# Storage Systems MOC > [!abstract] Thesis > Storage is not one decision, it is a set of contracts. Block, file and object each trade structure against scale. The right choice comes from **access pattern, growth and guarantees**, and the stack above (databases, observability, AI, sovereignty) inherits those choices. Part of [[085 Systems MOC]] · Method: [[Maps of Content (MOCs)]] ## Core note - [[Storage Systems]]: the three models, the seven forces, and the decision logic ## 1. Models and platforms **Block** - [[Longhorn]]: Kubernetes-native distributed block - [[Ceph]]: unified block, object and file from one substrate - [[OpenStack]]: Cinder (block) and Swift (object) as the cloud API over storage **Object** - [[MinIO]]: S3-compatible, pure object, high throughput - [[Ceph]]: RadosGW as the S3 face of the same cluster - [[data lakes]]: compute on top of cheap blob storage **File and high-performance** - [[Storage Solutions]]: NVMe vs spindle vs high-performance systems - [[Barcelona Supercomputing]]: parallel file systems at HPC scale ## 2. Speed, fabric and the memory hierarchy - [[NVMe Fabric]]: disaggregated storage over the network - [[SPDK]]: microsecond-latency storage building block - [[Cache Coherent Interconnect (CXL)]]: blurring memory and storage - [[Memory Bandwidth]] · [[HBM]] · [[GDDR]] · [[SRAM]] - [[KV Cache Compression and Eviction]] - [[Von Neumann Architecture]] - [[NIC - Network Interface Cards]] ## 3. Reliability, capacity and cost - [[Data Centre Redundancy]] - [[Data Center Tiers]] - [[Capacity Managament]] - [[Puzzle of low data center utilisation]] - [[Cloud Compute Cost Drivers]] ## 4. Orchestration and the cloud stack - [[Open Source Hyperscaler MoC]]: where storage sits as Layer 3 - [[Kubernetes]] - [[Bare Metal]] · [[VMs]] · [[Docker Containers]] - [[Clustering]] · [[Scheduling]] - [[multi-tenancy]] · [[Noisy Neighbour Problem]] - [[Convenience-Control Tradeoff]] - [[Apache CloudStack]] ## 5. Observability on object storage - [[Loki]] · [[Thanos]] · [[Mimir]] · [[Tempo]] · [[Jaeger]] · [[Prometheus]] ## 6. Data centre and infrastructure context - [[Data Center MoC]] - [[Modular Data Centers MoC]] - [[The Data Center is the Computer - How the Network Shapes Performance]] - [[Next Generation Data Center]] - [[Kinds of Data Centers]] - [[Neoclouds]] - [[AI Inference Infrastructure]] - [[Power Consumption x Scalability]] ## 7. Sovereignty, security and Navon - [[Navon MoC]] - [[Navon Thesis]] - [[Navon Sovereign Vaults]] - [[Data Embassies]] - [[Compute to data]] - [[Confidential Computing]] - [[Decentralized Compute]] - [[Sovereign AI Positioning]] - [[Hybrid Cryptography as the bridge to Post Quantum Security]] - [[Cryptography MOC]] · [[Quantum Security MOC]] · [[Security MoC]] - [[Nirvana_Africa_Sovereign_Storage_Strategy.docx]] ## 8. Storage for AI and agents - [[AI Agents Stack]] - [[Persistent Memory]] - [[Context Layers MOC]] - [[The AI Stack - Building Blocks]] - [[System Design in the Age of AI]] - [[Latest Stack]] ## 9. Thinking tools - [[First Principles and Mental Models MoC]] - [[Concepts MOC]] ## Arrows of progress - **Disks in servers → disaggregated pools** joined by NVMe fabric - **One tool per model → one substrate, three interfaces** (Ceph) - **Retention priced by compute → retention priced by object storage** (Loki, Thanos, Mimir) - **Location of data → jurisdiction and key custody of data** (Sovereign Vaults) - **Retrieval-only agent memory → persistent, stateful memory** ## Open questions - Where does a distributed file system end and object storage with a POSIX gateway begin for AI training? - How much of data sovereignty is storage design versus key custody? - Does CXL make the block / memory boundary irrelevant for inference? #moc #storage