perf(core): eliminate avoidable per-call allocation in hot logging/config paths

[swicken]

Proposed Changes

• Logger lazy Supplier overloads are now actually lazy — the three debug(…, Supplier) and two info(…, Supplier) overloads called message.get() unconditionally; they now check the level first. ~900 call sites use suppliers specifically to skip message building, including ones that did reflectionToString on Contentlets via PermissionBitFactoryImpl debug logging on the reindex path.
• Config int/long/float/boolean getters no longer throw per read — the DOT_* env-alias probe used commons-configuration's throwing 1-arg getter wrapped in vavr Try, so every read of a non-env-overridden key allocated and discarded a NoSuchElementException (with full stack trace). Applied the containsKey guard that getBooleanProperty(name, default) already had.
• Logger.isVelocityMessage memoized via ClassValue — previously allocated a lowercased class name on every Class-keyed log call; ClassValue lookups are allocation-free, keyed per actual Class (no FQN conflation across OSGi classloaders), and entries are released with their class.
• LeakyTokenBucketImpl no longer walks Config per request-cost increment — the four config reads (enabled, refill, max bucket, add-header-info) are TTL-memoized (30s). Runtime overrides via system table / config refresh still take effect within the TTL, preserved deliberately so rate limiting can be toggled during an incident; a test with an injectable TTL pins that contract.

Why

Allocation profiling of a running instance (Pyroscope TLAB profiles under mixed REST/GraphQL/page-render load) attributed the Throwable.fillInStackTrace, Method.copy/Field.copy, and StringLatin1.toLowerCase allocation hotspots to these four spots; Config.get* subtrees alone were ~19% of sampled allocation in a mixed window (e.g. one full Config walk — exception included — per expired session via ClickstreamListener, per 250ms reindex-pause iteration, per request-cost increment). Expected effect is reduced allocation → less GC pressure, not a single-endpoint latency win.

Checklist

• Tests
• Translations
• Security Implications Contemplated (none — no input handling or auth paths touched; rate-limit values remain config-driven)

Additional Info

New/extended unit tests: LoggerTest (suppliers not evaluated when level disabled), LeakyTokenBucketImplTest (runtime config override reaches a running bucket; reads within TTL stay cached), ConfigTest (getLongProperty missing-key default; env-alias-present-but-unparseable falls back to default for int/long/float). The deprecated no-default Config getters still throw NoSuchElementException for missing base keys — that documented contract is unchanged (and pinned by existing tests).

Closes #36145

[github-actions]

🤖 Codex Review — openai.gpt-5.5

**[> [🔴 Critical] dotCMS/src/main/java/com/dotcms/cost/LeakyTokenBucketImpl.java:121 — The new debug lambda captures currentCount, but currentCount is reassigned inside the retry loop, so it is not effectively final and this class will not compile. Copy it to a final local before the lambda or keep the eager string logging.

---

_{Run: #27548822095 · tokens: in: 4536 · out: 4535 (reasoning: 4447) · total: 9071}

[claude]

Claude finished @swicken's task in 1m 11s —— View job

---

Rollback Safety Analysis

• Read rollback-unsafe categories reference
• Get full PR diff
• Analyze diff against unsafe categories
• Apply appropriate label

---

Result: ✅ Safe To Rollback

All changes in this PR are pure in-memory performance optimizations with no impact on stored data, database schema, Elasticsearch mappings, or API contracts.

Files analyzed:

File	Change	Rollback Risk
Logger.java	Lazy Supplier overloads now check log level before calling message.get(); isVelocityMessage memoized via ClassValue	None — pure JVM behavior change, no stored state
Config.java	containsKey guard added before the throwing 1-arg getter in int/long/float/boolean env-alias probes	None — eliminates exception allocation, behavior semantics unchanged
LeakyTokenBucketImpl.java	Four Config reads replaced with TTL-memoized Supplier instances (30s expiry via Guava)	None — same config keys, same defaults, same runtime behavior within TTL window
Test files	New unit tests for Logger, Config, LeakyTokenBucketImpl	N/A

Categories checked and cleared:

• C-1 Structural Data Model — no runonce tasks, no schema changes ✅
• C-2 ES Mapping — no ESMapping* changes ✅
• C-3 Content JSON model version — no CURRENT_MODEL_VERSION change ✅
• C-4 DROP TABLE/COLUMN — none ✅
• H-1 through H-8 — no data migrations, renames, PK changes, field types, storage providers, stored procedures, NOT NULL columns, or VTL viewtool contract changes ✅
• M-1 through M-4 — no column type changes, bundle format changes, REST/GraphQL contract changes, or OSGi interface changes ✅

The label AI: Safe To Rollback has been applied.

[swicken]

🤖 Codex Review — openai.gpt-5.5

**[> [🔴 Critical] dotCMS/src/main/java/com/dotcms/cost/LeakyTokenBucketImpl.java:121 — The new debug lambda captures currentCount, but currentCount is reassigned inside the retry loop, so it is not effectively final and this class will not compile. Copy it to a final local before the lambda or keep the eager string logging.

Run: #27548822095 · tokens: in: 4536 · out: 4535 (reasoning: 4447) · total: 9071

Not even sure how it came up with this :) There is no retry loop, it must be hallucinating.