Actuator and Micrometer observability

Latest stable Based on Bluetape4k release 1.11.0

Enable the endpoint

The nearcache endpoint bean is registered when Actuator is on the classpath and an EntityManagerFactory bean exists. Add it to management exposure for HTTP access.

management:
  endpoints:
    web:
      exposure:
        include: health,info,metrics,nearcache

GET /actuator/nearcache
GET /actuator/nearcache/{regionName}

The first returns Map<String, RegionStats>. A missing individual region returns null.

Read RegionStats

Field	Source	Nullable behavior
`regionName`	RegionFactory	Never null
`localSize`	Caffeine local cache	`0` if no cache is available
Local hit rate, hit/miss/eviction	Caffeine stats	`null` when local stats are disabled
L2 hit/miss/put	Hibernate statistics	`null` when statistics are off or lookup fails

Distinguish 0 from null: zero may be a measured value, while null means that statistic was not collected.

The endpoint wraps factory unwrapping and statistics lookup in runCatching. A different RegionFactory or an unwrap failure yields an empty map or null. Endpoint output alone does not prove backend health.

Micrometer gauges

After singleton initialization, the binder registers:

lettuce.nearcache.active.regions: current RegionFactory cache-map size
lettuce.nearcache.total.local.size: total local entries across all regions

The gauges read current RegionFactory state when sampled, so new regions affect the value. They do not provide per-region hit rate or Redis latency.

GET /actuator/metrics/lettuce.nearcache.active.regions
GET /actuator/metrics/lettuce.nearcache.total.local.size

Dashboard signals

Aggregate gauges can hide a hot region’s eviction spike. Combine them with:

active regions and local size per application instance
Hibernate L2 hits, misses, and puts per region
Redis command latency, errors, and reconnects
database query count and latency
cache configuration and deployment events

Local eviction cleanup may be asynchronous. The 1.11.0 integration test therefore checks a non-negative local size rather than an exact instantaneous count. Use trends and correlation with database load.

Registration failures

The binder logs at debug and registers nothing when a different RegionFactory is active. Unwrap or gauge-registration failures produce a warning but do not fail startup. Alert separately on missing observation and cache-backend failures so operators can distinguish the cause.