Reading · The Architect’s Notebook

By Amit Raghuvanshi | The Architect’s Notebook

🗓️ Jul 30, 2026 · Deep Dive ·

This post is the sixth installment in our deep-dive series on Amazon S3, a journey we started last year to understand the architecture behind one of the most influential distributed storage systems ever built.

So far, we have explored the original vision behind S3, its core architecture, request flow, partitioning and auto-scaling mechanisms, storage internals, REST APIs, authentication, data distribution strategies, parallel reads, erasure coding, durability models, metadata architecture, and the operational components that keep the system running at massive scale. Along the way, we have uncovered how S3 balances performance, availability, durability, and cost while serving trillions of objects worldwide.

In Part 6, we shift our focus from internals to real-world applications. We will explore how S3 powers global content delivery through CloudFront, serves as the foundation of modern data lakes, enables enterprise-grade backup and disaster recovery strategies, and acts as scalable application storage. More importantly, we will connect these use cases back to the architectural principles we have studied throughout this series.

For readers joining us midway through the series, here are the links to all the previous parts, providing a complete walkthrough of Amazon S3’s architecture and design.

S3’s architecture supports advanced features through its control plane (for configuration) and background operations (for automation like replication and transitions). These leverage S3’s distributed storage for efficiency.

What it involves: Turn a bucket into a web server for static content (HTML, CSS, JS, images). No server-side code, perfect for portfolios or docs. S3 serves files via HTTP/HTTPS (with CloudFront for HTTPS).
Tie to architecture: Enabling hosting configures the bucket’s metadata in the control plane. Requests route through the API gateway, but website endpoints bypass some REST API limits (e.g., supports index/error docs). Content is served from edge-cached storage nodes for low latency.
Why it matters: Cheap (pay only for storage/requests) and scalable, handles unlimited traffic via S3’s global infrastructure.

Real-World Example: Host a personal blog. Upload index.html (home page), styles.css, and blog-post.html. When users visit the endpoint, S3 serves index.html automatically. If they go to /blog-post.html, it loads directly. Background integrity checks ensure files aren’t corrupted. Costs: ~$0.50/month for 1GB + low traffic.
For HTTPS/custom domain, add CloudFront (CDN) and Route 53 (DNS)—S3 alone is HTTP-only.

What it involves: Automatically copy objects (and metadata/tags) from a source bucket to a destination bucket in another region. Supports versioning and encryption.
Tie to architecture: Uses S3’s background operations for asynchronous replication across regions (beyond AZs). The control plane manages rules; storage nodes handle the copy. Metrics track progress via CloudWatch.
Why it matters: Enhances disaster recovery, compliance (e.g., store data in specific regions), and low-latency access (e.g., replicate to regions near users).
Real-World Example: A global e-commerce site stores product images in a US East bucket. Use CRR to replicate to EU (Ireland) for EU users (low latency) and compliance (GDPR data residency). Prefix filter "images/" replicates only images. If US East fails, EU bucket serves seamlessly. Costs: Storage in both regions + inter-region transfer (~$0.02/GB out).

What it involves: Rules to automate object management, transition to cheaper storage classes, expire (delete), or abort incomplete uploads. Applies to current/noncurrent versions.
Tie to architecture: Background operations scan objects daily, using metadata (e.g., creation date) from the control plane. Transitions use erasure coding/replication for efficiency; deletions are permanent after grace periods.
Why it matters: Saves costs (e.g., move old data to Glacier) and enforces retention (e.g., delete logs after 90 days).
Real-World Example: A logging app uploads daily logs to logs-bucket. Policy transitions logs >30 days to IA (cheaper for infrequent access), >90 days to Glacier (archive), and deletes after 365 days (retention policy). Background scans detect eligibility; S3 handles transitions transparently. Costs drop 70%+ for old logs without manual intervention.

Scenario: A media company serves millions of video files to users worldwide, needing fast, secure, and scalable delivery.

How S3 Fits: S3 stores video files in buckets, with data distributed across multiple AZs for durability and parallel reads for performance. It integrates with Amazon CloudFront, a Content Delivery Network (CDN), to cache content at edge locations worldwide, reducing latency. S3’s metadata (e.g., content type, ETag) ensures efficient serving, and signed URLs provide secure, temporary access to premium content.

Implementation Details:

Store Video Files: Upload videos to an S3 bucket with logical folder structures (e.g., videos/premium/movie.mp4). S3 splits large files into chunks, distributes them across storage nodes, and uses erasure coding (for cost-efficient classes) or replication for durability.
CloudFront Integration: Configure CloudFront to use the S3 bucket as the origin. CloudFront caches videos at edge locations (e.g., in Tokyo for Asian users), fetching from S3 only when needed. S3’s parallel read capability ensures fast retrieval from storage nodes.
Signed URLs: Generate temporary URLs for premium content, restricting access via S3’s authentication layer in the request pipeline. URLs expire after a set time (e.g., 1 hour).

Tie to S3 Architecture:

Geographic Distribution: Videos are replicated across AZs in a region (e.g., US East) for durability. CloudFront pulls from the closest AZ or cached copy.
Parallel Reads: Large videos are split into chunks across devices, enabling simultaneous reads for high throughput.
Request Pipeline: Signed URL requests are validated by S3’s API gateway and control plane, ensuring only authorized users access files.
Strong Consistency: Uploaded videos are immediately available for CloudFront to serve, thanks to S3’s read-after-write consistency.

Real-World Example: A streaming service like Netflix stores HD videos in S3. A user in London requests a video; CloudFront serves it from a nearby edge location. If uncached, S3’s parallel reads fetch chunks from multiple AZs, ensuring fast delivery. Signed URLs restrict premium content to subscribers. Costs: ~$0.023/GB (S3 Standard) + CloudFront fees (~$0.085/GB in US).

Scenario: An analytics company stores and processes petabytes of data (e.g., web logs, sales data) from various sources for querying and analysis.

How S3 Fits: S3 acts as the storage layer for a data lake, organizing massive datasets in buckets with partitioned structures for efficiency. Its scalability handles petabytes, while storage classes optimize costs. Integration with AWS Athena enables SQL queries directly on S3 data, leveraging metadata for fast access. Lifecycle policies automate cost savings.

Implementation Details:

Partitioning: Organize data by attributes (e.g., year/month/day/hour) to improve query performance. S3’s control plane stores metadata for each partition’s objects.
Storage Classes: Use S3 Standard for frequently accessed raw data, Standard-IA for processed data, and Glacier for archives, managed via lifecycle rules.
Athena Integration: Athena reads S3 data directly, using metadata to locate objects without scanning the entire bucket.
Lifecycle Management: Automatically transition older data to cheaper classes or expire it.

Tie to S3 Architecture:

Device-Level Distribution: Large datasets are split into chunks across storage nodes, enabling parallel reads for Athena queries.
Metadata Management: S3 stores partition metadata (e.g., object keys, sizes), which Athena uses to optimize queries.
Background Operations: Lifecycle transitions and data integrity checks run automatically, ensuring cost efficiency and durability.
Strong Consistency: New data uploads are immediately queryable by Athena.

Example Structure:

Data Lake Structure:
s3://analytics-data-lake/
├── raw-data/
│   ├── year=2024/month=01/day=15/hour=14/
│   │   ├── web-logs-001.json.gz
│   │   ├── web-logs-002.json.gz
│   │   └── ...
├── processed-data/
│   ├── daily-aggregates/
│   └── monthly-reports/
└── archived-data/
    └── year=2023/

Real-World Example: A retail company stores web logs in analytics-data-lake. Analysts use Athena to query sales trends (e.g., SELECT SUM(sales) FROM raw-data WHERE month='01'). Raw logs transition to Standard-IA after 30 days, Glacier after 90, and expire after a year, saving ~70% on storage costs. S3’s parallel reads speed up Athena queries, and metadata ensures efficient partition access.

Scenario: A financial services company needs secure, compliant backups of critical data (e.g., transaction records) with geographic redundancy.

How S3 Fits: S3’s high durability, versioning, and replication make it ideal for backups. Cross-region replication (CRR) ensures data exists in multiple regions, while Glacier Deep Archive offers ultra-low-cost storage for long-term retention. MFA delete and encryption add security, managed via the control plane.

Implementation Details:

Multi-Region Replication: CRR copies data to another region for disaster recovery. Background operations handle asynchronous replication.
Versioning: Tracks object versions, protecting against accidental overwrites/deletions. Metadata stores version IDs.
MFA Delete: Requires multi-factor authentication for deletes, enforced in the request pipeline.
Glacier Deep Archive: Stores rarely accessed backups at ~$0.00099/GB/month, using erasure coding for efficiency.

Tie to S3 Architecture:

Geographic Distribution: CRR extends replication beyond AZs to other regions, using storage nodes for redundancy.
Metadata Management: Stores version IDs, replication status, and encryption details in the control plane.
Background Operations: Manages replication, integrity checks, and transitions to Deep Archive.
Strong Consistency: Ensures deleted versions are immediately unavailable, critical for compliance.

Real-World Example: A bank stores transaction logs in my-backup-bucket (US East). CRR replicates to EU (Frankfurt) for compliance. Versioning protects against overwrites; MFA delete secures against unauthorized deletion. After 180 days, logs move to Deep Archive, costing ~$1/TB/month. If US East fails, the EU bucket ensures recovery. S3’s erasure coding and integrity checks maintain data reliability.

Scenario: A SaaS app (e.g., document management) stores user-uploaded files with varying access patterns, requiring security and cost optimization.

How S3 Fits: S3’s Intelligent-Tiering automatically moves objects between access tiers (frequent/infrequent) based on usage, leveraging background operations. Server-side encryption secures data, and IAM policies control access via the request pipeline. Event notifications trigger workflows (e.g., Lambda processing), integrated with the control plane.

Implementation Details:

Intelligent-Tiering: Auto-transitions objects to save costs without manual rules. Metadata tracks access patterns.
Server-Side Encryption: Applies AES256 encryption during upload, stored in metadata.
IAM Policies: Fine-grained access control, enforced in the request pipeline.
Event Notifications: Trigger Lambda or SQS when objects are created/updated, managed by S3’s control plane.

Tie to S3 Architecture:

Device-Level Distribution: Files are chunked across nodes for parallel reads/writes, speeding up uploads/downloads.
Metadata Management: Stores encryption settings, storage class, and user metadata (e.g., user-id).
Background Operations: Intelligent-Tiering monitors access; notifications trigger workflows.
Strong Consistency: Ensures uploaded files are immediately available for processing.

Real-World Example: A SaaS app lets users upload PDFs to app-user-data. Files use Intelligent-Tiering, moving to infrequent access after 30 days of inactivity, saving ~30% on costs. AES256 encryption secures data; IAM restricts users to their folders (e.g., user123/*). Uploads trigger a Lambda to generate thumbnails, using S3’s event system. Parallel writes ensure fast uploads, and consistency allows immediate processing.

Content Distribution: S3 + CloudFront delivers content globally with signed URLs for security, leveraging parallel reads and edge caching.
Data Lake: Partitioned buckets and lifecycle policies optimize analytics with Athena, using metadata for efficient queries.
Backup/DR: CRR, versioning, and Deep Archive ensure compliance and recovery, with erasure coding for cost efficiency.
App Storage: Intelligent-Tiering and encryption optimize costs/security; notifications enable automation, all managed via S3’s scalable architecture.
Costs: Vary by class (Standard: ~$0.023/GB, IA: ~$0.0125/GB, Deep Archive: ~$0.00099/GB) + transfer fees.

Ep #130: The Engineering Behind Amazon S3 Part 5: Different Use cases

Comments