problem

1. Give the exporter's HttpServer an explicit bounded executor (httpServer.setExecutor(Executors.newFixedThreadPool(2))) so one slow scrape can't serialize/queue all others.
2. Add a short TTL/in-flight guard around updateMetrics() (e.g., skip recompute if last run was < N seconds ago, or synchronize so concurrent scrapes share one in-progress computation) so scrape frequency can never multiply backend load.
3. Instrument: log/measure updateMetrics() wall-clock time so the reporter (and CI) can confirm which sub-metric collector is actually slow and verify the fix closes the growth.

additional comments:

1. Stale dynamic config (CONFIRMED) — capacity.calculate.workers is a runtime-dynamic setting, but the new shared executor only reads it once at first creation; live changes are silently ignored until a restart.
2. Swallowed capacity-recalculation abort (CONFIRMED per the extra verify pass) — shutdown racing an in-flight recalculation throws RejectedExecutionException, caught by the blanket catch(Throwable) in recalculateCapacity(), silently skipping storage/IP/VLAN updates for that cycle.
3. Unsynchronized race on _capacityExecutorService (CONFIRMED per the extra verify pass) — can leak a freshly-recreated pool that's never shut down again.
4. Shared fixed-size pool serializes previously-independent callers (PLAUSIBLE) — rolling-maintenance host-drain gating can now queue behind the hourly timer or API-triggered recalculations.
5. Pool no longer bounded to actual task count, so it can stay oversized/stale relative to fleet size (efficiency).
6. Bundling this executor-lifecycle rewrite into what the reported bug (Scrape duration using prometheus is increasing continuously until restart #13586) only needed a one-line fix for (altitude/scope creep).
7. Inconsistent lazy-vs-eager thread-pool lifecycle pattern within the same class (reuse/convention).
8. Minor: the synchronized getter is called per-loop-iteration instead of hoisted once (efficiency).