S3-Native Architecture

Enterprise KV Store
Built for Scale

The only distributed key-value store with S3 as source of truth. No hot partitions. Fast global reads. 90% lower cost.

< 5ms
Read Latency
90%
Cost Savings
Billions
Records

The Traditional Tradeoff is Over

Choose performance, durability, AND cost efficiency

In-Memory Stores

Redis, Memcached

  • ✗ $0.05/GB/hour
  • ✗ Data loss on crashes
  • ✗ Complex clustering at scale
  • ✓ Fast reads (< 1ms)
  • ✓ BYOC option (your VPC)
☁️

Managed Cloud KV

DynamoDB, Cosmos DB

  • ✗ $0.25/GB/month
  • ✗ Hot key/partition throttling
  • ✗ Expensive bulk load operations
  • ✓ Managed & durable
  • ✗ No BYOC option
Best Choice
🎯

BoulderKV

S3-Native Design

  • ✓ $0.023/GB/month
  • ✓ 11 nines durability
  • ✓ Linear scaling
  • ✓ Fast reads (< 5ms)
  • ✓ BYOC (your VPC)

Built for Enterprise Workloads

Everything you need for mission-critical applications

Flexible Performance

Hot data served from cache at <5ms, cold data from S3 at 80-200ms. Control your cache budget to optimize cost vs. performance for your workload.

You choose the tradeoff
💎

S3-Backed Durability

Every write lands in S3 first. 99.999999999% durability guarantee with automatic replication across regions.

11 nines durability
💰

Predictable Economics

Object storage pricing with no per-read costs. Scale to billions of features without breaking the budget.

90% cost reduction
📈

Linear Scaling

Handle billions of features with predictable performance. Add regions and capacity independently.

Horizontal scalability
🔄

Async Replication

Sync updates across online and offline feature stores. One source of truth for training and serving.

Unified feature store
🛠️

Developer-Friendly

Simple HTTP/REST API with SDKs for Python, Go, and Node.js. Integrate in minutes, not weeks.

Fast integration

Proven Use Cases

Trusted by teams building at scale

🤖

ML Feature Stores

Unified storage for training and inference features. Low-latency serving with historical consistency.

Real-time serving Batch training Time travel
🚀

Inference Caches

Cache model predictions at massive scale. Reduce GPU costs by up to 80% for repeated queries.

LLM caching Embedding lookup GPU savings
🎮

Game State Storage

Store player progress and game state with low-latency access. Scale to millions of concurrent players.

Player profiles Leaderboards Session data
🌍

CDN & Edge KV

Distribute configuration, routing rules, and metadata globally. Instant propagation without restarts.

Feature flags Routing config A/B tests

Architecture That Makes Sense

Built for the object storage era

The Write Path

1
Durable Write to S3
Every write lands in S3 first for immutable durability and 11 nines guarantee
2
Async Index Update
Regional RocksDB indexes sync asynchronously (typically < 30s)
3
Global Availability
Data propagates to all configured regions automatically

The Read Path

1
Cache-First Lookup
Hot data served from local disk cache at <5ms. You control cache budget to balance cost and performance.
2
S3 Fallback
Cold data fetched directly from S3 when not cached. Typical latency 80-200ms with minimal cost.
3
Intelligent Tiering
System automatically keeps frequently accessed data in cache based on your budget

💡 Consistency Model

BoulderKV favors durability and cost efficiency over immediate read-your-writes consistency. Writes are durable immediately in S3, with eventual consistency to regional indexes (typically < 30s). Perfect for read-heavy workloads where strong consistency isn't required.

🔨 Building This?

We're adapting RocksDB for object storage from first principles and need engineers who care about microsecond-level performance. Join us →

Frequently Asked Questions

Everything you need to know about BoulderKV

What technology powers BoulderKV? +

BoulderKV is built on RocksDB, the proven LSM-tree database used by Facebook, LinkedIn, and many others. We've re-architected it from first principles to use S3 as the source of truth, combining proven reliability with object storage economics.

What is the consistency model? +

BoulderKV uses an eventually consistent model optimized for durability and cost efficiency. Writes are immediately durable in S3 (100% reliable), and propagate to regional indexes typically within 30 seconds. This makes it perfect for read-heavy workloads where strong consistency isn't required, such as ML feature stores, caching, and configuration data.

What latency should I expect for reads and writes? +

Reads: Hot data served from cache at <5ms (P99). Cold data fetched from S3 at 80-200ms. You control your cache budget to optimize the cost vs. performance tradeoff for your workload.

Writes: Typically 10-50ms for S3 durability, with async replication to indexes completing within 30 seconds.

How does caching work and can I control it? +

BoulderKV uses intelligent RocksDB-backed caching to keep frequently accessed data on local disk. You set your cache budget (size limit), and the system automatically maintains hot data based on access patterns. This gives you full control over the cost-performance tradeoff: larger cache = better latency but higher costs, smaller cache = more S3 reads but lower costs.

When should I NOT use BoulderKV? +

BoulderKV isn't ideal for:

  • Strong consistency requirements - Use traditional kv stores if you need read-your-writes guarantees
  • Sub-millisecond latency needs - Cache hits are <5ms, but not sub-millisecond like pure in-memory stores
  • Very small datasets (<10GB) - Traditional solutions may be simpler and cheaper at small scale
Can I deploy in my own VPC (BYOC)? +

Yes! BoulderKV supports Bring Your Own Cloud deployments. You can run BoulderKV servers in your own VPC using your own S3 buckets, giving you complete data sovereignty and control. We provide the software, management plane, and support while your data never leaves your infrastructure.

What happens if S3 is unavailable? +

If S3 becomes unavailable (extremely rare with 99.99% SLA), cached data remains readable with no interruption. Writes will fail until S3 recovers, ensuring we never lose data or create inconsistencies. Given S3's track record of 11 nines durability and high availability, outages are exceptionally rare.

How is pricing calculated? +

Detailed pricing information is coming soon. We're finalizing our pricing model to ensure it works perfectly for teams of all sizes. Early access customers will get special introductory rates and direct input on the pricing structure. Request early access to learn more.

Is there a free tier or trial? +

Early access customers receive credits and special introductory pricing to test BoulderKV with their workloads. Request early access to learn more about trial options.

What are the size limits for keys and values? +
  • Max key size: 1KB
  • Max value size: 5MB (S3 single PUT limit)
  • Batch operations: Up to 1000 keys per batch request
  • Total dataset: Unlimited (scales with S3)

For larger values, we recommend storing references to S3 objects directly.

How do I migrate my existing data? +

BoulderKV provides bulk import tools for common sources:

  • Direct S3 import from existing buckets
  • Redis dump file import
  • CSV/JSON batch upload via API
  • Streaming import from other KV stores (DynamoDB, etc.)

Our team provides migration support for early access customers.

How does it compare to Redis and DynamoDB? +
FeatureRedisDynamoDBBoulderKV
Read Latency<1ms~10ms<5ms (cached)
Bulk OperationsExpensive at scaleHigh per-read costNo per-read fees
Hot KeysClustering neededPartition throttlingNo throttling
DurabilityAt-risk3 nines11 nines
ConsistencyStrongEventual/StrongEventual
BYOC
What regions and cloud providers are supported? +

AWS: All S3-enabled regions worldwide (primary support)

Multi-cloud: Support for GCS and Azure Blob Storage coming soon

BoulderKV can replicate data across multiple regions automatically for global low-latency access.

Pricing Designed for Predictability

Built on principles that make sense for modern infrastructure

💰

Cost Predictability

No surprise bills from read charges. Storage-based pricing that scales linearly with your data, not your traffic.

  • ✓ Object storage economics
  • ✓ No per-read fees
  • ✓ Transparent cost model
📊

Massive Savings

Up to 90% cost reduction compared to traditional KV stores. Perfect for read-heavy workloads that would break the bank elsewhere.

  • ✓ 90% cheaper than in-memory
  • ✓ No vendor lock-in
  • ✓ Bring your own cloud
🎯

Simple Model

Pay for what you store and what you write. Read billions of records without worrying about usage spikes or throttling.

  • ✓ Storage + write charges
  • ✓ Unlimited reads
  • ✓ No capacity planning
Coming Soon

Detailed Pricing Available for Early Access

We're finalizing our pricing tiers to ensure they work perfectly for teams of all sizes. Early access customers will get special introductory rates and direct input on our pricing structure.

Questions about pricing? Contact us

Ready to Transform Your KV Infrastructure?

Join the private beta and experience S3-native storage built for read-heavy workloads

Questions? Email us at hello@boulderkv.com