The hectic logger opens fd 3 as a dup of stderr. Child processes inherit
this fd, and daemonized PostgreSQL/PostgREST keeping it open prevents the
terminal from returning to the prompt after the spawning script exits.
- Add with_closed_fds helper that runs commands in a subshell with fds
3-9 redirected to /dev/null
- Inline the helper into both database and postgres-init builds
- Wrap pg_ctl start and postgrest with the helper
Move postgres-init-hectic-inheritance test (13 cases) to
migrator/init-hectic-bundle since the bundle is now applied by `migrator init`
instead of `postgres-init`. Drop init-migrator-with-inherits since
`--inherits` is now a deprecation warning, not an error. Add db-tool
hydrate-hook test (5 cases) covering --no-hook skip, default apply,
idempotency, and HECTIC_DOTENV_FILE. Augment init-migrator with
hectic.version and hectic.secret table assertions.
Add a hectic.immutable parent table. Tables inheriting it get auto-attached
BEFORE INSERT/UPDATE/DELETE/TRUNCATE row+statement triggers that block DML
unless the session sets hectic.migration_mode='on' (intended use: SET LOCAL
inside a migration transaction). Same exemptions as the rest of the bundle
apply (hectic schema, partitions, temp tables, GUC-excluded schemas).
database diff now appends an --- IMMUTABLE TABLE DATA --- section to its
output, with per-table unified row diffs of every table inheriting
hectic.immutable, surfacing drift in 'frozen' reference data alongside schema
drift. Subcommand exits non-zero when either schema or data differs.
Test postgres-init-hectic-inheritance extended to 10 cases covering
immutable triggers, DML blocked outside migration_mode, SET LOCAL allowing
DML inside a transaction, GUC not leaking past COMMIT, and TRUNCATE under
migration_mode.
Adds a SQL bundle plus event triggers that enforce `INHERITS (hectic.created_at)`
on every user CREATE TABLE and auto-attach a BEFORE UPDATE row trigger when a
table inherits `hectic.updated_at`. Always-exempt: `hectic`, `information_schema`,
`pg_*`, declarative partitions, temp tables. Per-DB opt-out via the GUC
`hectic.inheritance_extra_excluded_schemas`.
Exposed three ways:
* `pkgs.hectic.hectic-inheritance` — derivation with the SQL at
$out/share/hectic/hectic-inheritance.sql
* `self.lib.hecticInheritance.{sql,path}` — pkgs-free Nix surface
* `postgres-init` opt-in via `PG_HECTIC_INHERITANCE=1` (HECTIC_INHERITANCE_SQL
overrides the default)
Test postgres-init-hectic-inheritance covers all six branches: bootstrap,
non-inheriting reject, accepting inheritance, auto updated_at trigger fires,
GUC exclusion, declarative partition exemption.
Previously when PG_REUSE=1 and PG_VERSION existed but the target database had
never been successfully created (e.g. devshell exited mid-init in a prior run),
postgres-init skipped createdb and the subsequent psql connection failed with
'database "<db>" does not exist'.
Now on reuse path we probe pg_database and create the target DB if missing,
making postgres-init fully idempotent across stale-state recovery.
Adds postgres-init-reuse-missing-db test.
Wrap db-tool, postgres-init, postgres-cleanup with lib.makeOverridable so
consumers can inject extension-enabled PostgreSQL via .override { postgresql = ...; }.
Add PG_CONF_FILE: when set, replaces script-generated postgresql.conf entirely
(runtime port and unix_socket_directories still appended/overridden).
Previous basename guard only ran the function when invoked as the
postgres-init binary, but consumer shellHooks source the script to
inherit POSTGRESQL_HOST/PGURL/etc. Sourcing left the function defined
but never called, leaving WIPE_PGURL/PGURL unset and breaking devshell
entry in proxydoe and ruststats.
yukkop
