v0.2.0

Released: August 2026

v0.2.0 is a major milestone release that introduces production-grade Redis Cluster compatibility, high-availability replication with automated failover, online slot migration, and deep storage and concurrency performance optimizations.

A single FyroDB node outperforms a 12-master Redis Cluster while consuming roughly half the memory. In cluster mode, FyroDB provides drop-in compatibility with standard cluster-aware Redis clients (such as redis-py, ioredis, go-redis, jedis, and redis-cli -c).

---

Highlights

• Full Redis Cluster Compatibility: 16,384 hash slots, CRC16 hashing, hash tags ({...}), automatic MOVED and ASK client redirections, and multi-key CROSSSLOT validation.
• High-Availability Replication & Failover: Master-replica replication stream backed by an in-memory replication backlog ring buffer, asynchronous PSYNC / full RDB resync, consensus-based failure detection (PFAIL/FAIL), and automated failover election.
• Online Slot Migration: Live slot migration supporting MIGRATING and IMPORTING states, cluster state synchronization, and ASKING command semantics.
• Flat 16,384-Entry Routing Table: Zero-allocation slot lookup resolving through a shared flat routing table, cached per connection and validated with a single atomic epoch comparison.
• Tagged-Pointer Hash Probing: Embedded 7-bit hash tags in CustomMap slot metadata skip non-matching entries during linear probing without dereferencing entry cache lines.
• Contention Spin Backoff: Exponential backoff and thread yields under spinlock contention eliminate cache-line bouncing and negative scaling on hot keys.
• Allocation-Free Pops & Lazy Buffers: LPOP and RPOP serialize directly into connection output buffers; accepted idle connections defer buffer allocation until active.
• Unified Key-Capacity Admission: Default FYRODB_MAX_KEYS set to 0 (unlimited, matching Redis maxmemory 0).

---

Redis Cluster & Topology Management

16,384 Hash Slot Architecture

• Standard CRC16 slot computation mapping all keys to slots 0..16383.
• Hash tag extraction: curly braces ({user:100}:profile) hash only the substring inside the brackets, guaranteeing related keys map to the same slot and master.
• Multi-key commands (MGET, MSET, DEL, transactions) automatically enforce single-slot locality, returning CROSSSLOT Keys in request don't hash to the same slot if keys span multiple slots.

Inter-Node Cluster Bus & Failure Detection

• Dedicated cluster bus port (port + 10000) for inter-node communication with custom high-performance binary framing.
• Heartbeat gossip mechanism (PING/PONG) propagating cluster topology, node flags, and epoch metadata.
• Decentralized failure detection: nodes track PFAIL (possible failure) states and transition to FAIL upon consensus from a majority of masters.

High-Availability Replication & Auto-Failover

• Streaming master-replica replication engine (ReplicationEngine) with an in-memory replication backlog buffer.
• Asynchronous partial resync (PSYNC) and full resynchronization via streaming RDB snapshots.
• Automated failover election: when a master is flagged as FAIL, healthy replicas request votes from remaining masters and promote themselves upon majority consensus.
• Replica auto-reconnect and state reconciliation.

Online Slot Migration

• Dynamic slot state transitions: SETSLOT <slot> MIGRATING <node-id>, SETSLOT <slot> IMPORTING <node-id>, SETSLOT <slot> STABLE, and SETSLOT <slot> NODE <node-id>.
• Client ASK redirection during migration windows with full support for the ASKING command flag.
• Per-slot key tracking in storage engine for COUNTKEYSINSLOT and GETKEYSINSLOT.

Cluster Configuration Persistence (nodes.conf)

• Automatic state persistence conforming to the Redis nodes.conf file format.
• Node identity tracking (myid), current config epochs, and slot allocation tables.
• Automatic configuration reloading on startup and manual flush via CLUSTER SAVECONFIG.

---

Cluster Commands

FyroDB v0.2.0 implements the full suite of Redis Cluster commands:

Command	Description
CLUSTER INFO	Cluster state, assigned slots, node count, current epoch
CLUSTER MYID	Returns the unique 40-character hexadecimal node ID
CLUSTER NODES	Complete cluster topology and node state in nodes.conf format
CLUSTER SLOTS	Nested array mapping slot ranges to master and replica endpoints
CLUSTER SHARDS	Modern shard-level topology description with health status
CLUSTER KEYSLOT <key>	Computes the CRC16 hash slot for a key
CLUSTER COUNTKEYSINSLOT <slot>	Returns the number of keys currently stored in a slot
CLUSTER GETKEYSINSLOT <slot> <count>	Returns up to count keys present in a slot
CLUSTER MEET <ip> <port> [cport]	Connects to a new node and initiates cluster gossip handshake
CLUSTER FORGET <node-id>	Removes a node from the cluster state table
CLUSTER ADDSLOTS <slot...> / ADDSLOTSRANGE	Assigns slot ranges to the receiving node
CLUSTER DELSLOTS <slot...> / DELSLOTSRANGE	Removes slot ranges from the receiving node
CLUSTER SETSLOT <slot> <action>	Sets slot status to MIGRATING, IMPORTING, STABLE, or NODE
CLUSTER REPLICATE <node-id>	Configures the receiving node as a replica of a target master
CLUSTER RESET [HARD|SOFT]	Resets node cluster configuration and node identity
CLUSTER SAVECONFIG	Forces an immediate write of nodes.conf to disk
CLUSTER FAILOVER [FORCE|TAKEOVER]	Manually initiates a replica failover
CLUSTER BUMPEPOCH	Advances the node configuration epoch
ASKING	Sets the connection asking flag to permit execution after ASK redirects

---

Storage & Concurrency Performance

Tagged-Pointer Hash Probing

• In customhash::CustomMap, each slot's state word embeds a 7-bit hash tag in previously unused high bits.
• During linear probing, entries that do not match the key's hash tag are rejected immediately without fetching the map entry into CPU cache lines, reducing memory bandwidth pressure during lookups.

Contention Spin Backoff

• Per-key spinlocks now implement exponential backoff with thread yields when a key is heavily contended by multiple worker threads.
• Eliminates CPU core thrashing and negative throughput scaling on hot keys under 100+ concurrent clients.

Flat Routing Table & Standalone Fast Path

• Slot ownership is looked up in a shared 16,384-entry flat routing table.
• Worker connections cache a local view and validate freshness with a single atomic load (epoch == local_epoch), eliminating locking overhead on the request dispatch hot path.
• When cluster mode is disabled (ClusterRouting::Disabled), all cluster checks are bypassed completely with zero branching penalty.

Allocation-Free Pops & Lazy Connection Buffers

• LPOP and RPOP serialize the popped element directly into the connection's TCP output buffer without temporary Vec or string allocations.
• Accepted idle connections defer I/O buffer allocation until the first frame is received, slashing baseline memory usage when maintaining thousands of idle connections.

Unified Key-Capacity Admission

• Centralized key limit accounting across all hash map shards (src/storage/capacity.rs).
• FYRODB_MAX_KEYS defaults to 0 (unlimited live keys, matching Redis's default maxmemory 0).

---

Benchmarks

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

Whole-Machine: FyroDB (1 node, 12 workers) vs. Redis Cluster (12 masters)

Benchmark	FyroDB (1 node, 12 workers)	Redis Cluster (12 masters)	Speedup
Pipeline-64 SET	9.00M ops/sec	5.26M ops/sec	1.7×
Pipeline-100 SET	22.12M ops/sec	7.39M ops/sec	3.0×
Pipeline-100 GET	26.61M ops/sec	8.26M ops/sec	3.2×
Pipeline-100 LPUSH	18.73M ops/sec	6.54M ops/sec	2.9×
Pipeline-100 LPOP	20.45M ops/sec	6.81M ops/sec	3.0×
Pipeline-100 SADD	19.82M ops/sec	7.11M ops/sec	2.8×
Pipeline-100 ZADD	10.24M ops/sec	4.35M ops/sec	2.4×

Memory Footprint

State	FyroDB (1 node)	Redis Cluster (12 masters)
Idle RSS	5 MB	~120 MB (total)
Loaded RSS (1M keys)	57 MB	120 MB (total)
After FLUSHALL	16 MB	120 MB (total)

---

Upgrade Notes

• Default Key Capacity: FYRODB_MAX_KEYS now defaults to 0 (unlimited keys). Set it explicitly if a fixed key ceiling is required.
• Cluster Mode: Disabled by default. Enable cluster mode by setting FYRODB_CLUSTER_ENABLED=true or providing --cluster-enabled yes.
• Backward Compatibility: Fully compatible with existing RDB snapshot files from v0.1.1 and v0.1.2.
• Client Compatibility: Works with standard Redis standalone clients and cluster clients (redis-py, ioredis, go-redis, jedis, redis-rs, redis-cli -c).