NMS-19573: Creating metadata cache

[christianpape]

• JIRA: https://opennms.atlassian.net/browse/NMS-19573

[marshallmassengill]

I updated the title to fix the typo on the NMS issue, just FYI. I'm reviewing this one now.

[marshallmassengill]

A few things for the review:

There are some updates for F2024 that will make it so this doesn't build as it is.

Something to check is where this is wired in. I think Core gets this from Spring (applicationContext-entity-scope-provider.xml declares plain EntityScopeProviderImpl as the primary bean). Suggestion from Claude here:

The cache belongs at the Spring primary bean, where it also helps pollerd's residual per-poll builds after NMS-20009 and collectd's per-cycle parameter interpolation. Also, look at NMS-20007, where thresholding will build at most one scope per service per cycle.

It's got unbounded caches of Hibernate entity graphs by default. Default maximumSize=-1 means four unbounded caches, each value an ObjectScope pinning a loaded node, its ~60-field asset record, and metadata collections for a rolling 10 minutes. This should probably be a bounded default, and ideally the cached value should be a materialized MapScope snapshot rather than entity-wrapping scopes to also remove the detached-lazy-proxy hazard.

EntityScopeProviderImpl.getScopeForNode(null) returns EmptyScope.EMPTY; the decorator calls nodeScopes.getUnchecked(null) and Guava throws NPE on null key.

CachedEntityScopeProviderIT extends EntityScopeProviderIT and constructs the decorator with expireAfterWrite=0 — immediate expiry, caching effectively disabled. It proves delegation preserves semantics, which is worth having, but the suite passes even if the cache never caches. This is missing a second-lookup-doesn't-invoke-the-loader test, expiry behavior, and the null-key case.

refreshAfterWrite without an async reload() override refreshes synchronously on the caller's thread — the knob implies more than it delivers.

No recordStats()/metrics for what is explicitly a performance feature, so field hit-rates are unobservable.

maximumSize=0 also means unbounded (> 0 check) — doc trap.

Check the license header.

[marshallmassengill]

Not sure if this is intentional but the cfg file only tunes the blueprint instance; the Spring instance (Collectd, poller client, core thresholding) is hardcoded 900s and can't be disabled.

These seem like they are worth looking at:

• refreshAfterWrite=0 throws at construction (Guava requires > 0; guard is >= 0).
• testCache is timing-sensitive (5s window across DB saves) — a Ticker constructor arg would make it deterministic.

We also may want to put something in the release notes since this is a behavior change I would think.

[cgorantla]

This probably can be termed as metadata cache for all components not just for Thresholding.

[cgorantla]

This is only relevant in Sentinel