AI/ML Storage Design

This skill helps design storage architecture for AI and ML workloads on Google Cloud, covering both Training and Serving (Inference) phases.

0. Critical First Step: Check Existing Infrastructure

Before proposing new infrastructure, ALWAYS ask the user if they have existing storage resources. This prevents redundancy and leveraged existing investments.

Ask:

• "Do you already have a Managed Lustre instance running for your training pipeline?"
• "Are your datasets or models already stored in Cloud Storage buckets?"

List Lustre instances:

gcloud lustre instances list --location=<REGION>

Decision Rule:

• If Managed Lustre exists: Strongly consider using it for both Training and Serving (especially for single-zone serving) to maintain consistency and performance.
• If GCS is the only existing storage: GCS + FUSE is the default starting point unless performance requirements force a move to Lustre.

---

1. Storage Selection: Training

Training involves two main activities: Data Loading and Checkpointing.

A. Data Loading (Reading Training Data)

Feature	Cloud Storage (GCS) + FUSE	Managed Lustre
Best For	Large files (≥ 50 MB), Cost efficiency, Durability priority.	Small files (< 50 MB), Max performance, POSIX compliance.
Latency	Tens of milliseconds.	Ultra-low (< 1 ms).
Throughput	High (with FUSE + Anywhere Cache).	Extremely high.

Recommendation:

• Use Managed Lustre if: Files are small (< 50 MB), strict POSIX compliance is needed, or latency must be sub-millisecond.
• Use GCS + FUSE if: Files are large (≥ 50 MB), or you prioritize durability/HA over raw random I/O speed.

B. Checkpointing (Saving State)

Recommendation:

• Use Managed Lustre if: You are already using it for data loading, or require frequent, high-performance checkpoints to minimize training pauses.
• Use GCS if: You are using GCS FUSE for data loading. Use Hierarchical Namespaces on the bucket to speed up atomic renames and checkpoint saves.

---

2. Storage Selection: Serving (Inference)

When moving to production serving, the choice depends on your deployment architecture.

A. Cloud Storage (GCS)

Best for: Scalable, multi-region serving with dynamic fleets.

Recommendation: Use GCS + Cloud Storage FUSE + Anywhere Cache.

Select if ANY of the following apply:

• Dynamic Scalability: The number of inference nodes changes dynamically (autoscaling).
• Infrequent Updates: Model files are updated infrequently.
• Multi-Region/Zone: Serving models from multiple zones or regions.
• High Availability: Priority is durability and availability against regional/zonal failures.

Key Optimization:

• Cloud Storage FUSE: Enable parallel downloads for faster model loading.
• Anywhere Cache: Essential for high throughput (>1 TB/s) or large fleets (>100 nodes).

B. Managed Lustre

Best for: High-performance, single-zone, low-latency serving.

Select if ANY of the following apply:

• Existing Workflow: Training/checkpointing already uses Managed Lustre (avoids data movement).
• Single Zone: Serving is restricted to a single zone.
• Performance Sensitive: Requires consistent, ultra-low latency (<1ms).
• Frequent Updates: Model files are updated frequently.

C. Implementation (GKE)

When using Managed Lustre on GKE, use the Lustre CSI driver.

PV Example:

csi:
  driver: lustre.csi.storage.gke.io
  volumeHandle: <PROJECT>/<LOCATION>/<INSTANCE>
  volumeAttributes:
    filesystem: "models"
    ip: "10.x.x.x"

---

Discovery Questions

If requirements are unclear, ask these questions (after checking for existing infra):

1. **"What is the typical file size of your training data? Are they many small files (<50MB) or fewer large files?"** (Small -> Lustre; Large -> GCS)
2. "Are you serving the model in a single zone, or across multiple regions?" (Multi-region -> GCS)
3. "Do you need to scale the number of inference nodes dynamically?" (Yes -> GCS)
4. "How strict are your latency requirements for data loading? Do you need sub-millisecond access?" (Yes -> Lustre)

Optimization Tips

• Training: For GCS, always use Cloud Storage FUSE. For high throughput requirements, enable Anywhere Cache.
• Model Load Time: Optimize for rapid model loading to minimize accelerator (GPU/TPU) idle time.
• Archive: Use GCS Autoclass or Lifecycle Management to automatically transition older training data/models to cheaper storage classes.