Collapses the eight open Renovate PRs for this stack into one change, and
answers what two of them turned out to carry.
The eight tags: qbittorrent 5.2.3 -> 5.2.4, bazarr 1.6.0 -> 1.6.2,
sonarr 4.0.19 -> 4.0.20, flaresolverr v3.5.0 -> v3.5.2,
recyclarr 8.7.1 -> 8.7.3, prowlarr 2.5.2 -> 2.6.5, radarr 6.3.0 -> 6.4.4,
seerr v3.4.1 -> v3.5.0. gluetun, the busybox init image and the
digest-pinned mediathekarr have no update pending and are untouched.
Major rather than patch, because prowlarr 2.6.3 and radarr 6.4.3 are not
plain bumps: they removed the hard-coded trusted proxy networks
(10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, fc00::/7, fe80::/10, read in
Startup.cs at the previously pinned tags) and replaced them with a
TrustedNetworks setting whose own default is string.Empty -
ForwardedHeadersConfigurator returns before registering a single
KnownNetwork while it is empty. Behind an in-cluster ingress controller
that means the forwarded headers are ignored: the client address the app
reports is the controller's pod IP, the scheme reads as http behind TLS
termination, and under "Disabled for Local Addresses" a proxied request
stops counting as local.
So a <app>.trustedNetworks value is added for prowlarr, radarr AND sonarr,
rendered as <APP>__SERVER__TRUSTEDNETWORKS and only when non-empty - an
empty variable cannot be told apart from an unset one by the app, but looks
handled to a reader of the manifest. The env form was verified per app
rather than inferred: all three bind ServerOptions to <App>:Server next to
a plain AddEnvironmentVariables(). A helper suppresses the chart's entry
when the app's own env block already carries the variable, for the reason
the UrlBase helper states - two identically named variables in one
container are not a defined state.
sonarr is included because it is already in the empty state: 4.0.19, the
version this chart shipped before, has no hard-coded networks either. It
arrived in what the convention classes as a plain app-version bump, so
nothing reviewed or documented it. Repairing two and knowingly leaving the
third would be the same silent pattern.
The default is empty, and this therefore does NOT restore the behaviour
from before the upgrade. That is deliberate and follows upstream's reason -
trusting X-Forwarded-For from any private network is a spoofing vector -
and it is made visible instead of reinstated: NOTES.txt warns in exactly
the combination that costs something, the app published through an Ingress
or an HTTPRoute and the value empty, and stays quiet otherwise.
AllowedHosts, added by the same upstream releases, is deliberately not
exposed. Its default is string.Empty as well and the app then accepts any
host, so there is nothing to repair; the trap is documented instead,
because if it is ever set the app appends only localhost, 127.0.0.1, [::1]
and its own hostname - neither the pod IP nor the Service DNS name - and
the kubelet probes the pod by IP, which an incomplete list answers with
HTTP 400.
At default values the rendered manifest changes in exactly eight lines, the
image tags; the new value renders nothing until it is set.
MediathekArr fakes a Usenet indexer towards Prowlarr and a SABnzbd downloader
towards Sonarr/Radarr and fetches from the ARD and ZDF media libraries. Both
consumers already share gluetun's network namespace, so they reach it over
127.0.0.1 - no service, no DNS, no ingress detour. Purely additive: the flag
defaults to false and a default render is byte-identical to 3.0.5 (verified
against the previous chart for both the default values and the CI values).
The four things that fail silently without them:
- A START GATE on gluetun. MediathekArr fetches its ruleset once at start and
never retries; started before the tunnel it answers every Newznab search with
0 results while the container is up, Ready and quiet. Kubernetes cannot order
two regular containers and an initContainer would deadlock, so the wait sits
in front of the image entrypoint: poll gluetun's health server on
127.0.0.1:9999 and exec the entrypoint afterwards; on timeout the container
exits rather than starting blind, because a crash loop is visible and an empty
indexer is not. The startup probe budget (300s) deliberately outlasts the gate
(180s).
- incomplete/ AND complete/ ON THE SHARED DOWNLOADS CLAIM, not in an emptyDir.
An emptyDir counts against the container's ephemeral-storage limit, and one
large download would have the kubelet evict the whole pod with all ten
containers - the trap ytdlp-web-player documents.
- THE COMPLETE PATH SONARR SEES. The claim is mounted at /downloads exactly as
sonarr/radarr mount it, and the chart sets DOWNLOAD_COMPLETE_PATH itself, so
the SABnzbd complete_dir it reports is the same absolute path on both sides.
- TWO PORTS: 5007 (SABnzbd role and UI, primary) and 5008 (Newznab role for
Prowlarr), both declared to gluetun's firewall and both on one Service. The
SABnzbd API is at /download/api - /api on 5007 answers with a misleading
Newznab validation error.
Further details:
- Image pinned by digest (tag + digest): "latest" is a 2024 build, "beta"
moves. Single-arch linux/amd64. The built-in helm-values manager cannot
follow that pair, so renovate.json5 gets the chart's first customManagers
block; verified by dry-run with two fixtures, since a matching and a broken
rule look identical while the pin is current.
- Hardened as measured: non-root, uid/gid from puid/pgid,
readOnlyRootFilesystem: true, all capabilities dropped, no PUID/PGID.
- Probes on measured 200 endpoints (there is no health endpoint): readiness and
startup on /download/, liveness as an exec covering BOTH ports, because one
container runs two processes and a wedged Newznab half is exactly what makes
searches come back empty.
- No ingress by default: the UI links its assets with root-absolute paths, so it
can only be served at the root of a host; on the shared domain that root
belongs to seerr. A sub-path fails the render instead of producing a broken
UI.
- values.yaml and NOTES.txt document the geo MARKER PROCEDURE (not a table of
series), the escape route of a German exit and its price for the whole stack,
the fail_message: null trap, and the HLS question as an open question.
- CI: mediathekarr enabled in ci/freesailarr/test-values.yaml, install timeout
raised to 1200s for the extra image and the gate.
Fixes#245
Adds a second, optional ingress shape and fixes the download target of
qBittorrent. The existing per-app-host shape stays the default and renders
unchanged.
Shared domain (#167)
Setting ingress.host moves every app without a host of its own into one
Ingress object with one path each (<app>.ingress.path, default /<app>), one
TLS secret and therefore one DNS name and one certificate for the whole
stack. A per-app host always wins and keeps its own object, so mixed
operation needs no extra values.
The apps differ in what they can do under a sub-path, and the chart follows
that rather than pretending they are alike:
- radarr, sonarr, prowlarr get <APP>__SERVER__URLBASE from the chart, but
only when their own env block does not already carry the variable.
- bazarr has no environment variable for its base path and stays one
documented manual step (Settings -> General -> Base URL).
- qbittorrent has no base path at all and renders a second object on the
same host with a prefix-stripping rewrite. It repeats the same TLS secret
name and deliberately omits the cert-manager annotation, so both objects
state their own TLS while only one certificate is requested.
- seerr has no base path either and cannot be rewritten, so it is pinned to
"/" and a fail refuses any other path. A second fail refuses two apps
claiming the same path.
qBittorrent save path (#166)
Without Session\DefaultSavePath qBittorrent writes to its built-in default
/config/Downloads - the 1Gi config volume. It fills within minutes, every
torrent goes to "errored" and qBittorrent can no longer write its own
metadata to BT_backup, so a restart in that state costs the torrent list.
The init container (renamed qbittorrent-config-interface ->
qbittorrent-config-seed, since it no longer only sets the interface) now
seeds qbittorrent.savePath, default /downloads. Unlike the interface binding
it is inserted ONLY when the key is absent and never replaced: a path changed
in the WebUI is a legitimate operator decision and must survive a restart.
This is the reason for the major bump: an installation that never set the
path writes somewhere else after the upgrade.
Documentation
NOTES.txt and the media block in values.yaml now spell out why a shared claim
with three subPaths does not produce hardlinks: the kernel compares the
vfsmount in do_linkat, not the superblock, so two bind mounts of one
filesystem are EXDEV - while stat reports the same device number on both
sides and suggests the opposite. The layout that would work (one claim, one
mount, categories as sub-directories) is named as a possibility; the chart
cannot express it today.
README.md gains the shared domain as the second admissible ingress shape of
the style guideline.
media.downloads, media.movies and media.tv rendered three fixed volume
entries, one per category. Two categories on the same existingClaim
therefore produced two pod volumes with an identical claimName, and such a
pod never starts: it stays in Init:0/1 without a single kubelet event
(measured on RKE2/Longhorn, RWO and RWX alike).
The volume list is now deduplicated by claim name. A new mediaVolumeName
helper names the entry after the first ENABLED category in the fixed order
downloads, movies, tv that carries the claim; the other categories mount
that same entry and differ only in their subPath. The restriction to
enabled categories matters because the three entries render under
different conditions - a canonical owner whose entry is not rendered would
be a dangling reference.
With three distinct claims the chart renders byte-identically to 2.0.1,
hence a patch bump.
Also corrected in the values documentation, after measuring instead of
assuming: the kubelet creates missing subPath directories itself, but they
inherit the claim root's mode, and fsGroup cannot be relied on to fix that
- Longhorn's fsGroupPolicy (ReadWriteOnceWithFSType) skips fsGroup on RWX
claims entirely. What has to hold is that the claim root is writable for
PUID/PGID. No chown or init container was added.
Retrofits the repo-wide groundwork that #59 deliberately left out to keep the
reviewed scope. Chart 2.0.0+up1.0.0 - a major, because seccompProfile changes
runtime behaviour; appVersion stays the freely chosen marker 1.0.0.
values.schema.json (#68). The largest schema in the repo: nine app blocks with
probes, resources, security contexts, env and image. additionalProperties is
false on every chart-owned level; the objects that go verbatim into the pod
spec stay open. It rejects a mistyped app name, a mistyped key at any depth and
ingress.domain - the key #59 removed, which until now would have been accepted
in silence.
Null safety (#99), both halves, because they cover different cases. The
defaultedDict helper catches a key WITH a chart default, where Helm's
coalescing deletes the null before the schema ever sees it; the widened
["object","null"] types catch a key WITHOUT one, where the null survives into
validation. Measured on the old chart: radarr.securityContext: null rendered
"securityContext: null" - the whole hardening gone - and
securityContext.capabilities: null dropped the capability block entirely, so
the container kept every default capability instead of dropping ALL. Both
render green.
The fallback table reproduces THIS chart's defaults including its deviations:
the five linuxserver.io containers keep runAsNonRoot unset and a writable root
filesystem so their s6 init can drop to PUID/PGID, flaresolverr keeps its
numerically pinned uid, gluetun keeps NET_ADMIN, and recyclarr keeps its exec
probes. Restoring the "clean" standard there would put the linuxserver
containers back on the image's uid 911 - the failure calibre-web-automated hit.
The sync between table and values.yaml is not left to discipline: rendering
with every block of the silent class explicitly set to null is byte-identical
to rendering the defaults, which is how a drift shows up.
seccompProfile RuntimeDefault (#84) in the pod context, so it covers the init
container too. Trivy's KSV-0104 drops from ten findings to none.
NOTES.txt (#43) covering what actually bites: the pod is not Ready until the
tunnel is up, the two fields the 1Password item needs, the warning when
firewallOutboundSubnets is unset, the immutable volumeClaimTemplates, and the
measured fact that imports do not hardlink across the media categories.
Also, from point 5 of the plan: checksum/config does not apply - the chart has
no ConfigMap - but checksum/secret-ref does, and was missing. Helm never sees
the secret's contents, yet it does see which op:// item the OnePasswordItem
points at, so repointing secrets.gluetunSecretRef now restarts the pod instead
of leaving the old credentials running.
README: the #43 status line was stale (it still claimed only homeassistant
ships NOTES.txt), and the mandatory-icon rule now records freesailarr as a
documented exception - #59 dropped the icon and did not amend the rule.
The chart bundles nine independently versioned upstreams, and eight of them
were already pinned in values.yaml while only gluetun took its tag from
appVersion. That split had no upside: it made gluetun the one image Renovate
could not reach through the built-in helm-values manager, and it implied a
"main app" the stack does not have.
gluetun now carries gluetun.image.repository/tag like every other image, and
the StatefulSet renders it from there. No tag comes from appVersion any more.
Renovate picks all nine up through helm-values - the manager the repo already
relies on for sidecar images - and ci/renovate-bump-chart.sh bumps the chart
version with each update.
Consequences, all documented where they bite:
- renovate.json5 deliberately gets no app-version block for this chart. One
would either duplicate the helm-values updates or rewrite an appVersion that
selects no image. The absence is commented so it does not read as an
oversight.
- freesailarr leaves the change-app-version.yml choice list. That workflow
rewrites appVersion and republishes, which for this chart would change no
deployed image at all.
- appVersion becomes a hand-maintained marker for the gluetun line. Since
renovate-bump-chart.sh bumps the patch level on an unchanged appVersion, a
gluetun update has to be paired with a hand edit of appVersion, or the +up
suffix stops describing what the chart deploys. Noted in Chart.yaml,
values.yaml and README.
FIREWALL_INPUT_PORTS does not open the app ports inside a pod network. gluetun
detects the pod's own subnet and inserts an INPUT accept for it ahead of the
per-port rules, and every packet arriving at the pod IP matches that rule by
destination. Measured with the packet counters: 50 fresh requests moved the
subnet rule from 306 to 356 packets while every --dport rule stayed at 0 since
container start. A port that listens but is not in the list (gluetun's control
server on 8000) is reachable from another pod all the same.
The variable is kept - it is the explicit statement of which ports are meant to
be reachable, and the subnet detection is gluetun's behaviour rather than a
guarantee of this chart - but the comments no longer claim the traffic would be
dropped without it.
Second correction, same run: the /tmp/gluetun emptyDir fixed the "open ...
permission denied" failure, but gluetun also chowns the file, which needs
CAP_CHOWN. With all capabilities but NET_ADMIN dropped that logs "chown
/tmp/gluetun/ip: operation not permitted" on every VPN restart. The file is
written and the public IP is persisted, so this is a log line rather than a
failure, and granting CHOWN to silence it would widen the container's
capabilities for nothing. Noted where the mount is declared.
Three review changes to the freesailarr chart.
Ingress: the chart no longer derives hosts as <app>.<ingress.domain> from one
Ingress object. The global ingress.domain is gone; every app carries its own
<app>.ingress.host and renders its own Ingress <release>-<chart>-<app>-ingress
with its own TLS secret tls-<release>-<chart>-<app>-ingress. The hosts are
therefore independent and may live in unrelated domains. The host is required,
but the required() call sits inside the loop over the *enabled* app ingresses,
so switching an app's ingress off must not make the chart demand a host for it.
clusterIssuer, className and annotations stay global defaults and are
overridable per app; an explicitly set per-app value wins (nil check instead of
"default", so an empty override is honoured).
Media: media.<category>.subPath mounts a sub-directory of the claim instead of
its root, default null (claim root = previous behaviour). The same
existingClaim may back all three categories with different subPaths. That is
the point rather than cosmetics: Radarr and Sonarr import by moving, and a move
within one filesystem can be a hardlink instead of a copy - across volume
boundaries every import duplicates the file.
Services: the collective Service with named ports is replaced by one ClusterIP
Service <release>-<chart>-<app>-service per enabled app that listens on a port,
flaresolverr included; recyclarr serves no port in daemon mode and gets none.
Each ingress points at its own app's service. All containers share gluetun's
network namespace, so the services differ only in their port - they provide
stable per-app DNS names, not network separation. The StatefulSet's governing
serviceName no longer names a removed object.
gluetun logs "clearing public IP data: open /tmp/gluetun/ip: permission
denied" every few seconds and never persists its public IP. It creates
/tmp/gluetun with mode 0644, i.e. without the execute bit needed to enter
the directory; it runs as root, but with all capabilities except
NET_ADMIN dropped there is no DAC_OVERRIDE left to paper over the missing
bit. An emptyDir provides the path with usable permissions.
qBittorrent had no network interface configured, so libtorrent listened
and announced on every interface of the shared pod network namespace -
including eth0 (the pod IP) and loopback. gluetun's kill switch drops
exactly those packets, so trackers answered with "skipping tracker
announce (unreachable)" or timed out, no peers were found and torrents
stalled at 0 bytes in stalledDL. Beyond the stall this is a leak: without
the binding qBittorrent also listens on the pod IP, i.e. reachable past
the tunnel.
The setting only exists in the persisted config file, which qBittorrent
creates itself on first start, so an init container seeds it before the
app comes up. The edit is idempotent and preserves existing user
settings: it replaces the keys if present, inserts them into the
[BitTorrent] section if missing, and creates file or section if needed.
The interface is configurable via qbittorrent.networkInterface (default
tun0); emptying it drops the init container.
With readOnlyRootFilesystem the s6-overlay init cannot rewrite
/etc/passwd, so the containers silently kept the image's built-in
uid/gid 911 instead of the configured PUID/PGID (verified at runtime) -
breaking ownership on the shared media volumes. Drop the now-unneeded
/run and /tmp emptyDirs; deviation documented in values.yaml.
Replace the upstream-recommended postStart hook, which runs
concurrently with the entrypoint and loses the race (observed: it
deleted the ip rule the restarted gluetun had just created, leaving the
container in a permanent CrashLoop behind the kill switch). A command
wrapper now removes stale table-51820 rules and the leftover tun
device deterministically before exec'ing /gluetun-entrypoint, covering
container restarts after unclean exits.
An uncleanly exited gluetun leaves its ip rule in the shared pod
network namespace; the restarted container then fails forever with
'adding ip rule 101: file exists' while the kill switch keeps blocking
all egress. Apply the upstream-recommended Kubernetes fix: a postStart
hook dropping table-51820 rules.
The image copies and patches the chromedriver binary to the hardcoded
path /app/chromedriver at startup (no env override exists) and writes
chromedriver caches under /app, so readOnlyRootFilesystem cannot hold;
deviation documented in values.yaml. Reverts the HOME=/tmp workaround,
which only covered the cache path, not the binary patching.
The image's home is /app on the read-only root filesystem, but
flaresolverr writes chromedriver caches under $HOME (/app/.local);
pointing HOME at the /tmp emptyDir keeps the root filesystem read-only.
jklein