perf(precompute): jemalloc allocator + Arc<str> group-key interning (2.2× on KLL window-close)

[zzylol]

What

Two CPU optimizations for the precompute engine hot path, found by profiling the tumbling-window KLL sketch workload (perf -F499, 4000 groups). Builds on the window-close move (#379, now in main).

1. jemalloc global allocator (tikv-jemallocator, default jemalloc feature, #[global_allocator] in lib.rs). The hot path was dominated by malloc/free churn (43% of CPU) from per-window sketch buffers; jemalloc's per-thread arenas / size-class caching recycle them, dropping allocator cost to ~7%.
2. Arc<str> group-key interning. group_states is now HashMap<u64, HashMap<Arc<str>, GroupState>>, so the per-sample/per-batch lookup borrows by &str with zero allocation; the key string is allocated once per group. Also removes a redundant per-batch hashmap lookup in process_group_samples.

Why

Profiling showed allocator churn was the #1 CPU cost. jemalloc is the single biggest lever; the key interning removes the remaining per-sample String allocation/clone on the lookup path.

Measured

High cardinality, 4000 groups, 2.4M window-closes, mean of 3 reps, relative to the pre-#379 clone + system-malloc baseline:

Variant	Wall	Throughput	Speedup
clone + system malloc	7.95s	301k closes/s	1.00×
move (#379) + system malloc	6.33s	380k closes/s	1.26×
move + jemalloc	3.71s	646k closes/s	2.14×
move + jemalloc + Arc<str>	3.61s	665k closes/s	2.20×

Testing

• All 473 lib tests pass (cargo test --release -p query_engine_rust --lib).
• Production binary (precompute_engine) builds with jemalloc.

Not in scope (follow-ups)

• fmod/ceil in sketchlib_kll_update (~17%) — algorithmic KLL level-math in the sketch library.
• Keyed aggregations (CMS / HydraKLL / MultipleSum) allocate a KeyByLabelValues per sample in extract_aggregated_key_from_series; the single-subpopulation KLL workload doesn't exercise this.

🤖 Generated with Claude Code