Joseph Doherty
9675823cd6
Two disposable docker fixtures (postgres:17-alpine on 15432, gvenzl/oracle-free: 23-slim on 11521) shaped like the existing mssql one — project label, distinct container name, healthcheck, one-shot seed service — plus a live suite per provider, gated on SQL_PG_* / SQL_ORACLE_* and skipping clean offline. Each suite covers what a DIALECT can get wrong, rather than re-running the provider-agnostic read semantics already pinned against SQL Server and SQLite: liveness + catalog load, a KeyValue read, the provider's own row-limit syntax via topByTimestamp (LIMIT 1 / FETCH FIRST 1 ROWS ONLY, where T-SQL emits TOP 1), identifier case-folding through the design §8.1 catalog gate, and the AdminUI browse walk driven through SqlDriverBrowser. Two seed decisions are load-bearing rather than scaffolding: - The Postgres schema is owned by a role the suite is NOT; the connecting login holds USAGE + SELECT only. That is the exact provisioning PostgresDialect's ListSchemasSql exists for (information_schema.schemata lists only schemas the role OWNS, so a granted-not-owning poller would see none and every tag would be rejected). A fixture that owned its own schema would pass either way. - Every identifier in both seeds is created UNQUOTED, so the catalogs hold lower-case (Postgres) and UPPER-case (Oracle) spellings. The case-folding legs author the opposite case and pin SqlCatalog.TryMatch's REAL contract — exact ordinal first, then a UNIQUE case-insensitive match whose catalog spelling is substituted — which on both providers is what makes mixed-case authoring work at all, since both dialects emit case-sensitive double-quoted identifiers. Connection strings are composed through dialect.Factory.CreateConnectionString- Builder() and the abstract DbConnectionStringBuilder indexer, so the csproj keeps its no-direct-provider-reference discipline; every key used was verified accepted through that indexer. The csproj gains only Driver.Sql.Browser, for the browse leg. Live-run evidence (local OrbStack, both images arm64): Postgres 5/5 green; Oracle 6/6 green after the ORA-01745 dialect fix this suite found. Offline the whole project is 33 skipped / 0 failed.
88 lines
5.2 KiB
SQL
88 lines
5.2 KiB
SQL
-- Seed for the Sql driver's PostgreSQL live gate (PostgresReadTests), applied once by the
|
|
-- compose `postgres-seed` one-shot as the superuser against the OWNED `otopcua-sql-pg` container.
|
|
--
|
|
-- The table shapes mirror the SQL Server fixture (SqlPollServerFixture) row for row — a present
|
|
-- value, a present row with a NULL cell, and a wide-row table whose NEWEST row is deliberately not
|
|
-- its first — so the same assertions can be made against both backends and a divergence between
|
|
-- the two provider paths shows up as a failing assertion rather than as two suites that quietly
|
|
-- test different things.
|
|
--
|
|
-- ── TWO THINGS HERE ARE LOAD-BEARING, NOT SCAFFOLDING ─────────────────────────────────────────
|
|
--
|
|
-- 1. The schema is owned by `sqlpoll_owner` and the suite connects as `otopcua`, which holds
|
|
-- USAGE + SELECT and nothing else. That is the exact provisioning `PostgresDialect.ListSchemasSql`
|
|
-- was written for: `information_schema.schemata` lists only schemas the current role OWNS, so a
|
|
-- granted-not-owning read-only poller would see an empty schema list, `SqlCatalogLoader` would
|
|
-- throw its "the catalog returned no schemas at all" grant diagnosis, and every tag would be
|
|
-- rejected. Sourcing the schema list from `information_schema.tables` instead is what makes this
|
|
-- work — and a fixture where the poller owned its own schema would pass either way, proving
|
|
-- nothing. Do not "simplify" this by creating the tables as `otopcua`.
|
|
--
|
|
-- 2. Every identifier is created UNQUOTED, so Postgres folds it to lower case in the catalog. The
|
|
-- suite's case-folding leg authors the same names in UPPER case and expects them to resolve
|
|
-- through `SqlCatalog.TryMatch` (exact-ordinal first, then a UNIQUE case-insensitive match, which
|
|
-- substitutes the catalog's own spelling). Quoting a name here would make the catalog spelling
|
|
-- mixed case and that leg would stop meaning anything.
|
|
|
|
\set ON_ERROR_STOP on
|
|
|
|
-- ── roles ─────────────────────────────────────────────────────────────────────────────────────
|
|
-- \gexec runs the row the SELECT produced, and produces no row when the role already exists —
|
|
-- the idempotent CREATE ROLE idiom. A DO block cannot be used for the login role: psql does not
|
|
-- interpolate :'app_password' inside a dollar-quoted body.
|
|
SELECT 'CREATE ROLE sqlpoll_owner NOLOGIN'
|
|
WHERE NOT EXISTS (SELECT 1 FROM pg_roles WHERE rolname = 'sqlpoll_owner')
|
|
\gexec
|
|
|
|
SELECT 'CREATE ROLE otopcua LOGIN'
|
|
WHERE NOT EXISTS (SELECT 1 FROM pg_roles WHERE rolname = 'otopcua')
|
|
\gexec
|
|
|
|
-- Set (or re-set) the password every run, so re-seeding with a different SQL_PG_PASSWORD converges
|
|
-- rather than leaving the suite unable to log in.
|
|
ALTER ROLE otopcua LOGIN PASSWORD :'app_password';
|
|
|
|
-- ── schema ────────────────────────────────────────────────────────────────────────────────────
|
|
-- Drop-and-recreate rather than merge-or-insert: the seed is an assertion contract, and a leftover
|
|
-- row from an older revision of this file would make a passing test meaningless. The blast radius
|
|
-- is bounded to this fixture-owned schema.
|
|
DROP SCHEMA IF EXISTS sqlpoll CASCADE;
|
|
CREATE SCHEMA sqlpoll AUTHORIZATION sqlpoll_owner;
|
|
|
|
-- Everything below is created AS THE OWNER, not as the superuser — see note 1 above.
|
|
SET ROLE sqlpoll_owner;
|
|
|
|
-- Key-value (EAV) source: tag_name → num_value, with a source timestamp.
|
|
CREATE TABLE sqlpoll.tag_values (
|
|
tag_name varchar(128) NOT NULL,
|
|
num_value double precision NULL,
|
|
sample_ts timestamp(3) NOT NULL
|
|
);
|
|
|
|
INSERT INTO sqlpoll.tag_values (tag_name, num_value, sample_ts)
|
|
VALUES ('Line1.Speed', 42.0, TIMESTAMP '2026-07-24 10:00:00'),
|
|
('Line1.Temp', NULL, TIMESTAMP '2026-07-24 10:00:01'),
|
|
('Line1.Pressure', 3.5, TIMESTAMP '2026-07-24 10:00:02');
|
|
|
|
-- Wide-row source: one row per station, several value columns. Station 8 is inserted LAST and is
|
|
-- also the newest sample, while station 7 is first — so a topByTimestamp plan that lost its
|
|
-- ORDER BY … DESC (or its LIMIT 1) returns a different, wrong value rather than a coincidentally
|
|
-- right one.
|
|
CREATE TABLE sqlpoll.latest_status (
|
|
station_id integer NOT NULL,
|
|
oven_temp double precision NULL,
|
|
pressure double precision NULL,
|
|
sample_ts timestamp(3) NOT NULL
|
|
);
|
|
|
|
INSERT INTO sqlpoll.latest_status (station_id, oven_temp, pressure, sample_ts)
|
|
VALUES (7, 180.5, 1.2, TIMESTAMP '2026-07-24 10:00:00'),
|
|
(9, NULL, 1.6, TIMESTAMP '2026-07-24 10:00:01'),
|
|
(8, 210.25, 1.4, TIMESTAMP '2026-07-24 10:00:02');
|
|
|
|
RESET ROLE;
|
|
|
|
-- ── grants: USAGE + SELECT and nothing else ───────────────────────────────────────────────────
|
|
GRANT USAGE ON SCHEMA sqlpoll TO otopcua;
|
|
GRANT SELECT ON ALL TABLES IN SCHEMA sqlpoll TO otopcua;
|