Palworld Server Stability in 2026: Memory Leaks, Restart Schedules, and World Save Protection

Palworld’s server binary has a well-known characteristic that server owners learn the hard way: it allocates memory for base Pals, captured Pals, and world chunk data and never fully returns it to the OS. Over a 24–48 hour window the resident set size (RSS) climbs steadily. At some point the kernel OOM killer, or the process’s own internal limits, brings the session down mid-save. This article is about the operational controls that keep a Palworld server up: what the memory growth actually looks like, how to size RAM so restarts are predictable, how to schedule restarts without corrupting the world, and how to protect save data with versions you can actually roll back to.

Everything below assumes a dedicated box or a KVM VPS with a fixed RAM allocation and local NVMe storage — not shared hosting. If you want to see how RAM-to-player-count math works across hosting tiers, the dedicated server plan breakdowns on the main site are a reasonable reference point for sizing.

How Palworld Actually Leaks Memory

Memory growth in Palworld comes from three distinct places, and they behave differently:

  • Base worker Pals. Each working Pal in each base holds a state object and a pathfinding context. A base with 15 Pals is fine; a server with 30 players all running maxed bases with 15 Pals each is 450 active contexts plus the idle ones parked in Palbox storage.
  • Chunk streaming. The world is divided into chunks that load on demand. Unloading is best-effort. Players who travel widely (exploring, boss hunting) cause more chunks to be retained than players who stay home.
  • Save serialization. The periodic autosave builds a full serialized copy of the world in memory before writing. On large worlds, that transient allocation can be 1–2 GB on top of steady-state RSS.

The practical upshot: a 4-player server may sit at 2.5 GB and grow to 3.5 GB over a day. A 32-player server may start at 7 GB and hit 16 GB by hour 30. Growth rate scales roughly with player count and with base density, not with map exploration as much as people assume.

Measuring RSS Is Not Optional

You need a sampled time series, not a spot check. Log RSS every 60 seconds with a one-liner cron job:

*/1 * * * * ps -o rss= -p $(pgrep -f PalServer-Linux-Shipping) | awk '{print strftime("%F %T"), $1/1024 " MB"}' >> /var/log/palworld/rss.log

After 48 hours you have a growth curve. Fit a straight line to it in a spreadsheet: the slope in MB/hour is your leak rate. That single number drives every other decision in this guide.

Choosing Your Restart Interval From the Leak Rate

Do not use a generic “restart every 4 hours” rule. Derive it. If your leak rate is L MB/hour, your process starts at S MB after a warm-up period, and you want a safety margin M of 20% of total RAM R, then:

hours between restarts = (0.8 × R − S) / L

Worked example: 16 GB box, warm start 7 GB, leak 280 MB/hour. (0.8 × 16384 − 7168) / 280 = (13107 − 7168) / 280 ≈ 21.2 hours. Round down to 20 hours or, more practically, to 12 hours so restarts land at predictable times twice a day instead of drifting.

A quick sizing reference for the RAM side of that equation:

Concurrent playersTypical warm RSSObserved leak rateMinimum RAMRecommended RAM
1–42.0–3.0 GB60–90 MB/h6 GB8 GB
5–124.0–5.5 GB120–180 MB/h10 GB12 GB
13–246.0–8.0 GB200–320 MB/h14 GB16 GB
25–328.0–11.0 GB300–450 MB/h20 GB24 GB
Numbers from community-reported Palworld 0.4.x–0.6.x builds on Linux; your mods and base density will shift them.

Note that CPU scales far more gently than RAM. A modern 4-core/8-thread Xeon or Ryzen handles 16–24 players with tick time around 25–40 ms. Do not over-buy cores based on player count; buy RAM and NVMe, then buy single-thread speed if tick time is your bottleneck.

Driving Restarts Without Losing the World

A restart is only dangerous when the save is in flight. The sequence that avoids corruption has four steps and must be scripted, not typed by hand each night:

  1. Send an in-game broadcast warning 120, 60, and 15 seconds ahead. Palworld’s own /Broadcast command works, but the server RCON interface is more reliable because it does not depend on a GM being logged in.
  2. Send Save via RCON and wait — do not proceed until the RCON reply returns. Serialization of a large world can take 5–20 seconds.
  3. Send Shutdown via RCON, or SIGTERM to the process. Give it 30 seconds. Do not follow with SIGKILL faster than that.
  4. Verify the new save file exists and is larger than the previous one before starting the process again.

Step 4 is the one people skip. A truncated save passes a naive existence check but fails on next boot with a world-load error. Comparing file size and modification timestamp costs nothing and catches the failure mode that matters.

World Save Protection: Rotation, Checksums, Offsite

Palworld’s save directory (Pal/Saved/SaveGames/0/<WorldGUID>/) holds the level file, player files, and a players mapping. Every restart should produce a snapshot; keep at least seven rotations and at least one copy offsite.

A rotation scheme that survives the common disasters:

  • Pre-restart snapshot — taken after step 3 above, before the process restarts. Keep 7.
  • Hourly snapshot — cheap if your filesystem supports reflinks (Btrfs, ZFS) or copy-on-write snapshots. Keep 24.
  • Daily snapshot — copied to a second disk on the same machine, or better, pushed to object storage. Keep 30. This is what saves you from “the last week of saves were all corrupt because the corruption happened before we noticed.”
  • Weekly snapshot — offsite in a different city or region. Keep 8. Ransomware, host account compromise, and disk-destroying provider mistakes all defeat on-box rollback.

Tar the save directory with --exclude on the backup subfolder and pipe through zstd -3. A 300 MB world compresses to roughly 40–60 MB, which makes 30 daily offsite copies cost almost nothing in storage.

Watchtower: Deciding When to Intervene Mid-Session

A scheduled restart at 04:00 is not always enough. If a session is unusually long — a weekend raid night with 30 concurrent players for eight hours — RSS can exceed your projection. Set two external guards:

  • Soft ceiling at 75% of RAM. Log a warning, post to your Discord, and increase the frequency of RCON saves so that the most recent snapshot is never more than five minutes old.
  • Hard ceiling at 88% of RAM. Trigger an emergency restart using the same save → verify → restart sequence. Do not let the OOM killer choose the moment; the kernel does not save first.

Both guards should be external to the game process. A Prometheus node_exporter plus a simple alert rule on process_resident_memory_bytes is enough. Do not rely on the game server’s own monitoring to report that it is about to die.

The Settings That Reduce Memory Pressure

Several server settings have a disproportionate effect on RSS versus tick rate:

  • BaseCampMaxNumInGuild — lowering from 4 to 3 cuts base-related contexts meaningfully on populated servers. Players complain about it, so this is a negotiated change, not a default.
  • bIsUseBackupSaveData — keep enabled. It costs a little I/O and saves whole afternoons.
  • AutoSaveSpan — the default 30 seconds is aggressive on large worlds. 120–180 seconds reduces serialization churn and transient memory spikes, at the cost of a slightly older save on a crash.
  • DropItemMaxNum and DropItemAliveMaxHours — unbounded dropped item accumulation in a busy base area is a slow memory and entity-count tax. Cap it.

Storage and Filesystem Choices That Prevent Corruption

Save corruption on Palworld is almost always a crash-during-write problem compounded by a write cache. Two defensive measures matter more than the filesystem choice itself:

  • NVMe with power-loss protection, or at minimum a consumer NVMe with fstrim and a high commit interval. On ext4, data=ordered (the default) plus periodic sync is fine; do not use data=writeback for a game server.
  • Do not put the save directory on a network filesystem. NFS or SMB with a cache and a flaky link is a reliable way to produce half-written level files.

If your host’s storage behaviour is unknown to you, that is a reason to ask. The hardware notes on bestdedicatedwebhostingserver.com cover disk types and commit behaviour per plan, which matters more for Palworld than for most game servers.

A Twelve-Point Checklist Before You Announce the Server

  • RSS sampled every 60 s, log rotated and retained for at least 14 days.
  • Leak rate measured over 48 hours, restart interval derived, not guessed.
  • Restart script: broadcast → save (wait for RCON reply) → shutdown → verify → boot.
  • Seven pre-restart save rotations, verified non-empty.
  • Hourly local snapshots if the filesystem supports them.
  • Daily snapshot copied off the game disk, 30 retained.
  • Weekly offsite copy to a different region, 8 retained, restore tested.
  • Soft memory alert at 75%, hard emergency restart at 88%.
  • External monitoring with alerting to Discord or email.
  • AutoSaveSpan raised to 120–180 s on worlds larger than ~250 MB.
  • Dropped-item limits configured.
  • One documented restore drill performed — actually restore a snapshot to a scratch directory and boot it.

The drill is the step that separates a backup plan from a backup. A rotation directory full of tarballs that have never been restored is a hypothesis, not a safety net. Once the drill passes, restarts become boring — which is exactly what you want from an operational routine.

Leave a Reply