v0.3.1

Released: November 2026

v0.3.1 brings the Redis Lua scripting engine to FyroDB, alongside major write-path performance enhancements, direct in-process L1 cache routing for the PostgreSQL proxy, memory footprint right-sizing, and pub/sub TCP throughput scaling.

Highlights

Redis Lua Scripting Engine

v0.3.1 introduces a fully Redis-compatible Lua 5.4 scripting runtime powered by mlua 0.12.2:

• Commands: Full support for EVAL, EVALSHA, SCRIPT LOAD, SCRIPT EXISTS, and SCRIPT FLUSH.
• redis.call and redis.pcall: Re-entrant invocation directly into FyroDB's internal command execution pipeline.
• Pre-compiled SHA1 Digest Cache: Scripts are compiled once, cached in memory by their SHA1 digest, and executed with zero overhead on repeated invocations.
• Generational Garbage Collection & Isolated State: Each script runs within an isolated state with generational garbage collection to prevent memory accumulation across long sessions.
• Zero-Allocation Type Conversions: Fast conversion paths between RESP frames and Lua values without intermediate heap strings or vectors.

DB Cache Proxy: In-Process L1 Cache Routing

The PostgreSQL query caching proxy now supports direct in-process result routing between the proxy and the KV storage engine:

• In-Process Cache Hits: Cache hits bypass the TCP loopback network stack entirely, retrieving query results directly from memory.
• 8-Slot Hot Query Cache: Micro-cached queries eliminate repeated lookups for high-frequency queries.
• Zero-Allocation Parameter Decoding: Extended protocol parameter decoding eliminates heap allocations during query inspection.
• Unnamed Portal Fast-Path: Direct execution for unnamed portals common in ORM connection pooling traffic.

Zero-Allocation Write Run Parsing & Pipelined Execution

• try_fast_tail Parsing: Pipelined write commands (ZADD, SADD, HSET, PFADD, GEOADD, XADD, LPUSH/RPOP) verify wire-level command prefixes via a single 64-bit comparison and extract arguments into fixed stack arrays ([RawSlice; QUEUE_RUN_MAX * 3]), bypassing general parse_one() overhead and eliminating argument vector allocations.
• Pipelined JSON.GET Run Dispatcher: Added try_fast_json_get_key_at with batch slot and entry prefetching (dispatch_json_get_run) and direct raw bulk byte emission (write_bulk_raw). Batch window for JSON.SET expanded to 256.
• Greedy SET + EXPIRE Pair Absorption: try_fast_set_expire_pair_at greedily absorbs adjacent SET key val and EXPIRE key ttl command pairs directly out of socket buffers into fused atomic writes under a single lock.

High-Throughput Pub/Sub TCP Scaling

• Batched Fanout Delivery: Subscriptions to the same channel receive batched fanout delivery in the pub/sub worker, drastically reducing event-loop context switching.
• Uniform Cached Frame Encoding: Pre-encodes outbound pub/sub messages into CachedFrame shared across all channel subscribers.
• Socket Buffer Tuning: Subscriber sockets configure a 1 MB SO_SNDBUF to maximize TCP pipeline throughput under heavy fanout bursts.
• try_fast_publish_at: Dedicated fast path for PUBLISH command extraction during event loop processing.

Storage Shard Capacity & Memory Right-Sizing

• Right-Sized Shard Capacity: Reduced default INITIAL_SHARD_CAPACITY from 131,072 to 32,768, saving over 100 MB of empty table pointers across 128 shards while maintaining dynamic growth.
• Adaptive Write Buffer Retention: Standard connections retain a 32 KB write buffer between read cycles, while dedicated pub/sub subscriber connections retain up to 256 KB.
• Lookahead Cache-Line Prefetching: CustomHash probe sequences and shard rebuilding tables issue _mm_prefetch hints to overlap memory latency during resizing.

Collection Fast Paths & Buffer Streaming

• Sequential Integer Set Fast Path: Monotonic sequential integer insertions in Set storage execute via an $O(1)$ push.
• Streamed Set Flag Responses: sadd_added_flags_into formats :1\r\n and :0\r\n responses directly into connection write buffers, eliminating intermediate Vec<bool> allocations.
• Pre-Allocated ZSet Index: ZSetData::with_capacity pre-allocates skip-index entries when capacity exceeds 128, right-sizing chunks to CHUNK_MAX = 64.
• Fast Float Serialization: Integrated ryu for zero-allocation floating point formatting.

Bug Fixes

• CustomHash Spinlock Deadlock: Resolved a spinlock deadlock on heavily contended single keys by introducing exponential backoff with thread yielding.
• Shard Capacity Growth: Fixed capacity growth boundary checks in CustomHash shards to prevent premature table rebuild stalls.
• Pub/Sub Protocol Alignment: Fixed SUBSCRIBE / UNSUBSCRIBE responses to return per-channel message frames conforming to the Redis specification without temporary Vec allocations.
• Pub/Sub Frame Off-by-One: Fixed an off-by-one error in total byte length calculation in encode_message().
• Store::hset_bytes Counter: Corrected return counter semantics in Store::hset_bytes to properly report newly created versus modified fields.

Upgrade

Update your docker-compose.yml:

image: rana718/fyrodb:0.3.1
# or
image: rana718/fyrodb:latest

No breaking changes. All existing configuration, Redis commands, and client connections continue to work seamlessly.

Summary