mirror of
https://github.com/shankar0123/certctl.git
synced 2026-06-07 16:01:30 +00:00
shankar0123
a91197014f
The shipped quickstart instructs operators to copy deploy/.env.example to
deploy/.env, edit POSTGRES_PASSWORD, and run docker compose up. On the
*first* boot of a fresh checkout this works. On the *second* boot — i.e.,
when an operator first booted with the default POSTGRES_PASSWORD=certctl,
then edited .env and re-ran up — the certctl-server container picks up the
new password (env interpolated at every container start) but postgres does
not. The postgres docker-entrypoint runs initdb only when the data dir is
empty; on subsequent boots the persistent named volume postgres_data is
non-empty so pg_authid retains the password baked in on first boot. The
server connects with the new credentials, postgres rejects them, and the
operator sees an opaque `pq: password authentication failed for user
"certctl"` in the server log with no pointer to the actual cause. New-
operator onboarding gets blocked on the documented production path.
Why a doc fix alone is not sufficient. Operators don't reread the docs
after a successful first boot — the trap fires on the *second* up, when
they think they've already learned the system. The opaque pq error is
indistinguishable in the log from a typo'd password or a misconfigured
secret store. The diagnostic has to fire at the moment the failure is
observed.
Why we don't try to fix the bootstrap. The env-vs-pg_authid divergence is
intrinsic to how the official postgres image bootstraps (see
docker-entrypoint.sh: initdb runs only if PGDATA is empty). Switching to a
bind mount or ephemeral volume breaks the production path; switching to
POSTGRES_PASSWORD_FILE + ALTER ROLE adds operator surface without
eliminating the divergence. The ergonomic fix is to surface the failure
mode loudly, with both remediation paths, at the exact log line where it
becomes visible.
Two remediation paths, surfaced together. Destructive: `docker compose
-f deploy/docker-compose.yml down -v && up -d --build` — wipes the
postgres volume so initdb re-runs with the new env value. Use this on
demos / first-time setup where data loss is acceptable. Non-destructive:
`docker compose exec postgres psql -U certctl -c "ALTER ROLE certctl
PASSWORD '<new>';"` followed by a server restart with the matching
POSTGRES_PASSWORD. Use this on any environment that holds data you want
to keep. Surfacing both means the operator can pick based on their
environment without us assuming.
Files changed:
- internal/repository/postgres/db.go — extract wrapPingError(err) helper.
errors.As against *pq.Error; on SQLSTATE 28P01 (invalid_password) emit
the multi-line guidance preserving the %w wrap chain. Non-28P01 errors
retain the original `failed to ping database: %w` shape so transient
connection-refused / timeout paths don't get noisy. Add
pgErrInvalidPassword = "28P01" constant. Convert blank
`_ "github.com/lib/pq"` import to direct import (driver registration
still works via init()) so we can name the *pq.Error type at compile
time. NewDB now calls wrapPingError(err) instead of inlining the wrap.
- internal/repository/postgres/db_test.go (new) — 4 internal-package
unit tests covering wrapPingError. AuthFailureGuidance pins the
contract substrings ("SQLSTATE 28P01", "POSTGRES_PASSWORD",
"first boot", "down -v", "ALTER ROLE"). NonAuthErrorPreservesOriginalWrap
pins the no-leak contract for SQLSTATE 08006 (connection_failure).
NonPqErrorPreservesOriginalWrap pins the network-level path.
NilReturnsNil pins defensive contract. All run in -short without
testcontainers — package postgres (internal) so the unexported helper
is callable directly.
- docs/quickstart.md — `> **Warning:**` callout immediately after the
`cp deploy/.env.example deploy/.env` block at lines 56-61. Names the
trap, names the SQLSTATE, gives both remediation paths. Uses the
in-file `> **Note:**` blockquote convention.
- deploy/ENVIRONMENTS.md — `**Stateful volume — first-boot password
binding (U-1)**` paragraph appended to the Postgres expert-note block.
Explains the env-vs-pg_authid divergence, points at wrapPingError as
the runtime diagnostic, lists both remediation paths. Uses the in-file
`**Expert note:**` convention.
Out of scope (separate follow-ups):
- deploy/helm/certctl/templates/postgres-statefulset.yaml has the same
root cause via PVC retention. The wrapPingError diagnostic covers the
Helm path because the same NewDB code runs at server startup; the
Helm-specific doc warning lands separately.
- /.env.example at repo root (line 16 hardcodes the password literally
inside CERTCTL_DATABASE_URL rather than interpolating) — adjacent
trap, separate fix.
- examples/{acme-nginx,private-ca-traefik,step-ca-haproxy,multi-issuer,
acme-wildcard-dns01}/docker-compose.yml all carry the pattern. The
diagnostic covers them; targeted doc warnings are scoped to the
canonical quickstart + ENVIRONMENTS docs.
Out of consideration:
- Switch to bind mount / ephemeral volume — breaks the production path.
- POSTGRES_PASSWORD_FILE + Docker secret + ALTER ROLE rotation — adds
operator surface without fixing the env-vs-pg_authid divergence.
Verification (all passing):
- go build ./...
- go vet ./...
- go test -short -race ./internal/repository/postgres/ — 4/4 new tests
pass plus existing tests
- go test -short ./... — every package green
- govulncheck ./... — no vulnerabilities in our code
- wrapPingError coverage 100%; postgres pkg total unchanged in shape
(NewDB/RunMigrations were 0% pre-fix, still 0% post-fix; new helper
adds 100%-covered statements)
Refs: coverage-gap-audit-2026-04-24-v5/unified-audit.md
§2 P1 cluster, cat-u-quickstart_postgres_password_volume_trap
GitHub Issue #10 (mikeakasully)