Storage Backends

AEGIS orchestrators support mounting a diverse array of storage backends directly into Agent containers while maintaining strict Phase 1 zero-privilege security guarantees.

This is achieved via the AegisFSAL (File System Abstraction Layer) which dynamically proxies NFS file operations (like read, write, getattr) to the appropriate underlying storage backend instance associated with the volume metadata.

1. SeaweedFS (Default)

SeaweedFS is the standard, distributed storage backend for AEGIS. It provides high-performance, clustered POSIX-like storage.

Mechanism

• Provider: SeaweedFSAdapter
• NFS Proxying: Requests are translated by the orchestrator into HTTP operations hitting the SeaweedFS Filer API (usually port 8888).
• Use Case: General purpose ephemeral scratch space, and standard persistent volumes.

Agent Manifest Example

By default, volumes use SeaweedFS. You can map them explicitly:

volumes:
  - name: shared-data
    type: seaweedfs
    storage_class: persistent
    mount_path: /workspace/shared-data
    access_mode: read-write

2. OpenDAL (Cloud APIs)

OpenDAL allows agents to mount broad cloud storage APIs natively (such as S3, GCS, Azure Blob, WebDAV) as if they were local POSIX filesystems.

Mechanism

• Provider: OpenDalStorageProvider
• Security: The agent container requires zero credentials. The Orchestrator resolves credentials (via OpenBao integration) and instantiates the opendal::Operator. All API translation and authentication are handled transparently by the orchestrator host proxying I/O over the NFS interface.
• Use Case: Direct loading of massive ML datasets from S3 without copying into SeaweedFS first; writing model checkpoints directly to cloud buckets.

Agent Manifest Example

volumes:
  - name: datasets
    type: opendal
    provider: s3
    config:
      bucket: my-ml-datasets
      endpoint: "https://s3.us-east-1.amazonaws.com"
      access_key_id: "secret:aws/s3-reader/access-key" # Vault resolved
      secret_access_key: "secret:aws/s3-reader/secret-key"
    storage_class: persistent
    mount_path: /workspace/datasets
    access_mode: read-only

3. Local Host Mounts

Often, an orchestrator operates on hardware with fast local NVMe drives or pre-downloaded shared weights. LocalHost proxies allow agents to mount a directory physically present on the orchestrator host.

Mechanism

• Provider: LocalHostStorageProvider
• Security: Bound by AegisFSAL's strict PathSanitizer. Agent NFS operations cannot perform path traversal escapes (../) out of the defined mount_path directory on the host.
• Use Case: Ultra-fast LLM weight loading, NVMe caching directories, or shared model caches pre-downloaded to the physical machine.

Agent Manifest Example

volumes:
  - name: local-weights
    type: hostPath
    config:
      path: /mnt/fast-nvme/llama-weights
    storage_class: persistent
    mount_path: /workspace/weights
    access_mode: read-only

4. SEAL Remote Volumes

Signed Envelope Attestation Layer (SEAL) provides secure, identity-verified node-to-node communication. For storage, it allows an agent running on Node A to interact with a volume physically managed by Node B.

Mechanism

• Provider: SealStorageProvider
• Security: POSIX operations (e.g., read_at) are wrapped into signed SealEnvelope API requests. The receiving orchestrator (Node B) verifies Node A's cryptographic Identity and Cedar policies before applying the read to its own local AegisFSAL layers.
• Use Case: Swarm coordination, cross-node pipeline execution without relying on a centralized storage backbone.

Agent Manifest Example

volumes:
  - name: remote-pipeline-data
    type: seal
    config:
      node_id: "node-b-uuid-1234"
      remote_volume_id: "vol-9876-abc"
    storage_class: persistent
    mount_path: /workspace/pipeline/input
    access_mode: read-only

All mount paths must be rooted at /workspace.

Security Model Overview

No matter which backend an agent request is routed to, it must pass through AegisFSAL first. This guarantees:

1. UID/GID Squashing: Cloud storage permissions and host root permissions do not leak into the Agent container. Ownership appears mapped to the Agent.
2. Quota Enforcement: OpenDAL S3 or HostPath directories are still subject strictly to the AEGIS volume size_limit, ensuring agents cannot fill up orchestrator host disks or bankrupt cloud accounts.
3. Audit Trail: Every request to every backend emits unified StorageEvent tracking for full observability.
4. No FUSE: Agents never need CAP_SYS_ADMIN capability.