Alibaba ACA Cloud Business cheat sheet

ECS instance families: g (general purpose), c (compute optimized), r (memory optimized), s (shared/burstable); choose c for CPU-bound, r for memory-bound workloads.

ESS (Auto Scaling) adjusts ECS capacity using scaling rules (target tracking, scheduled, or custom) and supports multi-AZ scaling groups.

OSS storage classes: Standard (hot), Infrequent Access (IA), Archive, Cold Archive, Deep Cold Archive; retrieval time and minimum storage duration increase down the list.

SLB types: Classic (L4), Application (L7), Network (L4, high performance); ALB supports HTTP/HTTPS and advanced routing, NLB is for extreme performance.

VPC provides isolated networks; each VPC has a CIDR block, vSwitches (subnets) in one AZ, route tables, and security groups acting as stateful firewalls.

RDS engines: MySQL, SQL Server, PostgreSQL, MariaDB, PPAS; supports read-only instances, backup/recovery, and multi-AZ high availability.

ApsaraDB for Redis is an in-memory cache/database; use for session caching, leaderboards, and pub/sub; supports cluster and read/write splitting.

Alibaba Cloud CDN caches content at edge nodes; origin can be OSS, ECS, or an external server; supports HTTPS, cache rules, and refresh/preload.

MaxCompute is a big-data warehouse for batch SQL; DataWorks provides data integration, scheduling, and governance; Realtime Compute for Apache Flink handles streaming.

High availability pattern: deploy ECS across multiple AZs in one region behind SLB, use multi-AZ RDS, and enable auto scaling to replace failed instances.

Decoupling pattern: use Message Queue (MQ) or Kafka to buffer traffic between tiers, preventing cascading failures and enabling asynchronous processing.

Multi-tier web architecture: SLB (public) → web tier (ECS in multiple AZs) → app tier → RDS; use NAT Gateway for outbound internet from private subnets.

Security design: security groups (stateful, instance-level) and network ACLs (stateless, subnet-level); use RAM for identity and access management with least privilege.

Disaster recovery: RDS cross-region backup and OSS cross-region replication; RTO/RPO drive choice between backup/restore, pilot light, warm standby, and active-active.

Cost optimization: use reserved instances or savings plans for steady workloads, spot instances for fault-tolerant batch, and auto scaling to match demand.

Scalability: vertical scaling (resize ECS) is limited; horizontal scaling (add instances behind SLB) and stateless design enable elastic growth.

Data durability: OSS provides 99.9999999999% (12 nines) durability; use versioning and cross-region replication for critical data.

Monitoring: CloudMonitor collects metrics and sets alarms; Log Service (SLS) centralizes logs; ActionTrail records API activity for audit.

Migration phases: assessment → design → migration (pilot, then bulk) → validation → cutover; use Alibaba Cloud Migration Hub to track progress.

Server migration: Server Migration Center (SMC) migrates physical/virtual/cloud servers to ECS via a client; supports incremental sync and cutover.

Database migration: Data Transmission Service (DTS) supports schema migration, full+incremental data sync, and continuous replication between homogeneous/heterogeneous databases.

Storage migration: ossutil or Data Online Migration service moves data from local or other clouds into OSS; supports large-scale parallel transfer.

E-commerce scenario: SLB + ECS auto scaling + Redis for session/cache + RDS + MQ for order decoupling; CDN for static assets to handle traffic spikes.

Hybrid cloud: Express Connect (dedicated line) or VPN Gateway connects on-premises to VPC; use for gradual migration and data residency.

Global business: use multiple regions with Global Accelerator or CDN to reduce latency; deploy data in-region to meet compliance.

Migration risk control: run pilot migrations, validate with load testing, keep rollback plans, and schedule cutover during low-traffic windows.

PAI (Platform for AI) provides model training, development, and deployment; supports notebooks, DLC (Deep Learning Containers), and EAS (Elastic Algorithm Service) for inference.

Model Studio (Bailian) offers access to Qwen foundation models and tools for RAG, agent building, and model fine-tuning via API.

Qwen is Alibaba Cloud's large language model family; available in different sizes (e.g., Qwen-Turbo, Qwen-Plus, Qwen-Max) trading off cost, speed, and capability.

RAG pattern: retrieve relevant documents from a vector store, augment the prompt, then generate an answer; reduces hallucination and grounds responses in enterprise data.

Vector retrieval: use embedding models to convert text to vectors, store in a vector database, and perform similarity search for RAG or recommendation.

Generative AI architecture: data ingestion → embedding → vector store → LLM inference (via API) → application layer; add guardrails and logging.

AI compute: GPU-accelerated ECS instances (e.g., gn series) for training/inference; PAI manages distributed training and auto scaling of inference endpoints.

Responsible AI: apply content moderation, access control (RAM), data encryption, and prompt/response logging to meet compliance.