[pditommaso]

What changes

Adds an optional L1 TTL to AbstractTieredCache, so a value rewritten in Redis by one instance reaches every other instance within a bounded time.

Today an L1 (Caffeine) entry is kept for as long as the value's own TTL. An instance that holds a key in L1 does not see a newer value another instance writes to L2 until that TTL passes. That's fine for values written once, but it breaks caches whose values are refreshed during their lifetime. For example, sched's CloudInfo snapshot store (seqeralabs/sched#1742) keeps each region's catalogue under a 7-day TTL as an outage safety net, and a cron on one pod rewrites it every hour. Without an L1 bound, every pod except the writer would keep serving its first copy for up to 7 days.

How

• New protected Duration getL1Ttl() hook, returning null by default, so existing caches behave exactly as before.
• When it is set, the L1 Caffeine cache uses expireAfterWrite(l1Ttl), and the next read after expiry rehydrates from L2. An entry's L1 lifetime is the shorter of the L1 TTL and the value's own TTL; L2 expiry is unchanged.
• The L1 TTL is included in the cache's startup config log line.
• README gains an "L1 TTL" section. The module goes to 1.2.0, with a changelog entry, so it is published on merge ([release]).

Testing

./gradlew :lib-cache-tiered-redis:test passes (Redis Testcontainer). Two new cases in AbstractTieredCacheTest:

• with an L1 TTL, a reader keeps its L1 copy until the L1 TTL passes, then serves the value another instance rewrote;
• without one, the reader keeps serving its L1 copy (the current default, pinned).

🤖 Generated with Claude Code

[jonmarti]

@pditommaso, approving:

• with no L1 TTL the cache is built exactly as before;
• the L1 TTL only bounds how long a copy is served locally (the value's own expiresAt check still applies);
• the removal listener, which now also fires on L1 expiry, only logs.

The first new test can't pass without the hook, since the reader would keep serving its first copy, as the second test pins. Two non-blocking notes inline.

[pditommaso]

Addressed both review notes in 69de38d:

• L1 TTL test timing widened (1s TTL, 1.5s sleep) to avoid CI flakiness.
• getL1Ttl() Javadoc now documents Duration.ZERO (L1 disabled, every read goes to L2).

./gradlew :lib-cache-tiered-redis:test passes.