Summary

prewarm() scheduled 40 dummy tasks far in the future, cancelled them, and purged the queue, in order to pre-grow the single-threaded ScheduledThreadPoolExecutor's work queue during Sentry.init. On-device measurement shows the cost it avoids is negligible-to-zero, so this removes it (and the now-unused constants) along with its two init call sites.

Why it's safe to remove

prewarm's only job was to pre-grow the executor's work queue so a later resize wouldn't happen "in an unexpected area of the SDK". Measured on a Galaxy A55 (Android 16), growing the queue the full 16→54 (all three array resizes) costs ~8µs total — and that is the ceiling: init only ever queues a handful of tasks, well under the default 16-slot capacity, so in practice no resize happens and the saving is 0µs.

PrewarmBenchmarkTest, n=1000 × 3 runs, median µs. Both paths schedule 40 tasks into an empty queue (identical per-insert work); they differ only in whether the backing array resizes:

It also can't help, and slightly regresses init

The executor is single-threaded and prewarm is submitted ahead of the first useful task (loadLazyFields, submitted right after in Sentry.init), so the 40-task loop can only delay that task, never speed it up. Thread creation and executor class-loading are paid by whichever task runs first, so loadLazyFields warms them identically with or without prewarm.

Measured end-to-end, the time until the first useful task finishes (from a fresh executor, n=400 × 3 runs) is ~100µs slower with prewarm than without, at every percentile; a plain submit(noop) is indistinguishable from doing nothing:

(PrewarmBenchmarkTest also records the per-op loop cost and the process-once class-loading cost, which is unavoidable regardless of prewarm; both confirm the above.)

Changes

• Remove prewarm() from ISentryExecutorService, SentryExecutorService (and the now-unused INITIAL_QUEUE_SIZE / dummyRunnable), and NoOpSentryExecutorService.
• Remove the two init call sites (Sentry.init re-init branch and SentryOptions.activate()).
• Drop the prewarm unit test and mock overrides; add PrewarmBenchmarkTest.

Original prewarm PR: #4606 — its only stated reason was the queue-growth cost measured above.

🤖 Generated with Claude Code