Redis

cosid-store-redis: RangeStore and SlotLeaseStore as Lua scripts, run through the minimal RedisScripts interface by Lettuce. It has no Spring dependency (D6); the Spring Data adapter is in the autoconfigure module. Redis 7.0 or later (D27). 6.2.24 passed the TCK once, which is not a support commitment.

Durability precondition

Uniqueness rests on "an acknowledged write is never lost" (已确认的写入不会丢失)

Run one of:

• a single node with AOF and appendfsync always, or
• replicas with WAIT covering every replica (options.wait set to the replica count).

Under asynchronous replication a failover can lose range and lease writes, and then hand out duplicate ids. The store uses a range or a ceiling only after its write is acknowledged, but it cannot make Redis keep it. For the same reason the keys must never be evicted: keep maxmemory-policy noeviction (the default). Where uniqueness is strict, prefer JDBC or ZooKeeper.

Driver version

cosid-store-redis is built and tested on Lettuce 7.9.0, which brings Netty 4.2.17; Spring Boot 4.1.2 manages Lettuce 7.5.2. Which one an application gets depends on its build. On Spring Boot 4.1.2 (CI job published-jars, run 40881059003):

Build	Lettuce
Maven, spring-boot-starter-parent or the BOM imported	7.5.2	the BOM's version wins; Netty is the BOM's 4.2.19
Gradle, platform("org.springframework.boot:spring-boot-dependencies")	7.9.0	the higher version wins

The store runs on both. For Lettuce 7.9.0 with Maven, set <lettuce.version>7.9.0.RELEASE</lettuce.version>.

Configuration

cosid:
  stores:
    redis:
      type: redis
      timeout: 1s           # every command, and the WAIT timeout
      options:
        uri: redis://redis:6379
        wait: auto

Option	Default	Meaning
uri		a Redis URI, when no client is given
cluster	false	uri names a Redis Cluster
wait	auto	auto: WAIT 1 after every write when the server has replicas (INFO replication at open; in a cluster, when every primary has one); off; or the replica count
sweepBatch	16	expired leases one acquire moves back to the free set, at most

What the store does with the client it is given (StoreConfig.client, or the context's bean in Spring):

Client	The store
RedisScripts (the Spring Data adapter, for a RedisConnectionFactory)	uses it as it is, WAIT included
RedisClient / RedisClusterClient	opens a connection of its own and closes only that one
StatefulRedisConnection / StatefulRedisClusterConnection	runs on it and never closes it; wait must resolve to 0, since WAIT must not block a shared connection
none	creates a client from uri and closes it

A write that fewer replicas acknowledge fails with RedisStoreException. Its outcome is unknown, so acquire, renew and reserve may be retried; a lost range is only a gap.

Keys

Each group of keys shares one hash tag, so every script runs on one Cluster node. Keys are never deleted (D23).

Key	Type	Content
cosid:{key}	string	the next value of a range key, decimal (exact beyond 2^53)
cosid:{key}:sealed	string	the seal marker
cosid:{ns}:meta	hash	limit, tolerance of the namespace
cosid:{ns}:slot	hash	slot → ownerKey|incarnation|lastOwnerKey|fence|expiresAt|floor|ceiling|nextFloor
cosid:{ns}:free	sorted set	free slots by number, prefilled 0..limit-1
cosid:{ns}:live	sorted set	live leases by expiresAt
cosid:{ns}:owner	hash	owner key → the slot it took last (sticky choice)
cosid:{ns}:held	hash	owner key → the slots whose row names it

An acquire sweeps at most sweepBatch expired leases into free, then tries the held, preferred, sticky and lowest free slot: a few hash and sorted-set operations, O(log n), never a scan (4.x scanned in Lua). The store clock is TIME inside the script; Redis 7 replicates script effects, so that is safe.