Amazon S3 Vectors & S3 Tables Overview
Amazon S3 has evolved beyond simple object storage into an AI-native and analytics-native storage platform with two groundbreaking services: Amazon S3 Vectors and Amazon S3 Tables. These services bring purpose-built capabilities directly into S3, eliminating the need for separate specialized databases while maintaining S3’s legendary durability, scalability, and cost efficiency.
- S3 Vectors (GA December 2025) – The first cloud object storage with native vector support for AI workloads, semantic search, and RAG applications
- S3 Tables (GA December 2024) – Fully managed Apache Iceberg tables optimized for analytics workloads with automatic maintenance
Amazon S3 Vectors
What is Amazon S3 Vectors?
Amazon S3 Vectors is the first cloud object store with native support to store and query vector embeddings, providing purpose-built, cost-optimized vector storage for AI agents, AI inference, and semantic search. It reduces the total cost of uploading, storing, and querying vectors by up to 90% compared to specialized vector database solutions.
Key Features
- Massive Scale – Store up to 2 billion vectors per index, 10,000 indexes per bucket (up to 20 trillion vectors total)
- Sub-second Performance – Warm query latency as low as 100ms; infrequent queries return in under 1 second
- 90% Cost Reduction – Dramatically lower costs compared to specialized vector databases
- Fully Serverless – No infrastructure to provision or manage; pay only for what you use
- Strong Consistency – Subsequent queries always include the most recently added data
- Metadata Filtering – Up to 50 metadata keys per vector (10 non-filterable) for combined similarity + attribute filtering
- Distance Metrics – Supports Cosine and Euclidean distance metrics
- Encryption – SSE-S3 (default) or SSE-KMS with custom keys at bucket or index level
- High Write Throughput – Up to 1,000 PUT transactions per second for streaming updates
- Top-K Results – Return up to 100 search results per query
Architecture Components
| Component | Description |
|---|---|
| Vector Bucket | A new S3 bucket type dedicated to vector storage with vector-specific APIs |
| Vector Index | Container for vectors within a bucket; defines dimensionality and distance metric |
| Vectors | Individual embeddings with key, data (float32), and optional metadata |
| Metadata | Key-value pairs attached to vectors; filterable and non-filterable types |
AWS Service Integrations
- Amazon Bedrock Knowledge Bases – Native integration for cost-effective RAG applications; select S3 Vectors as the vector store when creating a knowledge base
- Amazon OpenSearch Service – Tiered strategy: store long-term vectors in S3 Vectors and export high-priority vectors to OpenSearch for real-time, low-latency search
- Amazon SageMaker Unified Studio – Create and manage knowledge bases with S3 Vectors within the unified AI development environment
- AWS CloudFormation – Deploy and manage vector resources as infrastructure as code
- AWS PrivateLink – Private network connectivity for secure access
S3 Vectors Use Cases
- Semantic Search – Search across millions of documents, images, videos, and audio based on meaning rather than keywords
- Retrieval Augmented Generation (RAG) – Provide contextual information to LLMs from large document collections at reduced cost
- AI Agent Memory – Give AI agents lasting memory by storing every interaction and insight across petabytes of vector data
- Media Intelligence – Index video content at frame level (e.g., 5,400+ embeddings per hour of video) for instant scene retrieval
- Personalization Engines – Store user preference embeddings for recommendation systems
- Medical Image Similarity – Compare medical images (radiology, pathology) against vast databases for diagnosis assistance
- Code Search – Navigate large codebases using semantic similarity
S3 Vectors Pricing (US East – N. Virginia)
| Dimension | Rate |
|---|---|
| Storage | $0.06 per GB/month |
| PUT (Upload) | $0.20 per GB uploaded (min 128KB per PUT) |
| Query – API Fee | $2.50 per million queries |
| Query – Data Processed (first 100K vectors) | $0.004 per TB |
| Query – Data Processed (100K–10M vectors) | $0.002 per TB |
| Query – Data Processed (10M+ vectors) | $0.0004 per TB |
| Query – Data Returned | $0.01 per GB (first 512KB/query free) |
| Vector Bucket | Free to create |
Pricing Example: 10 million vectors (6.17 KB each) across 40 indexes, with 1M queries/month returning top 100 results = approximately $11.38/month. This demonstrates the dramatic cost savings compared to traditional vector databases.
Amazon S3 Tables
What are Amazon S3 Tables?
Amazon S3 Tables are fully managed Apache Iceberg tables that automate the operational burden of managing data lakes and lakehouses. They deliver storage specifically optimized for analytics workloads with automatic table maintenance, delivering up to 3x faster query performance and 10x higher transactions per second compared to self-managed Iceberg tables in general purpose S3 buckets.
Key Features
- Fully Managed Apache Iceberg – Automatic compaction, snapshot management, and unreferenced file removal
- Advanced Compaction Strategies – Binpack (default), sort compaction, and z-order compaction for multi-dimensional queries
- S3 Tables Intelligent-Tiering – Automatically moves data to the most cost-effective tier, reducing storage costs by up to 80%
- Automatic Replication – Cross-region table replication for reduced query latency and disaster recovery
- Iceberg REST Catalog API – Compatible with any Iceberg engine (Spark, Trino, Flink, Athena, Redshift, Snowflake)
- Apache Iceberg V3 Support – Deletion vectors for efficient batch updates, reducing write amplification
- Table-Level Security – Built-in access control, encryption, and lifecycle management per table
- AWS Analytics Integration – Native integration with AWS Glue Data Catalog, Lake Formation, Athena, EMR, and Redshift
- MCP Support – AI agents and LLMs can interact with S3 Tables via Model Context Protocol
- 11 Nines Durability – 99.999999999% durability with 99.99% availability
How S3 Tables Work
- Create a Table Bucket – A new S3 bucket type purpose-built for tabular data
- Create Tables – Define Apache Iceberg tables using SQL (via Athena) or Iceberg REST Catalog API
- Ingest Data – Write data using Spark, Flink, Firehose, or any Iceberg-compatible engine
- Automatic Maintenance – S3 continuously compacts, manages snapshots, and removes orphan files
- Query with Any Engine – Use Athena, Redshift, Spark, Trino, DuckDB, or Snowflake
S3 Tables Use Cases
- Data Lake Modernization – Migrate from Parquet/Hive/Hadoop to managed Iceberg tables with reduced complexity
- Streaming Analytics – Stream data from IoT sensors, transactions, and application logs with near real-time queryability
- Big Data Analytics – High-throughput workloads benefiting from 10x higher TPS
- AI-Powered Analytics – Query data using natural language through MCP for ad-hoc exploration
- Transactional Data Lakes – ACID transactions with time-travel and schema evolution
- Compliance & Audit – Immutable audit trails with snapshot history and data versioning
S3 Tables Pricing (US West – Oregon)
| Dimension | Rate |
|---|---|
| Storage (Standard) | $0.0265 per GB/month (first 50TB) |
| PUT Requests | $0.005 per 1,000 requests |
| GET Requests | $0.0004 per 1,000 requests |
| Object Monitoring | $0.025 per 1,000 objects |
| Compaction – Objects Processed | $0.002 per 1,000 objects |
| Compaction – Data Processed | $0.005 per GB |
| Replication – Table Updates | $0.010 per 1,000 table updates |
| Table Bucket | Free to create |
Pricing Example: 1 TB table with 30,000 new files/month, 500K GET requests, and automatic compaction = approximately $28.54/month (storage + requests + monitoring + compaction).
Vector Database Comparison: S3 Vectors vs OpenSearch vs Pinecone vs pgvector
| Feature | S3 Vectors | OpenSearch Service | Pinecone | pgvector (Aurora/RDS) |
|---|---|---|---|---|
| Max Vectors | 2B per index (20T per bucket) | Billions (distributed) | Billions (serverless) | Millions (instance-bound) |
| Query Latency | ~100ms (warm), <1s (cold) | Single-digit ms | ~10-50ms | ~10-100ms (depends on index) |
| QPS (Queries/Second) | Low-Medium (optimized for infrequent) | Very High (thousands) | High (thousands) | Medium (hundreds) |
| Cost at Scale | Very Low (90% cheaper) | High (compute + storage) | High (pod/serverless units) | Medium (instance cost) |
| Infrastructure | Fully Serverless | Managed (Serverless option) | Fully Serverless | Managed instances |
| Hybrid Search | Metadata filtering only | Full (keyword + vector + filters) | Metadata filtering | Full SQL + vector |
| Real-time Updates | Strong consistency, 1K TPS writes | Near real-time | Near real-time | Immediate (ACID) |
| AWS Integration | Native (Bedrock, OpenSearch, SageMaker) | Native AWS service | Third-party (API) | Native (Aurora/RDS) |
| Best For | Large-scale, cost-sensitive, infrequent queries | Real-time search, high QPS | Quick start, managed vector DB | Existing PostgreSQL apps, ACID needs |
When to Use Each Vector Solution
- Choose S3 Vectors when:
- You have large vector datasets (millions to billions) with infrequent query patterns
- Cost optimization is the primary concern
- You need long-term durable storage for vector embeddings
- Building RAG applications with Amazon Bedrock Knowledge Bases
- Implementing a tiered strategy (cold vectors in S3, hot vectors in OpenSearch)
- Query latency of 100ms–1s is acceptable
- Choose OpenSearch Service when:
- You need single-digit millisecond latency
- High query throughput (thousands of QPS) is required
- You need hybrid search combining keyword, vector, and structured filters
- Real-time applications like product recommendations or fraud detection
- Choose Pinecone when:
- You want a fully managed vector-only database with minimal setup
- Multi-cloud or vendor-neutral strategy is important
- You don’t need tight AWS service integration
- Choose pgvector when:
- You already use PostgreSQL and want to add vector search to existing data
- You need ACID transactions combining relational and vector data
- Vector dataset is relatively small (millions, not billions)
- You prefer a single database for both structured queries and similarity search
Analytics Storage Comparison: S3 Tables vs Athena vs Redshift
| Feature | S3 Tables | Amazon Athena | Amazon Redshift |
|---|---|---|---|
| Type | Managed Storage Layer (Iceberg) | Serverless Query Engine | Data Warehouse |
| Infrastructure | No compute to manage | Serverless (no provisioning) | Provisioned clusters or Serverless |
| Table Format | Apache Iceberg (managed) | Reads Iceberg, Hive, Delta, Parquet, CSV, JSON | Native columnar + Spectrum for S3/Iceberg |
| Table Maintenance | Automatic (compaction, snapshots, cleanup) | None (user-managed) | Automatic (VACUUM, ANALYZE) |
| Query Performance | Depends on engine (Athena, Spark, etc.) | Seconds to minutes (scan-based) | Sub-second to seconds (optimized) |
| Concurrency | 10x higher TPS than self-managed Iceberg | Limited concurrent queries | High (hundreds of concurrent queries) |
| Pricing Model | Storage + requests + compaction | $5 per TB scanned | Per-node-hour or RPU-hour (Serverless) |
| Best For | Managed lakehouse storage at scale | Ad-hoc queries, exploration | Complex analytics, dashboards, BI |
Understanding the Relationship
S3 Tables is a storage layer, not a query engine. It complements Athena and Redshift rather than replacing them:
- S3 Tables + Athena – Best for serverless data lake analytics with automatic table optimization. Athena queries S3 Tables directly with improved performance from automatic compaction.
- S3 Tables + Redshift – Best for high-concurrency BI dashboards and complex joins. Redshift reads S3 Tables through Spectrum or direct Iceberg integration.
- S3 Tables + Spark (EMR/Glue) – Best for large-scale ETL, streaming ingestion, and ML feature engineering.
When to Use Each Analytics Approach
- Choose S3 Tables when:
- You want managed Iceberg tables without maintaining compaction jobs
- Data lake with multiple query engines (Athena + Redshift + Spark)
- Streaming data that needs to be queryable in near real-time
- You want Intelligent-Tiering to automatically optimize storage costs
- Multi-region data access with automatic replication
- Choose Athena (with S3 Tables or general purpose S3) when:
- Ad-hoc exploration of data in S3
- Pay-per-query model is preferred (no idle costs)
- Simple queries that don’t require complex joins across large datasets
- Infrequent querying where provisioned resources would be wasteful
- Choose Redshift when:
- High-concurrency BI dashboards with sub-second response times
- Complex analytical queries with many joins across large tables
- Predictable, heavy workloads that justify provisioned compute
- Advanced features like materialized views, stored procedures, and ML integration
S3 Vectors vs S3 Tables – Quick Comparison
| Aspect | S3 Vectors | S3 Tables |
|---|---|---|
| Data Type | Vector embeddings (unstructured → vectors) | Tabular/structured data (rows and columns) |
| Query Type | Similarity search (nearest neighbor) | SQL analytics (filter, aggregate, join) |
| Primary Use Case | AI/ML, RAG, semantic search | Data lakes, analytics, BI |
| Bucket Type | Vector Bucket | Table Bucket |
| Open Standard | Proprietary S3 Vectors API | Apache Iceberg (open) |
| GA Date | December 2025 | December 2024 |
| Complementary Use | Store embeddings for AI agents and search | Store structured data for analytics queries |
Together: S3 Vectors and S3 Tables can work in tandem. For example, store product catalog data in S3 Tables for analytical queries while storing product description embeddings in S3 Vectors for semantic search. Both share S3’s durability, security model, and operational simplicity.
AWS Certification Exam Practice Questions
Question 1
A company needs to store 500 million vector embeddings for a semantic search application. The application receives approximately 100 queries per minute and requires results within 1 second. Cost optimization is the highest priority. Which solution best meets these requirements?
- Amazon OpenSearch Service with vector engine
- Amazon S3 Vectors
- Amazon Aurora PostgreSQL with pgvector extension
- Self-managed Pinecone on Amazon EC2
Show Answer
Answer: B –
Explanation: S3 Vectors is designed for large-scale vector storage (up to 2B per index) with sub-second query performance at up to 90% lower cost than specialized vector databases. With 100 queries/minute (infrequent pattern) and cost as the primary concern, S3 Vectors is the ideal choice. OpenSearch would provide faster latency but at significantly higher cost. pgvector would struggle at 500M vectors. Pinecone on EC2 is not a valid deployment model.
Question 2
A data engineering team manages an Apache Iceberg data lake on Amazon S3. They spend significant time running compaction jobs, managing snapshots, and cleaning orphan files. Queries are becoming slower as the data grows. Which solution would reduce their operational overhead while improving query performance?
- Move data to Amazon Redshift
- Use Amazon S3 Tables with automatic table maintenance
- Implement S3 Lifecycle policies to delete old files
- Add more partitions to the existing Iceberg tables
Show Answer
Answer: B –
Explanation: Amazon S3 Tables provides fully managed Apache Iceberg tables with automatic compaction, snapshot management, and unreferenced file removal. This eliminates the operational burden while delivering up to 3x faster query performance. Moving to Redshift changes the architecture entirely. Lifecycle policies don’t understand Iceberg metadata. Adding partitions doesn’t address compaction issues.
Question 3
An AI team is building a RAG application using Amazon Bedrock Knowledge Bases. They need to store vector embeddings for 50 million documents with the lowest possible cost. The application will query the knowledge base approximately 10,000 times per day. Which vector store should they select in the Bedrock Knowledge Bases configuration? (Select TWO benefits of this choice)
- Amazon OpenSearch Serverless – provides lowest latency
- Amazon S3 Vectors – reduces vector storage and query costs by up to 90%
- Amazon S3 Vectors – provides native integration with Bedrock Knowledge Bases
- Amazon Aurora PostgreSQL – provides ACID transactions for vectors
- Amazon OpenSearch Serverless – provides automatic scaling with no minimum charges
Show Answer
Answer: B, C
Explanation: S3 Vectors is natively integrated with Amazon Bedrock Knowledge Bases and reduces the cost of vector storage and querying by up to 90%. With 10,000 queries/day (~7 per minute), the infrequent query pattern is ideal for S3 Vectors. While OpenSearch provides lower latency, the question prioritizes cost. Aurora pgvector is not a native Bedrock Knowledge Bases vector store option.
Question 4
A company has Apache Iceberg tables stored in general purpose S3 buckets. They want to improve query performance, reduce storage costs with automatic tiering, and replicate tables to a second region for disaster recovery. Which combination of features is available with Amazon S3 Tables? (Select THREE)
- Automatic compaction with sort and z-order strategies
- S3 Glacier Deep Archive storage for table data
- S3 Tables Intelligent-Tiering for automatic storage cost optimization
- Cross-region table replication
- Real-time CDC (Change Data Capture) to Amazon DynamoDB
- Native integration with Amazon Kinesis Data Streams
Show Answer
Answer: A, C, D
Explanation: S3 Tables supports automatic compaction (including sort and z-order), Intelligent-Tiering that reduces storage costs by up to 80%, and automatic cross-region table replication. S3 Tables does not support Glacier storage classes, CDC to DynamoDB, or native Kinesis integration.
Question 5
A solutions architect is designing a tiered vector search architecture. Real-time product recommendations need single-digit millisecond latency, while a historical document search (accessed a few times per hour) needs to be cost-optimized. Which architecture should the architect implement?
- Use Amazon OpenSearch Service for all vector workloads with auto-scaling
- Store all vectors in Amazon S3 Vectors and optimize with metadata filtering
- Store hot vectors in Amazon OpenSearch Service for real-time queries and cold vectors in Amazon S3 Vectors for infrequent historical search
- Use Amazon Aurora pgvector for product recommendations and Pinecone for document search
Show Answer
Answer: C –
Explanation: The tiered strategy combining OpenSearch (for hot, real-time queries requiring single-digit ms latency) and S3 Vectors (for cold, infrequent queries where sub-second latency is acceptable) provides the optimal balance of performance and cost. S3 Vectors natively integrates with OpenSearch, allowing vectors to be exported from S3 to OpenSearch when demand increases. Using OpenSearch for everything would be costly; using S3 Vectors for everything wouldn’t meet the real-time latency requirement.
Frequently Asked Questions
What is Amazon S3 Vectors?
S3 Vectors is a purpose-built storage service for AI embeddings. It stores up to 2 billion vectors per index, supports 100ms query latencies, and costs up to 90% less than specialized vector databases — integrating natively with Bedrock and SageMaker.
What is Amazon S3 Tables?
S3 Tables provides first-class Apache Iceberg table support in S3. It manages compaction, snapshot management, and schema evolution automatically, enabling analytics queries via Athena, Redshift, and EMR with built-in Intelligent-Tiering.
When should I use S3 Vectors vs OpenSearch?
Use S3 Vectors for cost-optimized vector storage at scale (billions of embeddings) with simple similarity search. Use OpenSearch when you need hybrid search (vector + keyword), complex filtering, or real-time indexing with sub-10ms latency.