-grained Pod lifecycle management, ordered rollout, and enhanced failover capabilities.

Runtime Dynamic Configuration

  • Zero-Downtime Tuning: Adjust cache replicas, storage media tiers (SSD/HDD/RAM), and eviction policies without Dataset reconstruction or workload restart.
  • Hot Parameter Swapping: Runtime modification of cache engine configurations (e.g., Alluxio thread pool, Jindo worker threads) for traffic spike handling.

API Upgrade to v1alpha2

  • Standardized Conditions, ObservedGeneration, and phase transition semantics for improved GitOps and tooling compatibility.
  • Conversion webhook support for seamless v1alpha1 → v1alpha2 migration.

Validation Webhook

  • Admission-time CRD validation with auto-correction suggestions to prevent misconfigurations.
  • Policy enforcement for resource quotas and security constraints.

ThinRuntime Productization

  • Production-ready stability for large-scale deployments with minimum container privileges (eliminate privileged FUSE Pod requirements).

2. Data Anywhere

Objective: Achieve cross-region, cross-cluster, and cross-platform data mobility and accessibility.

LLM KV Cache Orchestration

  • Disaggregated KV Cache: Externalize vLLM/SGLang KV Cache to Fluid-managed distributed storage, enabling 10x+ throughput improvement for long-context inference.
  • Cross-Pod Cache Sharing: Live migration of KV Cache between inference instances for preemptive scheduling and spot instance tolerance.
  • Mooncake Integration: Official partnership for high-performance KV Cache backend with RDMA acceleration.

Efficient Data Prewarming & Migration

  • Distributed Prewarming: Maximize bandwidth utilization for fast data loading.
  • Throttling Control: Limit bandwidth usage during prewarming to avoid saturation.
  • Rsync Optimization: Improve cross-region sync efficiency.

JindoRuntime High Availability

  • Master Pod Crash Recovery: Automatic re-setup and state reconstruction after cache master failure without data loss.
  • Metadata Persistence: WAL-based metadata recovery for rapid failover.

Observability-Driven Optimization

  • Access Pattern Recognition: ML-based analysis to auto-inject acceleration strategies (prefetching, block size optimization).
  • Dataset Garbage Collection: Idle dataset detection via reference counting and access history analysis.

3. Data Anytime

Objective: Ensure real-time, adaptive, and intelligent data availability for workloads.

Temporal Workflow Integration

  • Kueue-Driven Pipelines: Trigger training/inference jobs automatically upon DataLoad completion; automate post-job cache eviction and data migration.
  • Event-Driven Policies: Flexible metadata synchronization triggered by workload lifecycle events.

Developer Experience

  • Fluid kubectl Plugin: Native CLI extension (kubectl fluid) for:
    • Dataset status inspection and health diagnostics
    • On-demand prewarming triggering (kubectl fluid warmup)
    • Cache performance profiling and bottleneck analysis
    • Runtime configuration hot-updates