RFC (OpenSearch 3.6.0): LRUCache lock-path hot-path optimization

[fiatlux]

Versions (relevant - OpenSearch/Dashboard/Server OS/Browser):

• OpenSearch: 3.6.0
• Server OS: macOS (local dev env)

Describe the issue:

• LRUCache.get(...) already holds the cache lock, but on hit it calls incRef(key), which re-enters lock logic and repeats key lookup.
• This adds avoidable overhead on a very hot path in file-cache-heavy workloads.
• Proposed change:
  • Add private incRefNode(node) for already-resolved entries.
  • Use it in get, addNode, and replaceNode.
  • Keep public incRef(key) behavior unchanged (delegates internally to the node helper after lookup).
• Safety:
  • No API or wire/protocol changes.
  • No behavior changes to refcount/LRU/stat semantics, only removal of redundant work.
• Local microbenchmark result:
  • FileCacheBenchmark.get : 8101.39 -> 13393.85 ops/ms (+65.33%).

Configuration:

• Benchmark: FileCacheBenchmark.get
• Params: maximumNumberOfEntries=65536, concurrencyLevel=8
• JMH profile: -wi 3 -i 5 -w 2s -r 2s -f 2 -t 8

Relevant Logs or Screenshots:

• Patch screenshot attached

Feedback requested:

• Do you see any concurrency or eviction edge cases this change could miss?
• Any concerns that stats accounting could diverge from existing behavior?
• Are there additional workloads/benchmarks you want to see before proposing upstream?