kv-store

Distributed key-value store in C++17. Built to understand how systems like Redis and etcd work under the hood — consistent hashing, gossip-based failure detection, Raft consensus, and async I/O via io_uring.

Architecture

• Server — io_uring event loop feeds a lock-free thread pool; text protocol parsed with string_view (zero-copy)
• Storage — 16-shard hash table backed by a custom slab allocator (mmap, per-size-class free lists)
• Cluster — consistent hashing ring (CRC32, virtual nodes) for key distribution across nodes
• Membership — UDP gossip with ALIVE/SUSPECT/DEAD state machine; failure detection without a central coordinator
• Consensus — Raft for strong-consistency writes: leader election, log replication, WAL-persisted to disk
• Consistency levels — ONE, QUORUM, ALL, STRONG (Raft path)

Build

Requires Linux 5.11+ (for io_uring), liburing, CMake 3.16+, GCC 10+ or Clang 12+.

apt install liburing-dev
cmake -B build && make -C build -j$(nproc)

Usage

./build/kvstore --port 6379

echo "SET foo bar" | nc localhost 6379
echo "GET foo"     | nc localhost 6379

Start a 6-node cluster:

./scripts/run_cluster.sh

Benchmarks

# kvstore vs redis side by side
./scripts/run_bench_compare.sh

# memory comparison
./scripts/measure_memory.sh

# perf flamegraph (requires FlameGraph + sudo sysctl -w kernel.perf_event_paranoid=1)
./scripts/run_perf.sh ~/FlameGraph

Single-node results (x86 dev server, 200k ops, 64b values):

system	ops/sec	p50	p99
kvstore	23,350	42.5 µs	50.3 µs
redis	12,556	63.8 µs	141.1 µs

Memory at 100k keys, 64b values:

system	RSS	bytes/key
kvstore	27 MB	281
redis	14 MB	139