UpdateUser wrote a new address but kept email_verified, so patching a verified
account asserted a proof of an address nobody had proven — and the
pre-session login mails and resolves on exactly that flag, so a typo'd edit
could hand the account's sign-in codes to the wrong mailbox.
Changing the address now clears the flag in the same write; a no-op patch that
passes the same value keeps it. The fake mirrors the semantics, and the pgint
suite pins both halves (same value keeps proof, new value drops it).
The design has claimed since migration 0010 that at most one account can hold
a PROVEN email address, with ErrEmailTaken as the 409 a second verifier sees.
Neither half ever shipped: no migration created users_verified_email_unique,
and VerifyEmailOTP had no guard at all — the sentinel was defined but never
returned, so two accounts could both verify one address. The damage is not
cosmetic: the pre-session login resolves accounts BY verified email, so the
duplicate decided which identity a mailed sign-in code belonged to.
- Migration 0020 creates the partial unique index (lower(email) WHERE
email_verified) the comments have been citing — the database-level backstop.
- VerifyEmailOTP now refuses the take-over with ErrEmailTaken BEFORE consuming
the code (the address, not the code, is the problem), charges no attempt,
and maps a lost cross-user race (unique violation) to the same answer.
- The verify handler answers 409 email_taken instead of a generic 500.
Covered by the pgint suite (sequential double-verify refused with the code
still live, a direct duplicate write still loses to the index, the refused
account can still prove its own address) and a hermetic 409 case.
The hermetic suites encode the store contracts against fakes; PGRepo drifted
behind them three times (attempt accounting, a missing JOIN, a missing FOR
UPDATE) while every unit test stayed green. This harness replays the real
embedded migrations onto a throwaway database — its name must contain "pgint"
or the harness refuses to run — and exercises the SQL directly: sessions, the
onboarding email-OTP lifecycle (supersede/expiry/lockout), the pre-session
login consume, the op-login state machine, link and bind-code redemption,
submissions, and builds with the image admission round trip.
Run it after touching SQL under internal/api/pgrepo.go, internal/submit, or
internal/build; CONTRIBUTING.md carries the one-liner.
- The control-plane probes (0c8e29b) and the user-modpack build lane
(f79e5eb + 02fd2de) get their live evidence recorded, including the three
drill-only defects the lane fixed (Job scheduling, Kaniko Dockerfile
ownership, Trivy DB egress).
- The stale 'unfixed' rows #7/#11-#15 are reconciled with their commits.
- Remaining-work list re-stated: image durability, PG contract tests, the
panel's /jobs block, multi-node reaper placement, alerting, and the newly
found player-visible build status gap.
The first live build (Kaniko v1.24, 4 vCPU / 5.5 GiB node) walked the new
transport end to end and hit three real defects, each invisible to unit tests:
- The Job requested its FULL limits (2 CPU / 4Gi per container), so the build
Pod never scheduled on the platform's own starter node: FailedScheduling /
Insufficient memory, Pending forever. Requests are now a small floor
(250m / 512Mi, never above a configured cap) while the limits stay the
safety caps.
- Kaniko re-copies the Dockerfile out of the context and chowns/chmods it to
the source owner; a 65532-owned context (the distroless felis image uid)
fails that under the pod's dropped capabilities ('copying dockerfile:
chown /kaniko/Dockerfile: operation not permitted'). The fetch container
now extracts as root — the uid Kaniko already runs as — so the copy
succeeds; the pod was root by necessity regardless.
- Trivy's DB fetch is exactly what the build egress lock denies: the scan
step failed closed on mirror.gcr.io. New [registry] trivy_db_repository
renders --db-repository, and docs/troubleshooting.md §8e now carries the
verified mirror recipe (docker pull/tag/push of aquasec/trivy-db:2 into the
internal registry; --insecure already covers its plain HTTP).
Verified live after this batch: fetch initContainer streamed the blob through
the API + netpol + token, Kaniko built and pushed registry.felis.svc:5000/
user-uploads/sub-<id>:latest, and Trivy scanned against the mirrored DB.
A submitted modpack was durable but unreadable: the uploads PVC cannot cross
namespaces (felis-api mounts it; Kaniko runs in felis-build) and the s3 lane
handed the sandboxed build Pod no credentials, so NO user build could ever
consume its context. The transport is now the API itself:
- submit: derived context refs become the internal-face URL
/api/v1/internal/submissions/{id}/context (service-token gated), and Blobs
gains Open (local + s3) with an ErrBlobNotFound sentinel for the route's 404.
- api: serves that route on the internal face only (openapi.yaml updated; the
route-coverage test enforces it).
- build: an http(s) context renders a context-fetch initContainer (the felis
image's new fetch-context entrypoint) that streams the blob with the
namespace-local service-token Secret — never mounted into Kaniko — and
extracts it under a zip-slip guard into a size-limited emptyDir that Kaniko
reads read-only as --context=/context.
- platform/install: the api Deployment carries its own internal base URL; the
build namespace gets the token Secret through the existing replica mechanism
(bootstrap.sh + felis setup); the build egress lock opens exactly the control
namespace on the internal port.
- cmd/felis: fetch-context entrypoint (registered, documented, unit-tested for
escapes/symlinks/non-gzip).
Tests cover rendering, hardening, the s3/local Open paths, and the route's
404/503 mapping. Verified next on the real single-node cluster with Kaniko.
The api, operator and registry Deployments shipped with no liveness/readiness
probes at all: a wedged process stayed 'Running' forever, and the operator had
no health listener to probe in the first place. Kaniko build evidence on a
fresh install showed the only cluster-wide red after a disk-pressure pass was
Deployment status that never reflected health.
- felis-api: readiness /readyz (DB + K8s API round-trip) and liveness /healthz
on the internal face (:8081), the only listener that serves both endpoints;
liveness deliberately avoids /readyz so a DB blip cannot restart the api.
- felis-operator: new --health-probe-bind-address (:8081) with controller-
runtime's /healthz + /readyz (registered ping checks; an unregistered handler
map would 404), plus the matching container port and probes.
- registry: /v2/ probes on the pinned port, so a broken storage backend stops
reading as 'Running'.
Tests pin paths, ports, and that each probe targets a declared container port.
- [registry] gains kaniko_image / trivy_image / build_cpu_limit /
build_mem_limit overrides; empty keeps the compiled-in defaults. An
air-gapped or mirrored install has no route to gcr.io/aquasec (the
build egress policy allows only DNS + registry + package mirrors), so
builds previously could not even start their executors.
- deferred-seams: the uploads-context entry now records WHY a mount is
impossible (PVCs cannot cross namespaces) and that the s3 lane also
lacks credentials in the build Pod — options captured for the real fix.
- troubleshooting 8e (executor ImagePullBackOff + the overrides),
13b rewritten (verified eviction refusal, 5m pressure-transition,
image-GC recovery), 15 (upgrade/rollback runbook for Recreate).
- Backup semantics decided and documented: a backup is the whole /data
volume (worlds + config + plugins + cache) and a restore rolls all of
it back — OpenAPI/README wording updated to match (same-tag images are
still watched for regressions by the openapi parity gate).
Following the first shield attempt (custom class, value 1e6) a live drill
showed the limit: kubelet evicted the game pods and then the api,
operator and registry anyway — evicting them was never what reclaimed
the disk — and with the images containerd-only, the GC stage left
everything in ImagePullBackOff. A custom class cannot be raised past 1e9
(the API caps user-defined values), while kubelet's eviction refusal
needs >= 2e9, so the control plane now uses the built-in
system-cluster-critical.
Re-drilled: disk filled to 1.7G free -> login/lobby evicted, and kubelet
logged "cannot evict a critical pod" for felis-api/operator/registry,
which stayed Running throughout. Recovery facts now in troubleshooting
13b: the DiskPressure condition lingers ~5m after space is freed
(--eviction-pressure-transition-period), and game images GC'd while
their pods were evicted need the documented re-import (verified: 25s to
Running).
A full disk made kubelet's node-pressure eviction pick control-plane pods
alongside game pods (both priority 0), and with the images existing only
in the node's containerd (air-gapped), losing the api meant a manual
image re-import. Every control-plane pod template (api/operator/reaper/
registry) now names the bundle's cluster-scoped felis-control-plane
PriorityClass: value 1,000,000, preemptionPolicy Never — eviction order
only, never preempting a running game server. The image-GC half is not
code-fixable on an air-gapped box; troubleshooting gains 13b with the
recovery path (re-run the installer to rebuild imports, or docker save |
k3s ctr images import - for one image).
Live drill found this: the reaper pod runs as uid 1000, k3s creates its
storage root /var/lib/rancher/k3s/storage 0700 root:root, so enabling
retention on a stock install made every archive fail
'lstat /worlds/<pvc>: permission denied' and skip the world (fail-closed,
but a silent no-op). bootstrap now grants traverse (setfacl u:1000:x,
else chmod o+x) when FELIS_WORLDS_HOST_PATH is set, the renderer's
precondition note names the requirement, and troubleshooting documents
both it and the multi-node nodeSelector fact.
Verified on the VM after granting the ACL: a 20d-idle world with a marker
file was archived into felis-backups (marker intact), its PVC and host
directory were reclaimed, world_backups got an inactive_15d row, and the
servers row/CR were retained.
Three faces of one gap, all on the supported install path:
- Backup/restore answered 503 out of the box: nothing ever rendered the
archive PVC, so FELIS_BACKUP_PVC was unset. The bundle now renders the
PVC (Minecraft namespace, RWO 10Gi, cluster default class) and
'felis manifests' names it by default (--backup-pvc= is the explicit
no-store shape); bootstrap passes it through so the generated felis.toml
[archive] local_path and the jobs' mount path come from one variable.
- Retention was unreachable: bootstrap never passed the reaper flags. It
now forwards FELIS_WORLDS_HOST_PATH/FELIS_ARCHIVE_LOCAL_PATH, so one
env enables the daily CronJob; unset keeps today's fail-safe (no reaper,
nothing deleted).
- Even when enabled it could not find a world on a stock install:
resolveWorldDir now also resolves the exact local-path directory
<pv-name>_<ns>_<pvc-name> read from the live PVC's volumeName (never a
glob, so a stale deleted PV's bytes can't be archived in place of the
current world). Reaper Role gains persistentvolumeclaims:get (weaker
than the delete it already held).
README (zh/en) stops promising automatic/scheduled backups and states
retention is opt-in. bootstrap_test covers the env->flag contract.
Backup and restore only enqueue a cluster Job; a later failure left its
only trace in that Job object, invisible without kubectl. Add
GET /api/v1/servers/{name}/jobs (owner-or-admin) projecting the newest
20 managed Jobs (felis-backup / felis-restore) as
running|succeeded|failed with message and timestamps. Nil reader -> 503
jobs_unavailable, mirroring the backup/restore feature gates. RBAC gains
jobs:list; OpenAPI parity updated.
The readiness gate is status-driven; at a 5s re-probe cadence the observed
'container Ready but API still 409 not_running' window was 6~10s. Halving the
cadence halves the worst case; probes still only run while unreachable.
Resolving a session cookie failed identically whether the credential was
missing or Postgres was unreachable: local_auth_enabled read errors fell into
the fail-closed 'disabled' branch and SessionUser errors into 'invalid
session', both surfacing as 401 'authentication required' — a lie that reads
as 'log in again' during an outage. Split the enabled-read into
(enabled, error), tag non-ErrNotFound store failures with errAuthBackend, and
map that to a new 503 auth_unavailable in requireExternal. Fail-closed is
unchanged: missing setting / bad value / missing session stay 401.
The setup URL carries a 43-char token and overruns 80 columns; the TUI
renderer clipped it. Break it at the query '=' boundary (token on its own
line) with a shared wrapDisplayURL helper used by both the Owner wizard and
the mc-bind wizard; unit test pins the no-loss concatenation.
Without SetLogger, the first reconcile prints
'[controller-runtime] log.SetLogger(...) was never called; logs will not be
displayed' followed by a full stack trace (live-observed in felis-operator).
Route it through logr.FromSlogHandler(slog.Default()) so its messages are
ordinary stderr lines; go-logr/logr promoted to a direct dependency.
Follow-up to the CronJob placement fix: a Pod cannot USE a ServiceAccount from
another namespace either (live drill: 'error looking up service account
minecraft/felis-reaper: serviceaccount not found'). Move the SA and its
RoleBinding subject to the Minecraft namespace alongside the CronJob.
A Pod can only mount PVCs from its own namespace; the CronJob referenced the
minecraft-namespace backup PVC while being rendered under ControlNamespace, so
it could never schedule — live drill: FailedScheduling 'persistentvolumeclaim
felis-backups not found'. The reaper Role/RoleBinding were already
minecraft-scoped (the objects it touches live there), so the CronJob was the
odd one out. The minecraft felis-config replica (felis setup, backup Job fix)
supplies its config mount.
6-way concurrent redeem of one code 500'd on users_username_key (each request
generated a fresh user id but the same uuid-derived username), plus the rarer
two-codes-one-uuid race. Same drift family as VerifyLinkCode, which already
locks its code row and handles the conflict.
- SELECT ... FOR UPDATE the code row: same-code racers serialise; losers exit
as ErrLinkCodeInvalid (400 invalid_code), no user row is attempted.
- INSERT users ... ON CONFLICT (username) DO NOTHING + re-read by username:
cross-code racers converge on the winner's row (role checked, staff still
refused) instead of a unique-violation 500.
- account_links ON CONFLICT (mc_uuid) DO NOTHING for the same race.
Verified live: same-code x6 = 1x200 + 5x400; two-codes x2 = 2x200 same user;
db clean; zero unmapped errors.
The interface doc promised 'each joined to its staff username', the fake and
the pending handler both project username and created_at, but the PG query
selected neither — live internal /op-login/pending returned username:"" and
created_at:0001-01-01. Same drift class as ConsumeLoginEmailOTP: fake-based
tests can't see PG-only regressions.
The PG implementation was a single UPDATE ... WHERE code_hash that returned
ErrNotFound on zero rows: every wrong, expired, replayed or superseded code on
the pre-session email-login door (and the op-login finish / migration confirm
doors) fell through to writeError's unmapped-error 500, and attempts were never
charged so otpMaxAttempts/ErrOTPLocked could not trigger. The fake repo and the
Repo interface ("SAME code lifecycle as VerifyEmailOTP") already documented the
intended contract; only the PG side had drifted.
Mirror VerifyEmailOTP's transaction without its users write: SELECT ... FOR
UPDATE the newest live row, expiry + attempt cap before the hash compare,
mismatch charges one attempt and returns ErrOTPInvalid without consuming,
match consumes and commits. Verified live on the VM: 5 wrong guesses return
400 and stop at attempts=5 (correct code then also refused, unconsumed);
fresh code redeems; replay returns 400.
The gofmt/staticcheck commit staged handlers_user.go (singular) for the
formatting fix but missed this sibling for its two struct-literal-to-
conversion cleanups.
Live verification of the previous commit showed the immediate retry STILL
stranded: deleting a finished Job leaves it terminating (job-tracking
finalizer), so the re-Create collided with the dying object and was
mapped to ErrAlreadyExists a second time. Poll until the name actually
frees (bounded, ~10s) and surface a 'retry shortly' error if a stuck
finalizer ever outlives the budget. Fake-client tests pin both the
replace-finished and coalesce-in-flight branches.
An E2E audit on a live install found that a FAILED restore held its
deterministic Job name for the rest of the 10-minute TTL, so the next
restore answered 202 'restoring' while nothing ran (ErrAlreadyExists was
treated as success unconditionally). K8sJobs now inspects the colliding
Job: in-flight still coalesces, finished (succeeded or failed) is
deleted and replaced. The minecraft-namespace Role gains jobs:get/delete
for exactly that replacement.
The same audit found the backup Job mounts the felis-config Secret but
the installer only provisions it in the control namespace, so every
backup Job stranded on FailedMount. felis setup now replicates it into
the minecraft namespace beside the service-token and forwarding
secrets.
govulncheck flagged pgx v5.7.1 (GO-2026-5004, SQL-injection class) as
reachable from pgrepo.go, plus the old x/net and x/text. Bump all three
to the fixed versions; go vet/test stay green.
Nine files had drifted from gofmt and nothing checked; nine staticcheck
findings were live (three dead symbols, capitalization, a redundant
Sprintf, two literal-to-conversion sites, a nil test context). Fix all
of them and make CI fail on unformatted Go so this cannot re-drift.
Previously /readyz only verified Repo != nil && Cluster != nil — a
process-liveness check, not a dependency-health check. The spec
requires the readyz probe to verify DB, K8s API, and CRD informer
are live before declaring the pod ready.
- Repo interface gains Ping(context.Context) error
- Cluster interface gains Ping(context.Context) error
- PGRepo.Ping delegates to sql.DB.PingContext
- K8sCluster.Ping lists MinecraftServer CRDs (Limit=1) in the
configured namespace, exercising both the API and CRD informer
- handleReadyz iterates ping checks; any failure returns 503 with
the failing dependency name in the error message
- fakeRepo and fakeCluster gain configurable pingErr for hermetic
test coverage of the failure paths
New test: TestReadyzPingsDependencies verifies 200 when healthy,
503 when DB or K8s API is down.
Previously a server whose pod was ready but RCON probe kept failing
would stay in Starting phase forever. The CRD defines TimeoutSeconds
and ReadinessTimeoutSeconds but the reconciler never checked them.
- PodNotReady path: if the pod stays not-ready past timeoutSeconds
(default 300s), transition to Failed
- RCON unreachable path: if RCON stays unreachable past
readinessTimeoutSeconds (default 300s), transition to Failed
- Helper methods startupTimedOut/readinessTimedOut compare
StartRequestedAt against the respective timeout, falling back to
300s defaults when unset
- 2 new tests: ReadinessTimeoutConvertsToFailed, StartupTimeoutConvertsToFailed
Fixes the scenario where a broken backend (bad jar, crash-looping
process) would permanently occupy a Starting server slot.
Add per-user resource quota enforcement across all four dimensions:
max_servers, max_cpu_milli, max_memory_mb, and max_storage_gb.
- Migration 0013: add cached_cpu_milli, cached_memory_mb, cached_storage_mb
columns to servers table for pure-SQL per-owner aggregation
- SeedServer now writes resource cache alongside server row
- QuotaCheck replaces QuotaAvailable at claim time, checking all four caps
against the owning user's cumulative usage
- handlePatchServer checks owner's quota before allowing memory/resource
changes on owned servers; unowned servers skip the gate
- handleInternalClaim mirrors the full quota check
- UpdateServerResources keeps the cache in sync after spec mutations
- Reaper zeros resource cache on ReleaseWorld so released resources
are not counted against a former owner
- quantityToMilli/quantityToMB helpers convert K8s quantities to
quota-comparable integers
19 test packages pass.
Adds empty-server auto-stop to the reconciler. When Idle.AutoStopEnabled
is true and the RCON player tally is zero for EmptySecondsBeforeStop seconds,
the operator flips desiredState to Stopped, which triggers the normal
graceful-shutdown path.
- Add EmptySince status field to track empty duration
- Clear EmptySince on stop and when players return
- Reuse existing RCON probe's player count (zero extra network cost)
- 4 new test cases covering timestamp, timeout, player-join reset, and
disabled-by-default
Lemon-miaow