Changelog v0.2.1

Pipelined batch coalescing, zero-allocation command paths, grouped Pub/Sub fan-out, scoped worker threads, and extreme throughput gains across all data types.

v0.2.1

Released: September 2026

v0.2.1 is a high-performance optimization release focusing on pipelined batch coalescing, zero-allocation execution hot paths, grouped lock-free Pub/Sub fan-out, and concurrency safety refinements.

In this release, a single FyroDB node delivers up to 50.11M ops/sec INCR, 44.07M ops/sec Hot Key under contention, 39.26M ops/sec LPUSH/RPOP, and 74.64M msg/sec Pub/Sub delivery, outperforming a 6-node Redis Cluster across all benchmarks while maintaining lower memory overhead and predictable latency.

Highlights

  • Pipelined Queue Coalescing: Consecutive same-key queue commands (LPUSH, RPUSH, LPOP, RPOP) in a pipeline run are executed under a single lock acquisition, producing byte-identical replies and eliminating queue lock contention.
  • Atomic SET + EXPIRE Coalescing: Adjacent pipelined SET and EXPIRE / PEXPIRE / EXPIREAT / PEXPIREAT pairs fuse into an atomic set-with-TTL write path, cutting lock overhead by 50%.
  • Grouped Pub/Sub Fan-Out: Same-channel PUBLISH runs batch into grouped per-worker fan-out queues, eliminating redundant per-message subscriber queue enqueue operations.
  • Zero-Allocation Command Paths: Direct serialization to TCP output buffers for HGETALL, SMEMBERS, LRANGE, ZRANGE, HSET, HMGET, SMISMEMBER, ZMSCORE, HGET, SADD, ZADD, JSON.SET, JSON.GET, LPOP, and RPOP.
  • Zero-Alloc Inline Keys: Optimized CompactKey and SmallStr representation up to 23 bytes inline, with run-of-1 fast path in worker dispatch.
  • Quadratic ZADD Scan Fix: Eliminated 37x warm-server stall on ZADD member re-insertion via score-ordered indexing optimizations.
  • Floor-Probe Hash Decomposition: Optimized linear open-addressing table probing in customhash with decomposed bucket floor probing.
  • Vectorized Cluster Transport: Inter-node cluster bus framing upgraded to vectorized writev I/O (2.6x improvement per frame).
  • Scoped Worker Threads: Migrated server worker thread pool to std::thread::scope for clean lifecycle management and panic safety.

Architectural Deep Dive

Pipelined Batch Coalescing

When clients send pipelined batches of commands over a single connection, acquiring and releasing shard/entry locks for every individual command introduces substantial synchronization overhead.

v0.2.1 introduces automatic batch coalescing:

  1. Queue Commands (LPUSH, RPUSH, LPOP, RPOP): When consecutive commands in a pipelined run target the same queue key, FyroDB acquires the entry lock once, processes the entire run of operations in-place, and serializes the exact sequence of RESP integer or bulk string replies.
  2. SET + EXPIRE Pairs: Common caching patterns (SET key val immediately followed by EXPIRE key ttl) are detected during pipelined frame parsing and executed as a single atomic write-with-TTL operation.
  3. PUBLISH Batching: When multiple PUBLISH operations target the same channel within an event loop cycle, subscriber fan-out is consolidated into a single grouped dispatch per worker thread.

Zero-Allocation Serialization

Previous releases allocated intermediate Vec<u8> buffers or cloned strings when formatting multi-bulk replies. In v0.2.1:

  • Collection reads (HGETALL, SMEMBERS, LRANGE, ZRANGE) format RESP headers and stream element bytes directly into the connection's TCP output buffer under the protection of EBR and seqlock snapshots.
  • Multi-lookup commands (HMGET, SMISMEMBER, ZMSCORE) write multi-bulk element headers directly without intermediate vector staging.
  • Single-key mutators (HSET, SADD, ZADD, JSON.SET) serialize integer or status replies directly into output slices.

Storage & Probing Optimizations

  • ZADD Warm-Server Acceleration: Resolved a quadratic scan during repeated member insertions into large sorted sets, preventing latency spikes on heavily populated zsets.
  • Floor-Probe Decomposition: Decomposed the hash probe step in customhash::CustomMap into aligned floor probes, reducing cache misses and branch mispredictions.
  • Allocator Backends: Added compile-time feature flags (jemalloc, system) to rust-zmalloc, while maintaining mimalloc as the default zero-overhead allocator.

Benchmarks

Benchmarked on a 6-core Intel i5-11400H (12 hardware threads), loopback TCP, 100 concurrent clients, 1M operations per run:

FyroDB (1 node) vs. Redis Cluster (6 nodes)

BenchmarkFyroDB (1 node)Redis Cluster (6 nodes)Speedup
Pipeline-64 SET8.95M ops/sec5.75M ops/sec1.6×
Pipelined SET20.08M ops/sec7.56M ops/sec2.7×
Pipelined GET28.16M ops/sec11.29M ops/sec2.5×
Pub/Sub publish1.49M ops/sec147.1K ops/sec10.1×
Pub/Sub delivery74.64M msg/sec7.35M msg/sec10.2×
Mixed SET/GET (50/50)20.90M ops/sec7.24M ops/sec2.9×
INCR (atomic counters)50.11M ops/sec6.32M ops/sec7.9×
HSET/HGET (sessions)33.07M ops/sec7.95M ops/sec4.2×
LPUSH/RPOP (queue)39.26M ops/sec7.19M ops/sec5.5×
SADD (per-client sets)27.87M ops/sec6.34M ops/sec4.4×
ZADD (per-client zsets)18.12M ops/sec4.55M ops/sec4.0×
JSON.SET/GET (documents)14.91M ops/sec3.38M ops/sec4.4×
SET+EXPIRE (cache TTL)10.51M ops/sec2.64M ops/sec4.0×
Hot Key (1 key, contention)44.07M ops/sec2.03M ops/sec21.7×
Producer/Consumer (50+50)36.20M ops/sec981.3K ops/sec36.9×

Resource Usage

MetricFyroDB (1 node, PID 2258)Redis Cluster (6 nodes, 7 containers)
Peak RSS294.19 MB767.63 MB
Avg RSS168.81 MB347.46 MB
Peak CPU79.6%442.1%
Avg CPU42.5%125.2%

Upgrade Notes

  • Full Backward Compatibility: Drop-in replacement for v0.2.0. No breaking protocol or storage format changes.
  • RDB Persistence: Fully compatible with existing RDB snapshot files.
  • Client Compatibility: Standard Redis standalone and cluster drivers work seamlessly without modification.