Every operator eventually meets the worst save in their life: a world file that is present, the right size, and corrupt. Disk failures are the easy case, because RAID covers them and you get a clean error. The hard case is a save written mid-crash, a plugin that flattened a region, or an admin who deleted the wrong directory at 2 a.m. and did not notice for four days. For that case the only thing that matters is how old your last good copy is and how long it takes to put it back into service — RPO and RTO, not the number of backups you have.
Defining RPO and RTO for a game world
Recovery Point Objective is how much progress you are willing to lose. For a 200-player survival server that saves every 5 minutes, an honest RPO of 5 minutes means losing at most one save cycle of building, farming, and inventory changes. Recovery Time Objective is how long the server is down from the moment you decide to restore. Neither number is a backup setting; both are promises you make to your players.
| Community size | Sensible RPO | Sensible RTO | What it costs |
|---|---|---|---|
| 20–50 players, casual | 1 hour | 4 hours | Nightly full to object storage; no hot spares |
| 200 players, active economy | 10–15 min | 60–90 min | Hourly incremental + daily full, tested restore |
| 500+, competitive or paid | 5 min | 30 min | Continuous WAL/log shipping, warm standby, rehearsed failover |
Notice that the cost of RPO is driven by how often you can afford to disturb the world, and the cost of RTO is driven by how fast your storage can read a few gigabytes back and how quickly you can bring the instance up. Both are much cheaper on dedicated metal than on a shared host, because you control the save cadence and you are not competing for IOPS during the restore.
Hot copy versus cold copy: the consistency problem
A file-level copy of a world that is actively being written is not a backup. Minecraft region files, Ark’s .ark save, and most survival engines do not write atomically: a 20 MB region file can be half-old and half-new while your copy runs. The result is a backup that restores into a world with chunk-level corruption, which is worse than no backup because you only find out under pressure.
The safe options, in order of preference:
- Quiesce, save, copy, resume. Send the save command over RCON, wait for the write to flush, copy, then release the world. Costs 5–30 seconds of frozen players at your chosen cadence.
- Snapshot at the filesystem layer. An LVM thin snapshot or a ZFS snapshot taken in the same instant as a save gives you a point-in-time view cheaply, and you copy from the snapshot rather than the live volume. Requires a filesystem that supports it and a small amount of planning when you build the box.
- Application-level dump. If the world lives in a database (many survival servers keep inventories in MySQL), dump with a consistent snapshot flag rather than a raw table copy.
- Never: rsync or cp -r against a live world directory and call it done.
The three-layer storage plan
The layers exist because each covers a different failure mode. RAID is not a backup; it is uptime insurance and it does nothing for corruption. Local snapshots cover fast rollback from a bad plugin or a bad save, and they are the first thing you reach for. Off-site object storage covers the loss of the machine, the datacenter, or a compromised admin account. Each layer has a different recovery story and a different retention period.
| Layer | Retention | Storage cost for 40 GB world | Covers |
|---|---|---|---|
| LVM/ZFS snapshot, local | Last 12–24 snapshots | Incremental, a few GB on the RAID array | Bad plugin, bad save, quick rollback |
| Daily full to off-site bucket | 14–30 days | ~$0.60–1.20/mo at standard object pricing | Machine loss, datacenter loss |
| Monthly archive, cold tier | 12 months | ~$0.10–0.30/mo | “Restore the world from six months ago” |
| Config and credentials to a separate location | Permanent | Negligible | Rebuilding from scratch |
These figures are illustrative rather than a quote; check your provider’s current object-storage rate and egress policy, because egress on a restore is the line item that surprises people. A 40 GB world restored from a cold tier is a few dollars of egress at most, which is a trivial price for a one-hour RTO.
Cadence for a 200-player server
A nightly full backup is the wrong primitive for a 200-player server. The world changes constantly, players lose an entire evening’s progress if the RPO is a day, and the nightly window lands in your evening peak in half the timezones you serve. Use a tiered schedule instead: full backup nightly during your quietest hour, incremental snapshots every 15–60 minutes, and a save-then-copy every 5–10 minutes if your engine’s save is fast enough to not disturb play.
Testing the restore, not just the backup
A backup you have never restored is a hypothesis. The only number that matters is the last time you proved the restore works and how long it took. Put a quarterly restore drill on the calendar: pull the most recent off-site copy onto a spare instance, start the server, load in, walk a player through the spawn point, and time it from “decide to restore” to “players can connect”.
In practice the restore is rarely the slow part. The clock is eaten by discovering that the credentials to the bucket expired, that the backup script silently started writing to a full disk three weeks ago, or that the world needs a version of the plugin that no longer exists in the package index. Every one of those is a finding from a drill and a disaster in a live incident. Restore also needs a place to land — a second instance or a container on the same dedicated box is enough, and it is why keeping a bit of spare capacity on the machine pays for itself.
For the underlying storage decisions — how many drives, which RAID level, how much NVMe you actually need for write bursts — the dedicated server configurations for game worlds are worth reviewing before you size the box, because the snapshot layer and the RAID layout are far easier to get right at build time than to retrofit later.
The minimum viable plan
If you do nothing else, do these five things. Snapshot the world at the filesystem layer in the same instant as a save, every 15 minutes, keeping 24 of them. Push a nightly full copy to object storage in a different provider than your host. Write the RCON save command and the restore command into a one-page runbook that a second admin can follow. Rotate and store the bucket credentials somewhere other than the game server. And put one restore drill on the calendar per quarter, with a stopwatch. That combination gets a 200-player server to a 15-minute RPO and a 90-minute RTO for a storage bill under a couple of dollars a month — the cheapest insurance in the whole stack.

Leave a Reply
You must be logged in to post a comment.