Commit Graph
3 Commits
Author SHA1 Message Date
jklein fc8f57e176 minecraft-server: null-safe hardening, probes and resources
The last open gap of #99. minecraft-server was the one chart of 21 where an
erased security context removed the hardening in silence: it never got the
defaultedDict helper, the StatefulSet rendered .Values.securityContext.pod
straight through toYaml, and the schema typed securityContext.pod as a strict
"object" without required. Measured against ci/minecraft-server/test-values.yaml
on 3.0.1:

  --set securityContext.pod=null --set securityContext.container=null
  -> helm exit 0, no schema error, "securityContext: null" for BOTH the pod and
     the container - no runAsNonRoot, no readOnlyRootFilesystem, no dropped
     capabilities, in a chart with a PVC.

Bitter place for it: the README lists this chart as the pilot for #68. It got
its schema early and was therefore never caught by the later null-safety wave.

Both halves are added, because they cover different cases. For a key WITH a
chart default Helm's coalescing deletes the null before validation ever sees
it, and only the defaultedDict fallback catches that; for a key WITHOUT one
(server.resources.limits) the null survives into validation, where the strict
"object" aborted the render. Now widened to ["object","null"].

The fallback table reproduces THIS chart's defaults, deviations included:
uid/gid pinned to 1000 (the image's minecraft user - starting as it directly
takes the entrypoint's non-root branch, which skips the recursive chown of
/data and the write to /etc/nsswitch.conf that a read-only root filesystem
would reject), fsGroupChangePolicy OnRootMismatch (a recursive chown of a 32Gi
world on every start costs minutes), the explicit empty capabilities.add list,
and all three probes as exec probes running the image's mc-health script - this
image has no HTTP endpoint, so a probe default rebuilt as httpGet would leave
the workload with no working probe at all. The startup budget stays 15s x 80.

server.resources goes through the same helper while keeping its memory request
deliberately unset, so it is still derived from the JVM heap.

Also, the same class one level down: init.sh chowned the migrated serverpack to
a hardcoded 1000:1000 while the uid it has to match was overridable. Both reads
now go through the same helper, so they cannot disagree. With the chart defaults
the rendered script is unchanged, so checksum/config does not move and no
existing pod rolls.

minecraft.resources.storage and .memory stay deliberately not null-safe: an
erased value yields an invalid PVC or an unparseable quantity, which is the loud
class #99 excludes.

Evidence, all against ci/minecraft-server/test-values.yaml unless noted:

  - the reported call now renders the defaults, byte-identical to the
    unmodified rendering. Fallback, not rejection.
  - rendering with securityContext, all three probes and server.resources set
    to null is byte-identical to rendering the chart defaults - which is what
    keeps the table and values.yaml in sync instead of leaving it to discipline.
  - no regression: rendering before and after the change is byte-identical, on
    the CI values, on the chart defaults, and with curseforgeServerpack enabled
    (the init-container path the CI values do not exercise). Confirmed
    structurally too, not only by raw diff.
  - partial nulling, 14 keys swept one at a time. On 3.0.1 twelve of them
    changed the rendering silently, including capabilities (the container kept
    every default capability instead of dropping ALL) and requests (cpu and
    ephemeral-storage disappeared entirely). On 4.0.0 all fourteen render
    identical to the defaults.
  - explicit false and 0 still win over the defaults; a scalar in a mapping slot
    is still rejected, now naming the values path.

Major bump: erasing one of these blocks used to change the rendered workload
and now does not, which is a runtime behaviour change. appVersion is untouched.

Refs #99
2026-08-24 16:53:25 +02:00
jklein 4874e5db1f Pin minecraft-server to dated image tags and fix the whitelist switch (2.0.0+up2026.8.2.java8)
Stage 3 of the Fleet migration. Two changes, both breaking, hence the major
bump from 1.0.0 to 2.0.0.

Dated image tags: appVersion moves from the rolling line tag "java8" to the
immutable upstream release tag "2026.8.2-java8". The rolling tags are rebuilt
continuously, so a pod restart could pull a different image than the previous
start without that being recorded anywhere. 2026.8.2 is the newest upstream
release (published 2026-08-23) and exists for all three Java lines.

Whitelist semantics: the chart now wires ENABLE_WHITELIST (and ENFORCE_WHITELIST)
in both positions, so minecraft.whitelist.enabled: false actually switches the
whitelist off. Previously the chart only emitted whitelist variables while
enabled was true and had no way to express "off" - and because server.properties
lives on the PVC, a release that had once written white-list=true kept it
forever. players stays guarded by enabled, since a non-empty WHITELIST enables
the whitelist in the image regardless of ENABLE_WHITELIST.

The default flips from enabled: true to enabled: false. Under the corrected
semantics the old default (enabled with an empty player list) would mean a
server nobody can join; false matches both the image and vanilla Minecraft.
2026-08-23 18:06:23 +02:00
jklein 227bcebeb6 Bring minecraft-server up to the chart style guidelines (1.0.0+upjava8)
Stage 2 of the Fleet migration for minecraft-server. Major bump: runtime
behaviour changes (probes, hardening) and the values interface breaks
(resource renames, strict schema).

Two real bugs first:
- The container declared port 35565 while the Service targeted 25565. It
  only worked because targetPort was numeric. Both are now server.port
  (25565) and the Service targets the named port.
- The download-serverpack init container was guarded by the map
  .Values.curseforgeServerpack instead of .enabled, so it rendered for
  every release and silently no-opped on an empty wget URL.

Conventions:
- Resources renamed to <release>-<chart>-<kind>, templates to
  <app>.<kind>.yaml. The PVC deliberately keeps its historical name -
  renaming it would provision a new, empty volume and orphan the worlds.
- Probes (liveness/readiness/startup) via the image's own mc-health, so
  Ready means the server answers a server-list ping. Startup budget 20 min.
- Hardened, values-configurable security contexts: non-root uid 1000,
  read-only root filesystem with an emptyDir at /tmp, all caps dropped.
- Resources from values with the limitFactor helper; an unset memory
  request is derived from the JVM heap so the two cannot drift apart.
- Additive env block rendered through tpl, pinned init image, storage
  class as a value, RCON wiring only when actually enabled.
- curseforge/ftb/curseforgeServerpack are now mutually exclusive.
- minecraft.version is pinned to the Java line instead of LATEST.
- The mc-router auto-discovery annotation is gone; the service type it
  implicitly switched is now the explicit value server.service.type.

Schema pilot (#68): values.schema.json with additionalProperties false on
every chart-owned level, rejecting all three known dead keys.
2026-08-23 16:40:44 +02:00