TiKV — Distributed Transactional Key-Value Store on Raft

Introduction

TiKV is a distributed, transactional, strongly-consistent key-value store written in Rust and originally built to back TiDB. It replicates regions of ordered keys via Raft, provides ACID transactions with a Google Percolator-style MVCC layer, and scales horizontally by splitting and rebalancing regions automatically.

What TiKV Does

• Stores ordered byte keys across many nodes with auto-sharding into 96 MB regions.
• Replicates each region with Raft for strong consistency and automatic failover.
• Exposes two APIs: raw KV (single-key ops) and transactional KV (multi-key Percolator MVCC).
• Lets the Placement Driver (PD) schedule region splits, merges and replica placement for balance.
• Integrates with Prometheus, Grafana and jaeger-style tracing for deep observability.

Architecture Overview

Clients talk to PD to discover which TiKV node owns the region for a key, then to that node directly. Each region is a Raft group with 3+ replicas; the leader appends entries to a local RocksDB instance (RaftDB for log, KvDB for data). Writes become Raft log entries replicated to a quorum before applying. PD continuously monitors load and issues split, merge, leader-transfer and replica-move operations to keep the cluster balanced.

Self-Hosting & Configuration

• Deploy with TiUP, Kubernetes Operator (tidb-operator), or ansible for bare metal.
• Size 3 PD + 3+ TiKV nodes minimum; TiKV loves NVMe SSDs and 10 GbE networking.
• Tune storage.block-cache.capacity to ~45% of RAM on dedicated nodes.
• Use placement rules to pin replicas to AZs or racks for geo-distribution.
• Enable TLS between PD, TiKV and clients for production deployments.

Key Features

• Snapshot-isolation and repeatable-read transactions across thousands of nodes.
• Online schema and range changes without locking, inherited by TiDB.
• Native support for async commit and 1PC optimisations that cut latency.
• Coprocessor framework pushes filters/aggregates down to the storage layer.
• Multi-Raft architecture scales to PB-class clusters in production at PingCAP users.

Comparison with Similar Tools

• etcd — also Raft-based KV but single-group, small-data focus.
• CockroachDB KV layer — very similar design; TiKV is a standalone product plus TiDB SQL.
• FoundationDB — ACID KV with different transaction model; harder to operate.
• Cassandra / ScyllaDB — eventual-consistency wide-column; different consistency trade-off.
• DynamoDB — managed AWS equivalent; TiKV is self-hosted and open source.

FAQ

Q: Do I need TiDB to use TiKV? A: No — TiKV is a standalone KV store with Rust, Go, Java, Python and Node.js clients. Q: How does TiKV scale writes? A: Regions split at size/QPS thresholds and PD rebalances leaders across nodes. Q: What's the transaction model? A: Percolator 2PC with MVCC timestamps from PD, plus async commit for lower latency. Q: Is TiKV production-ready? A: Yes — CNCF graduated in 2020 and runs at PB scale at companies like Zhihu and Square.

Sources

• https://github.com/tikv/tikv
• https://tikv.org/docs/