Problem

ScheduledThreadPoolExecutor's internal DelayedWorkQueue is a min-heap that grows dynamically (50% per resize, starting from 16). prewarm() was introduced to pre-grow that array during Sentry.init — but init runs on the main thread, which is exactly when ANR risk is highest. The resize cost it avoids is only ~8µs; the prewarm itself adds ~100µs and is the more expensive operation.

More fundamentally, ScheduledThreadPoolExecutor's queue isn't replaceable, so there was no clean way to fix this without a custom implementation.

Solution

Replace SentryExecutorService with a custom implementation backed by:

• PriorityQueue<ScheduledTask<?>> pre-allocated to INITIAL_QUEUE_CAPACITY = 64 slots at construction — the backing array never resizes during normal SDK operation
• Single daemon worker thread that uses Object.wait(millis) / notifyAll() — sleeps precisely until the next task is due, wakes immediately when an earlier task is enqueued
• ScheduledTask<T> extends FutureTask<T> — gets the full Future<T> contract for free

prewarm() is now a documented no-op. The MAX_QUEUE_SIZE = 271 hard limit and purge-on-overflow behaviour are preserved unchanged.

What changed

Notes

• ./gradlew spotlessApply apiDump must be run locally before merging — sandbox can't run the JVM
• prewarm() is kept in the interface as @Deprecated for API compatibility; can be removed from ISentryExecutorService in a follow-up
• Supersedes / makes unnecessary: perf(core): Remove executor prewarm #5681

cc @romtsn

--

View Junior Session in Sentry