Summary
The periodic Housekeeping deleted-sweep purge runs a full world save and region-file deletion on the main server thread, producing a large MSPT spike each time it fires. Observed while spot-checking an unrelated settings-GUI perf fix (#3016): the GUI itself was no longer the cost, but a purge landing mid-measurement produced a ~181 ms /mspt max.
Evidence (server log)
08:38:32 [BentoBox] Housekeeping: saving all worlds before purge...
08:38:32 [BentoBox] Housekeeping: world save complete
08:38:32 [BentoBox] Housekeeping deleted sweep: starting across 2 gamemode(s)
08:38:32 [BentoBox] Purge deleted-sweep: 16 candidate region(s) ... caveblock-world
08:38:32 [BentoBox] Purge deleted: evicted 0 loaded chunk(s) from 12 target region(s)
08:38:32 [BentoBox] Now deleting region files for world caveblock-world
08:38:32 [BentoBox] Purge complete for world caveblock-world: 12 region(s) ...
08:38:32 [BentoBox] ... scanning 'Gusher' world 'gusher_world' ...
08:38:32 [BentoBox] Housekeeping deleted sweep: complete — 2 world(s) processed, 13 region(s) purged in 0s
/mspt immediately after (1-minute window still contains the purge):
08:38:49 avg/min/max, 5s/10s/1m: 7.8/1.5/72.0, 8.6/1.5/72.0, 8.1/1.4/181.0
/mspt before the purge (no purge in window): 1m max was 79.5 ms; the most-recent 5s max was 17.6 ms. The 181 ms peak coincides with the purge.
Suspected cost
saveWorld / "saving all worlds before purge" is a synchronous main-thread save.- Region-file deletion (
Now deleting region files for world ...) is main-thread file I/O.
- Chunk eviction from target regions.
All of this runs in one tick (log says "13 region(s) purged in 0s"), so it lands as a single heavy tick.
Ideas to investigate
- Spread the purge across ticks (budgeted per-tick region deletion) instead of one synchronous burst.
- Move region-file deletion and the pre-purge save off the main thread where the API allows (file deletion is filesystem-only; the save may need main-thread coordination but could use Paper async chunk-save paths).
- Rate-limit / batch region deletion (N regions per tick).
- Confirm whether "saving all worlds" is strictly necessary before every sweep, or can be scoped to affected worlds only.
Repro
Run a server with housekeeping deleted-sweep enabled and deletable islands present; watch /mspt (or spark) when the sweep fires. Relevant code: HousekeepingManager / island deletion + region purge path.
Notes
Found while working on #3016 (settings-GUI click-spam MSPT). Filed for later — not part of that PR.
Summary
The periodic Housekeeping deleted-sweep purge runs a full world save and region-file deletion on the main server thread, producing a large MSPT spike each time it fires. Observed while spot-checking an unrelated settings-GUI perf fix (#3016): the GUI itself was no longer the cost, but a purge landing mid-measurement produced a ~181 ms
/msptmax.Evidence (server log)
/msptimmediately after (1-minute window still contains the purge):/msptbefore the purge (no purge in window): 1m max was 79.5 ms; the most-recent 5s max was 17.6 ms. The 181 ms peak coincides with the purge.Suspected cost
saveWorld/ "saving all worlds before purge" is a synchronous main-thread save.Now deleting region files for world ...) is main-thread file I/O.All of this runs in one tick (log says "13 region(s) purged in 0s"), so it lands as a single heavy tick.
Ideas to investigate
Repro
Run a server with housekeeping deleted-sweep enabled and deletable islands present; watch
/mspt(or spark) when the sweep fires. Relevant code: HousekeepingManager / island deletion + region purge path.Notes
Found while working on #3016 (settings-GUI click-spam MSPT). Filed for later — not part of that PR.