we use Rust in TiKV 黄东旭 PingCAP 关于我 ● 黄东旭 Dongxu_Huang ● Open source hacker / Infrastructure Engineer ● MSRA / Netease / Wandoujia / PingCAP ● CTO of PingCAP ● Codis / TiDB / TiKV ● Go / Python 是否满足 Send + Sync TiKV ● 大规模分布式 Key-Value 数据库 ● 支持 ACID 跨行事务支持 ● 支持 MVCC 无锁的快照读 ● 构建于 Raft 之上，不依赖分布式文件系统 ○ 更少的第三方依赖 ○ 更高的性能（低延迟） ● 配合 TiDB 使用，需要有健全的逻辑实现 SQL 层的下推算子 TiKV 的前置需求 ● 极高的性能要求，尽可能低延迟，而且延迟需要稳定 ○ Multi-Raft ○ 分布式测试框架 ● 和 C 的模块大量交互 ● 开发人员和时间不足 ○ 5 人的团队，希望半年左右发布第一个可用版本 TiKV 的语言选型参考 ● Go ● C++11 ● Rust TiKV 整体架构 KV API Coprocessor Txn, Transaction MVCC Raft KV RocksDB Placement

right and make it fast. ● 测试很重要 ● 简单易用 ● 和社区结合 架构 TiKV TiKV TiKV TiKV Raft Raft Raft TiDB TiDB TiDB ... ... ... ... ... Placement Driver (PD) Control flow: Balance / Failover Metadata / Timestamp Storage stack 1/2 ● TiKV is the underlying storage layer ● Physically, data is stored in RocksDB ● We build a Raft layer on top of RocksDB ○ What is Raft? ● Written in Rust! TiKV API (gRPC) Transaction 2:[d-e] split 分裂: 1/4 TiKV1 Region 1:[a-e] TiKV2 Region 1:[a-e] TiKV3 Region 1:[a-e] raft raft Leader Follower Follower Raft group 分裂: 2/4 TiKV2 Region 1:[a-e] TiKV3 Region 1:[a-e] raft

Google Spanner Google F1 TiDB RDBMS NoSQL NewSQL Architecture TiKV TiKV TiKV TiKV Raft Raft Raft TiDB TiDB TiDB ... ... Placement Driver (PD) Control flow: Balance / Failover Metadata / Timestamp request Distributed Storage Layer ... TiKV - Overview • Region: a set of continuous key-value pairs • Data is organized/stored/replicated by Regions • Highly layered TiKV Key Space [ start_key, end_key) Raft 256MB PD - Overview TiKV TiKV TiKV TiKV … ... TiKV Cluster PD Node/Region Info Management Command TiKV Client Route Info ● Meta data management ● Load balance management TiKV - Multi-Raft

副标题 TiDB TiDB Worker Spark Driver TiKV Cluster (Storage) Metadata TiKV TiKV TiKV MySQL Clients Syncer Data location Job TiSpark DistSQL API TiKV TiDB TSO/Data location Worker Worker Worker Spark Cluster TiDB Cluster TiDB ... ... ... DistSQL API PD PD PD Cluster TiKV TiKV TiDB PD 天下大势，合久必分，分久必合 《三国演义》 图 + 关系型多模实现 01. 副标题 g_prefix Vertex_ID Tag_ID g_prefix src_Vertex_ID

protocol SQL NoSQL tidb-server tikv-server •No single point of failure •Automatic failover
 High Availability at Scale •Snapshot isolation @ TiDB •Linearizability @ TiKV
 Strong Consistency at Scale

+ Rust 插件 Gitlab CI/CD 心路历程 真刀实枪 – 文档工具 文档生成 CI/CD git page 作为文档 心路历程 吐槽点 ● Nightly 版本编译器： tikv ， raft-rs ● 无官方库，官方库支持不成熟 ○ Kafka （第三方： https://github.com/kafka-rust/kafka-rust) ○ PostgreSQL

Constructors Quat() = default; Quat(Scalar w, Scalar x, Scalar y, Scalar z) { wx_ = _mm_set_pd(x, w); yz_ = _mm_set_pd(z, y); } template Quat(const Q& rhs) : Quat(static_cast(rhs.w()) return _mm_cvtsd_f64(_mm_unpackhi_pd(SseWx(), SseWx())); } Scalar y() const { return _mm_cvtsd_f64 (SseYz()); } Scalar z() const { return _mm_cvtsd_f64(_mm_unpackhi_pd(SseYz(), SseYz())); } __m128d SseWx() __m128d w2x2 = _mm_mul_pd(lhs.SseWx(), rhs.SseWx()); __m128d x2w2 = _mm_shuffle_pd(w2x2, w2x2, _MM_SHUFFLE2(0, 1)); __m128d wx2wx2 = _mm_add_pd(w2x2, x2w2); __m128d y2z2 = _mm_mul_pd(lhs.SseYz(), rhs.SseYz());

Widget( double& d ) : pd_( &d ) {} double& get() noexcept { return *pd_; } double const& get() const noexcept { return *pd_; } private: double* pd_; }; Reference members

std::get_if(&v)) { std::cout << *pi << "\n"; } else if (double *pd = std::get_if(&v)) { std::cout << *pd << "\n"; } else if (auto *ps = std::get_if(&v)) { std::cout

linear_load(const double* base) { __m128i vindex {0x100000000ul, 0x300000002ul}; return _mm256_i32gather_pd(base, vindex * ElementSize, 1); } also implementable as gather operationPossible Extension: gather/scatter 4> result; __m128i vindex {0x200000000ul, 0x600000004ul}; stdx::__data(result) = _mm256_i32gather_pd( &(dataPtr + i.scalar_index(0))->x, vindex, 8); benchmark::DoNotOptimize(result); }, VectorResidualLoop);Possible