Optimize automatic cache persistence

[TypeThe0ry]

Summary

• Track cache modification and saved revisions so unchanged caches skip automatic persistence without losing changes made during an in-flight save.
• Mark use-times, revenue, history, favourites, and persistent placeholder mutations dirty.
• Feed dirty caches into one identity-deduplicated queue and drain it with a single cancelable async dispatcher instead of creating one delayed scheduler task per cache.
• Allow only one automatic save per cache at a time while preserving newer dirty revisions for a later cycle.
• Serialize saves per cache and only acknowledge a revision after every persistence step succeeds.
• Stop accepting new database work before shutdown waiting begins, so the wait is a real submission barrier.
• Bound database shutdown waiting to 30 seconds and dump named database worker stacks on timeout.
• Use Folia's async scheduler for delayed async work and align the auto-save.period-tick fallback with the configured 6000-tick default.

Root cause

Automatic saving queued a full rewrite for the global cache and every online player at once, even when nothing had changed. YAML serialization and disk I/O then competed for CPU and storage in one burst, causing periodic MSPT spikes. A per-cache delayed-task approach would also leave many tasks to track and could race with shutdown.

Impact

Most save cycles now avoid untouched players entirely. Dirty players are deduplicated and spread over time by one dispatcher, overlapping automatic saves are coalesced, and shutdown cannot accept new asynchronous database work after its drain phase begins.

Validation

• git diff --check
• Local compilation intentionally skipped per requester instruction.
• GitHub Actions build passed for the latest head, including Gradle build and artifact upload.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.