Serverless Architecture

This document provides a comprehensive technical overview of SkySQL Serverless architecture, explaining how it achieves true serverless capabilities while maintaining full compatibility with MySQL and MariaDB.

Design Philosophy

SkySQL Serverless is built on the principle: "Don't change what works". Instead of re-architecting the database engine like other cloud providers, SkySQL leverages cloud-native techniques to achieve serverless capabilities while preserving the mature, open-source database engine.

Core Principles

Preserve Open Source: Keep the proven InnoDB storage engine intact
Cloud-Native Approach: Use Kubernetes and containers for orchestration
No Forking: Maintain full compatibility with existing applications
Transparent Operations: Scaling and management should be invisible to applications

High-Level Architecture

graph TB
    A[Application] --> B[Intelligent Proxy]
    B --> C[Kubernetes Cluster]

    subgraph "SkySQL Control Plane"
        D[Resource Monitor]
        E[Scaling Controller] 
        F[Migration Controller]
        G[Backup Controller]
    end

    subgraph "Kubernetes Cluster"
        H[Database Pod 1]
        I[Database Pod 2]
        J[Database Pod N]
        K[Storage Volumes]
    end

    subgraph "Pre-Fabricated Pool"
        L[Warm DB Pool]
        M[Micro Databases]
    end

    C --> D
    D --> E
    E --> H
    E --> I
    E --> J

    H --> K
    I --> K
    J --> K

    L --> C
    M --> L

Core Components

1. Intelligent Proxy

The multi-tenant proxy is the cornerstone of SkySQL Serverless, providing:

Connection Management

Always-On Connections: Maintains application connections even when database scales to zero
Connection Pooling: Efficiently manages database connections behind the scenes
Load Balancing: Distributes requests across available database instances

Session State Management

The proxy tracks and preserves: - System variables: SET @@session.sort_buffer_size = X - User variables: SET @myvar = 'value' - Prepared statements and their definitions - Transaction isolation levels and other session settings

Failover and Recovery

Transparent Failover: Automatically handles database failures
Transaction Replay: Replays partial transactions to ensure data integrity
State Recreation: Re-establishes session state on new connections

2. Kubernetes Orchestration

SkySQL extends Kubernetes with custom controllers for database-specific operations:

Custom Resource Definitions (CRDs)

DatabaseService: Defines serverless database configurations
ScalingPolicy: Controls auto-scaling behavior and limits
BackupSchedule: Manages automated backup operations

Custom Controllers

Resource Monitor Controller: Tracks CPU, memory, and connection metrics
Scaling Controller: Implements vertical and horizontal scaling decisions
Migration Controller: Handles transparent live migrations
Pool Controller: Manages pre-fabricated database pools

3. Pre-Fabricated Database Pools

To achieve millisecond launch times, SkySQL maintains pools of ready-to-use databases:

Pool Management

Regional Distribution: Pools maintained in all supported regions
Dynamic Replenishment: Pools refilled based on demand patterns
Resource Optimization: Minimal resource allocation for pool databases

Database Initialization

When a user requests a database: 1. Check out a database from the appropriate pool 2. Resize according to service configuration 3. Execute security procedures (user creation, endpoint configuration) 4. Update control plane tracking 5. Database ready for use

4. Auto-Scaling Engine

SkySQL implements sophisticated auto-scaling across multiple dimensions:

Vertical Scaling Algorithm

Every 200ms:
  - Sample CPU usage (UsageCoreNanoSeconds)
  - Update 30-second sliding window

Scale-up decision (1-second window):
  - If CPU usage ≥ 90% of allocated budget → Scale up

Scale-down decision (30-second window):
  - If average CPU usage ≤ 20% → Scale down
  - If no active connections for 10s → Scale to 0

Resource Allocation

Granular Scaling: 0.5 vCPU and 2GB memory increments
Dynamic Limits: Free tier limited to 2 SCUs, paid tiers scale based on configuration
Linux cgroups: Uses cgroupsv2 for precise resource control

Database States and Lifecycle

Active State

Full Resources: CPU, memory, and connections allocated based on demand
Optimized Cache: Buffer pool sized and hydrated for current workload
Real-time Monitoring: Continuous performance and resource tracking

Suspended State

When activity drops to zero: 1. Scale Down Resources: Remove active threads, reduce memory allocation 2. Minimize Buffer Pool: Reduce to minimum required by InnoDB 3. Maintain Connections: Proxy keeps application connections alive 4. Quick Reactivation: Instant scale-up when activity resumes

Parked State

After extended inactivity (several hours): 1. Terminate Pod: Database pod completely removed 2. Preserve Storage: Volume remains attached to service 3. Proxy Management: Proxy tracks parked state 4. Automatic Recreation: Pod recreated when activity resumes

Buffer Pool Management

The database buffer pool is critical for performance. SkySQL implements intelligent buffer pool management:

Dynamic Sizing

Proportional Scaling: Buffer pool size adjusted with resource allocation
Performance Monitoring: Track cache hit ratios during scaling operations

Cache Hydration

When scaling up after a scale-down: 1. Page Tracking: Most frequently used pages identified during scale-down 2. SSD Storage: Page IDs stored on high-speed SSD storage 3. Background Loading: Frequently used pages reloaded into memory 4. Performance Consistency: Maintains cache hit ratios across scaling events

Implementation Details

-- Buffer pool hydration process
1. Before scale-down: SHOW ENGINE INNODB STATUS → identify hot pages
2. Store page IDs to fast SSD storage
3. During scale-up: Background process fetches pages from disk
4. Asynchronous hydration maintains query performance

Live Migration System

For horizontal scaling, SkySQL implements transparent live migrations:

Migration Triggers

High Watermark: Migration initiated at ~70% memory utilization
Automatic Provisioning: New instances created if needed
Workload Analysis: Least-used databases migrated first

Migration Process

Snapshot Creation: Create database snapshot on target instance
Replication Setup: Establish replication channel to source
Synchronization: Wait for replica to catch up
Proxy Redirection: Transparently redirect connections
Source Cleanup: Decommission source database

Zero-Downtime Guarantees

Differential Snapshots: Only changes copied after initial snapshot
Connection Preservation: No application connection drops
Session Continuity: All session state preserved during migration

Storage Management

Auto-Scaling Storage

SkySQL monitors and scales storage automatically:

Scaling Thresholds

Small Volumes (< 100GB): Scale at 60% capacity
Medium Volumes (100-500GB): Scale at 80% capacity
Large Volumes (> 500GB): Scale at 95% capacity

Implementation

Continuous Monitoring: PersistentVolumeClaim usage tracked
Proactive Scaling: Scaling initiated before capacity exhausted
Block Storage Integration: Currently uses cloud provider block storage (AWS EBS, Azure Disk, Google Persistent Disk)

Future Storage Innovations

SkySQL is evaluating distributed storage solutions: - Ceph/Rook Integration: Self-managed distributed storage - Multi-Cloud Storage: Storage spanning multiple cloud providers - Performance Optimization: Custom storage optimizations for database workloads

Security Architecture

Network Security

Private Networking: Database pods run in private subnets
Encryption in Transit: TLS/SSL for all connections
Firewall Integration: Cloud-native security group integration

Data Protection

Encryption at Rest: All storage volumes encrypted
Key Management: Integration with cloud provider key management services
Access Controls: Role-based access control (RBAC)

Isolation

Pod-Level Isolation: Each database runs in isolated Kubernetes pod
Network Policies: Kubernetes network policies for traffic control
Resource Isolation: cgroups ensure resource isolation between databases

Monitoring and Observability

Real-Time Metrics

Resource Utilization: CPU, memory, storage, and network metrics
Database Performance: Query performance, connection counts, cache hit ratios
Scaling Events: Auto-scaling decisions and their impact

Alerting System

Proactive Alerts: Performance degradation warnings
Scaling Notifications: Automatic scaling event notifications
Failure Detection: Immediate alerts for database failures

Integration

Prometheus: Metrics collection and storage
Grafana: Visualization and dashboards
Custom Metrics: Database-specific performance indicators

Other Considerations

Vendor Lock-in: SkySQL maintains portability
Cost Transparency: No hidden charges or surprise costs
Performance: Better performance due to no compute-storage disaggregation
Compatibility: Full compatibility with existing applications

Performance Characteristics

Scaling Performance

Vertical Scaling: Typically completes in < 5 seconds
Horizontal Scaling: Live migration in < 30 seconds
Cold Start: Database ready in < 100 milliseconds

Query Performance

OLTP Workloads: Equivalent to provisioned instances
Cache Performance: Maintained through buffer pool hydration
Connection Overhead: Minimal proxy overhead (< 1ms latency)

Scalability Limits

Vertical Scaling: Up to cloud provider instance limits
Horizontal Scaling: Unlimited through live migration
Storage Scaling: Up to cloud provider storage limits

Future Enhancements

Planned Features

Analytics Integration: On-demand OLAP with DuckDB integration
Global Distribution: Multi-region database deployment
Advanced AI: Machine learning-driven optimization
Edge Computing: Edge database deployments

Research Areas

Quantum-Ready Encryption: Future-proof security
Advanced Caching: Intelligent cache management algorithms
Distributed Consensus: Enhanced distributed database capabilities

Best Practices

Application Design

Connection Pooling: Use connection pooling in applications
Graceful Degradation: Handle temporary scaling events
Monitoring Integration: Implement application-level monitoring

Performance Optimization

Query Optimization: Optimize queries for scaling environments
Index Strategy: Maintain appropriate indexes for workload
Connection Management: Minimize connection overhead

Cost Optimization

Workload Analysis: Understand application usage patterns
Scaling Limits: Set appropriate scaling limits
Resource Right-Sizing: Monitor and adjust resource allocation

This architecture enables SkySQL Serverless to provide true serverless capabilities while maintaining the performance, reliability, and compatibility that enterprises require.