cpu-memory-review

name	cpu-memory-review
description	Use when reviewing CPU or memory performance findings; index of per-scenario decision rules for accepted patterns (lazy cleanup, mandatory .to_string() API, single-shot allocations, etc.).

CPU / Memory Performance Review

Files in this directory

File	Purpose
performance-review-methodology.md	General decision methodology — trace call chain, distinguish phases, verify mechanism
not-a-performance-issue.md	Legacy stub — contents split into cases/; maintained for link compatibility
cases/	Individual per-scenario closed-case files (one decision rule per file)

Quick routing

• Is this a general "is it a performance issue?" question? → performance-review-methodology.md
• Looking for a specific named scenario? → see the case table below or search cases/
• Checking a finding against a known accepted tradeoff? → find the matching row and load that file

Closed findings are recorded one-file-per-finding under the relevant skills/04-review/<topic>/ directory; use grep/ls to find them.

Case Index

File	Description
cases/metrics-label-to-string-api.md	.to_string() in metrics recording functions — mandatory for Cow<'static, str> API
cases/control-plane-status-write-counter-pre-caching.md	Pre-caching Counter handles for record_status_write on controller plane — not worth it
cases/get-raw-any-lookup-not-json-scan.md	get_raw_any / lookup_gateway is O(1) DashMap lookup, not JSON serialize + scan
cases/http-grpc-route-regex-matching-no-fallback.md	Hot-path Regex::new() fallback in match_header / match_query_param — removed; no re-add
cases/watcher-vec-lazy-cleanup.md	RwLock<Vec<WatchClient>> lazy cleanup — sufficient; bounded < 20KB
cases/webhook-read-then-truncate-16kb-cap.md	read_body_limited 1GB buffering claim — actual hard cap is 16KB
cases/oncelock-global-singleton-deliberate-design.md	GLOBAL_CONFIG_CLIENT OnceLock not releasing on reconnect — deliberate design
cases/small-object-clones-reconcile-config-sync.md	Small-object clones / config.clone() on reconcile / config-sync paths — not per-request
cases/oidc-jwt-claims-json-serialization-opt-in.md	OIDC/JWT claims serde_json::to_string per request — serialization is the opt-in feature itself
cases/extract-status-json-double-serialization-cold-path.md	extract_status_json called twice in process_resource_inner — cold path + API shape
cases/config-struct-field-ordering-padding-waste.md	Config struct field reorder for padding — field order is part of user-visible JSON schema
cases/hashmap-new-default-capacity-reconcile-paths.md	HashMap::new() without with_capacity on full_set / reconcile paths — cold event-level, tiny scale
cases/access-log-json-serialization-per-request.md	AccessLogEntry::to_json() per-request String alloc — channel boundary shape requires owned String
cases/access-log-request-uri-include-raw-uri-alloc.md	AccessLogEntry::from_context per-request URI String alloc under include_raw_uri=true — dwarfed by accepted to_json() alloc; proposed entry_raw_uri cache saves nothing
cases/record-cross-ns-refs-write-lock-per-item.md	record_cross_ns_refs one write lock per backend_ref — cold reconcile path, < 2.5 µs total
cases/bidirectional-ref-manager-no-gc-not-memory-leak.md	BidirectionalRefManager / SecretRefManager lacking GC — it is an index, not a cache
cases/workqueue-delay-heap-no-size-cap.md	Workqueue BinaryHeap<DelayedItem> no explicit size cap — dedup by key + per-Kind bounds
cases/conf-sync-watcher-mpsc-chain-not-backlog.md	conf_sync watcher mpsc(100) × 3 layers — items are batches, not single events; end-to-end backpressure
cases/status-diff-conditions-equal-on2-bounded.md	conditions_equal O(n²) scan — n ≤ 10; HashMap overhead exceeds linear scan benefit
cases/admin-state-cache-get-resource-full-list-scan.md	cache_get_resource full-list + linear-scan — admin/diagnostic path, bounded by human ops
cases/conf-sync-compress-event-vec-per-key.md	CompressEvent per-key Vec<ResourceChange> "unbounded / OOM / O(N) scan" — confirmed-real but won't-fix; control-plane, 1-byte enum, trivially bounded
cases/udp-established-session-route-pipeline-per-packet.md	UDP handle_client_packet ran full route pipeline per packet for established sessions — FIXED (F-016) with a client_sessions fast-path
cases/dsl-step-budget-not-cpu-bound.md	DSL VM max_steps counts opcode dispatches, not CPU time — per-opcode CPU is bounded by max_string_len + per-regex DFA limit + per-builtin output caps; the four-cap product is the runtime CPU bound
cases/gcir-stream-plugin-store-key-precompute.md	GlobalConnectionIpRestrictionPlugin per-connection format!("{}/{}", ns, name) in PM-store lookup — FIXED (F-008) by precomputing the key in new() and calling the _by_key getter on the accept path
cases/routes-http-listener-manager-per-request-lookup.md	pg_request_filter per-request get_or_create_listener_manager(&listener_key) — DashMap shard write-lock + ListenerKey(String) clone on hot path — FIXED (F-001) by caching Arc<HttpPortRouteManager> / Arc<GrpcPortRouteManager> on EdgionHttpProxy at construction
cases/domain-route-rules-inner-arcswap-removed.md	DomainRouteRules inner ArcSwap / ArcSwapOption fields paid 3 redundant per-request atomic loads — FIXED (F-003) by switching to plain Arc / Option<Arc<…>> because the outer ArcSwap<DomainRouteRules> already provides swap granularity
cases/invalid-rule-indices-vec-contains-config-load.md	HTTP/gRPC route compile loops call Vec::contains on invalid_rule_indices per rule — config-load path only, `
cases/tcp-log-entry-status-debug-format-replaced.md	TcpLogEntry::from_context per-connection format!("{:?}", TcpStatus) for a 12-variant unit enum — FIXED (observe-F012) by TcpStatus::as_str(&self) -> &'static str + locking test; the structural win is decoupling the access-log schema from the Debug impl