Every year on Double 11, Tmall’s turnover will set the world record time and time again, and 100% of Alibaba’s core systems will go to the cloud, supporting the world-class traffic peak of Double 11. In this traffic battlefield, Ali can be said to be the biggest winner of Huashan Lunjian. Today, I shared Ali's distributed cache documents for Taobao Double 11 trillion-level traffic, and comprehensively analyzed Ali's double 11 martial arts secrets.
Four application scenarios for Ali Double Eleven
MDB typical application scenarios
- Used for caching to reduce the access pressure to the back-end database.
- Temporary data storage, some data loss will not have a major impact on the business.
- Read more and write less, read qps up to ten thousand level.
FastDump typical application scenarios
- Periodically quickly import offline data into the Tair cluster, and quickly use new data.
- Reading has low latency and no glitches.
Typical application scenarios of LDB
- General kv storage, transaction snapshots, security risk control, etc.
- Store black and white single data, read qps very high.
- The counter function is updated very frequently, and the data cannot be lost.
RDB typical application scenarios
- Caching and storage of complex data structures.
- Such as playlists, live broadcast rooms, etc.
Double Eleven Technical Challenge
- Access peak growth rate: Tair peak> Transaction peak> Total GMV
- Multi-region and multi-unit
- Experience and cost
- Stability, high availability
Performance and cost
Memory data structure
- cache meta
- slab_mabager
- hashmap
- pages
Effect of lock
- Fine-grained locks
- Lock-free data structures (lock-free data structures)
- CPU local data structures (per-CPU data structures)
- Read Copy Update (RCU)
User mode protocol stack + memory merge
Client optimization
- Network framework replacement, adaptation coroutine mina=netty throughput increased by 40%+
- Serialization optimization integration of kryo and hessian throughput increased by 16%+
Memory grid
- Large amount of reads and writes in business scenarios A large number of local computing provides high-performance computing fast IO
- Characteristic data local read penetration Write Through / Write Behind/merge multi-unit replication
- The effect of reading down to 27.68% and writing down to 55.75%
Caching problems-hot spots
Cache breakdown
- Burst traffic hot products, store current affairs news and various stress tests
- Buffer breakdown current limit
- System-wide collapse
- Root access hotspot
Hotspot hash
- Hotspot type frequency hotspot traffic hotspot
Hot spot hashing effect
- Hotspot QPS number
- Hotspot traffic
- Hotspot cluster hashing effect comparison
Write hotspot
- Hot spot merge real-time recognition merge write engine
Alibaba’s double eleven trillion traffic distributed cache fully shows how to design to respond to hundreds of millions of requests.
Teach you how to build a high-concurrency, high-traffic system to withstand the real test of hundreds of millions of online user traffic.
Since the quality of the headline post picture may be affected (the picture will be compressed and the clarity will decrease), the full version of Ali's double eleven trillion traffic distributed cache design, I will do an open source sharing here to help more Many Internet practitioners who want to completely improve their hard power from the technical level of major manufacturers.
