AI summary by TwinMind
The Fastest SQL Database Ever Made
Theo - t3․gg
19m 38s
Performance Improvements with PlanetScale Metal
- T3 chat P99 latency dropped from 300ms to 45ms after February 28th migration
- P999 latency dropped to under 100ms average (previously multiple seconds)
- Occasional spikes to ~300ms but rare (single digit number of queries)
- No schema changes, indexes, or caching required - pure hardware upgrade
- Cost: $600/month for new database tier
Technical Architecture Comparison
Traditional AWS Aurora Setup
- Separate compute nodes and network storage layer
- CPU nodes are ephemeral, data stored on network-attached storage
- Network gap between query processing and data storage creates latency bottleneck
- Multiple compute nodes connect to shared network storage
PlanetScale Metal Architecture
- Each node has local NVMe SSD storage directly attached
- No network hop between compute and storage
- Uses Vitess for replication and consistency across nodes
- Built on custom hardware with fastest possible storage via PCI lanes
- All nodes <28 days old due to constant cycling/replacement
Vitess Technology Foundation
- Originally built at YouTube for MySQL scaling at massive scale
- Open sourced, team moved to GitHub, then founded PlanetScale
- Built for "most hostile possible environment" on Borg (pre-Kubernetes)
- Designed assuming any node could die at any time
- Handles sharding, scaling, replication automatically
- Used by Slack, Cash App (3 months), Intercom, Depot, Block
Performance Benchmarks vs Competitors
Test Configuration
- 4 vCPUs, 32GB memory
- 500GB database size
- Various connection counts (50, 100, 200)
AWS Aurora Results
- Non-proxied: Relatively stable but slower than PlanetScale
- Proxied: Terrible spikes every 75 seconds, P99 up to 2400ms
- Cost: $6,500/month vs PlanetScale $400/month for comparable setup
- Extreme config comparison: $122,000/month vs $3,000/month
Google Cloud SQL Results
- Extremely spiky performance with drops to near-zero QPS
- Takes 200-300 seconds to degrade, doesn't recover
- Cheaper for small storage but more expensive at scale
- Inconsistent latency makes it unsuitable for production
PlanetScale Metal Results
- 2x faster P99 latency than non-proxied Aurora
- Consistent performance without spikes
- "Unlimited IOPS" - not bottlenecked by network storage
- 2-6x cheaper than comparable AWS solutions
Product Positioning and Market Impact
- First SQL database faster than AWS offerings while running on AWS infrastructure
- Eliminates traditional tradeoff between developer experience and performance
- Features: branching, rollbacks, automatic dual writes during migrations
- 30-minute rollback window with automatic old version maintenance
- Better DX + faster performance + lower cost combination unprecedented
Use Case Considerations
- Not suitable if $600/month database cost is prohibitive
- Metal not always better: small datasets fitting in memory may perform better on traditional solutions
- IO bottleneck elimination is key benefit - unlimited IOPS advantage
- When IO isn't bottleneck, benefits diminish but consistency remains
