Storage Systems Overview

Introduction

Storage systems are fundamental to computer science and systems design. Understanding storage technologies, trade-offs, and architectures is critical for placement interviews, especially for roles in systems engineering, cloud infrastructure, and backend development.

The Storage Hierarchy

graph TD
    A[CPU Registers] --> B[L1/L2/L3 Cache]
    B --> C[Main Memory - RAM]
    C --> D[SSD / NVMe]
    D --> E[HDD - Magnetic Disk]
    E --> F[Object Storage - S3/GCS]
    F --> G[Tape / Cold Archive]

    style A fill:#ff6b6b,color:#fff
    style B fill:#ffa94d,color:#fff
    style C fill:#ffd43b,color:#000
    style D fill:#69db7c,color:#000
    style E fill:#4dabf7,color:#fff
    style F fill:#9775fa,color:#fff
    style G fill:#868e96,color:#fff

Key Concepts for Interviews

1. IOPS vs Throughput vs Latency

• IOPS (Input/Output Operations Per Second): How many read/write operations per second. Critical for transactional workloads (databases).
• Throughput (MB/s or GB/s): Data transfer rate. Critical for sequential workloads (video streaming, big data).
• Latency: Time for a single I/O operation. Critical for interactive applications.

graph LR
    A[Workload Type] --> B{Random or Sequential?}
    B -->|Random| C[High IOPS matters]
    B -->|Sequential| D[High Throughput matters]
    C --> E[SSD/NVMe preferred]
    D --> F[HDD can suffice]

2. Read/Write Patterns

• Sequential: Data accessed in order (log files, video). HDDs perform well.
• Random: Data accessed at arbitrary positions (databases, metadata). SSDs dominate.
• Write-Once-Read-Many (WORM): Ideal for logs, compliance data.
• Read-Heavy vs Write-Heavy: Impacts wear leveling (SSD) and caching strategies.

3. Durability vs Availability

• Durability: Data won’t be lost (measured in nines: 99.999999999% for S3).
• Availability: Data is accessible when needed (measured as uptime percentage).
• They are independent: You can have highly available but not durable storage (RAM cache), or highly durable but not always available (tape backup).

4. CAP Theorem in Storage

graph TD
    A[CAP Theorem] --> B[Consistency]
    A --> C[Availability]
    A --> D[Partition Tolerance]
    B --> E[All nodes see same data]
    C --> F[Every request gets a response]
    D --> G[System works despite network failures]

Distributed storage systems must choose between CP (consistent but may be unavailable) and AP (available but may return stale data).

5. RAID (Redundant Array of Independent Disks)

Storage Architecture Patterns

DAS vs NAS vs SAN

graph TD
    subgraph DAS[Direct Attached Storage]
        S1[Server] -->|Direct cable| D1[Disk]
    end
    subgraph NAS[Network Attached Storage]
        S2[Server] -->|NFS/SMB| FS[File Server] --> D2[Disks]
    end
    subgraph SAN[Storage Area Network]
        S3[Server] -->|Fibre Channel/iSCSI| SAN_SW[Switch] --> D3[Block Storage]
    end

• DAS: Simplest. Direct connection. Limited scalability.
• NAS: File-level access over network. Good for shared file systems.
• SAN: Block-level access over dedicated network. High performance, complex.

Interview Questions

1. Q: When would you choose HDD over SSD? A: For archival storage, sequential-heavy workloads (streaming), or when cost per TB is the primary concern. HDDs offer 3-5× lower cost per TB.

2. Q: Explain the difference between block storage, file storage, and object storage. A: Block storage provides raw disk blocks (like a virtual hard drive). File storage provides a hierarchical file system (NFS, SMB). Object storage provides a flat namespace with HTTP APIs and metadata (S3). Each suits different access patterns.

3. Q: What is write amplification in SSDs? A: Due to erase-before-write semantics, SSDs may need to write more data than requested. A 4 KB write might trigger a 256 KB erase+write cycle, reducing SSD lifespan and performance.

4. Q: How does RAID 5 differ from RAID 6? A: RAID 5 uses single distributed parity (tolerates 1 disk failure). RAID 6 uses double parity (tolerates 2 disk failures). RAID 6 has higher write overhead but better fault tolerance.

5. Q: What is eventual consistency in distributed storage? A: After a write, reads may return stale data temporarily, but eventually all replicas converge to the same value. Used by DynamoDB, Cassandra for high availability.

Common Mistakes

• Confusing throughput with IOPS — a system can have high throughput but low IOPS (sequential HDD) or vice versa (random SSD).
• Ignoring write amplification in SSD-heavy systems.
• Assuming RAID replaces backups — RAID protects against disk failure, not accidental deletion, corruption, or disasters.
• Overlooking tail latency (P99) — averages hide worst-case behavior that affects user experience.
• Not considering data locality in distributed systems — moving computation to data is cheaper than moving data to computation.

Summary

Storage systems span a wide spectrum from nanosecond-level registers to petabyte-scale object stores. The key is matching storage technology to workload characteristics: random vs sequential, read vs write, latency vs throughput, hot vs cold data. For interviews, understand the trade-offs between HDD/SSD/NVMe, the three storage types (block/file/object), RAID levels, and distributed storage consistency models.

Cross-References

• HDD Deep Dive — Magnetic disk internals
• SSD Deep Dive — Flash memory and wear leveling
• NVMe — Modern high-performance storage
• Object Storage — S3, GCS, and blob storage
• Block Storage — EBS, Cinder, raw volumes
• File Storage — NFS, HDFS, distributed file systems
• Distributed Storage — Consensus and replication
• Ceph — Unified distributed storage
• Erasure Coding — Space-efficient redundancy
• Cloud Overview
• Interview System Design
• DBMS Indexing
• Arch Memory Hierarchy