feat(providers): publish the nine wire-compatible engines a live probe measured (#424) - #426
Merged
Conversation
…e measured (#424) Eleven drivers, but a MariaDB user had no way to know that MySQL is the right button. This adds a compatibility registry, a hint in the connection dialog and a docs table - all from one source - covering the nine engines that were connected to and measured, and nothing else. The registry records what a probe found, not what a vendor claims: - `probedVersion` is the string the server itself returned. A unit test fails on an entry without one, which turns #424's "connects is not supported" rule into a CI gate instead of a paragraph. - `tier` exists because the data refuted a boolean. Citus answers all fifteen introspection surfaces; Materialize answers none while still running SQL fine. Publishing both as "compatible" is the overclaim the issue forbids, so query-only and partial engines are marked as such in the dialog itself. - `caveats` carry what breaks or misleads. Citus is the sharpest case: every surface answers, and the row counts it answers with are silently WRONG for a distributed table. A recorded assumption did not survive contact with the engine. #424 said CockroachDB's `pg_stat_*` surface was the risk; every pg_stat-driven surface answered, and the eight failures are all missing size/uptime/tablespace builtins (`pg_total_relation_size`, `pg_size_pretty`, `pg_postmaster_start_time`, `pg_tablespace_location`). That is a narrower and more fixable shape than the one we wrote down. Also fixes the one hole in the shipped set: `redis` had no `database-compose.yml` service, so its own gate-4 probe could not be reproduced on a clean machine. The compat instances sit behind `profiles: ["compat"]`, so a plain `up` is unchanged. Six names with a possible local instance stayed unprobed (image budget, cluster requirements, a licence key) and every managed-only service has no credentials here. None of them is ticked or published - untested is recorded as untested.
100 tasks
…#424) `dialog.locator("text=PostgreSQL")` matched a substring, so it broke the moment the dialog gained the compatibility hint: three probed version strings mention PostgreSQL ("citus 14.1-1 on PostgreSQL 18.4", "advertises PostgreSQL 9.5", "advertises PostgreSQL 13.14.0"), and Playwright refused four matches. The assertion always meant "the driver selector offers PostgreSQL", so it now says that: getByRole("button", { name: "PostgreSQL" }). Narrowing the version strings instead would have deleted measured facts to accommodate a fragile locator.
…try (#424) The compatibility table in docs/providers/README.md is hand-written prose beside machine-readable data, which is exactly how the two drift. The new test asserts every registry entry appears in the table with its probed version - the tri-sync rule applied to compatible names, which have no doc page of their own to sync. It found a drift on its first run: the table listed DragonflyDB's version as "df-v1.40.1" while the registry says "DragonflyDB df-v1.40.1". Fixed. Two other corrections from review: - database-compose.yml now says that CockroachDB's `network_mode: host` is Linux only. On Docker Desktop `host` is the VM, not the machine, so 127.0.0.1:26257 never answers there and the entry looks broken for a reason nothing explains. - sonar-project.properties excludes compatibility.ts from copy-paste detection, with the reason written down. Nine entries sharing one key sequence read as 56 duplicated lines; compressing them into tuples plus a mapper would satisfy the detector and make an auditable published claim unreadable.
… have (#424) Two corrections from review, both about a claim being wider than the thing it describes. The docstring said the README, docs and channel listings "all quote this number". They do not: they are markdown and quote it as prose, so the function has no runtime consumer at all. It stays because the arithmetic deserves one definition that a test pins rather than three hand-counts, but the comment now says that instead of implying a wiring that does not exist. The component fixture used TiDB as an invented query-only engine with an invented version. TiDB is a name #424 records as explicitly NOT measured, so the fixture contradicted the record it sits beside and read like a probe result to anyone skimming. Renamed to ExampleStore, which cannot be mistaken for a product.
|
This was referenced Aug 20, 2026
cevheri
added a commit
that referenced
this pull request
Aug 21, 2026
…ts to 32 (#454) The login hero derived its engine pills from `DB_UI_CONFIG` and so was already current - Apache Cassandra and Trino were on the page. Two things were not. The claim counted the wrong set. `HERO_CLAIMS` read `listShowcaseDatabases().length`, which includes `libredb`: the embedded store this app carries, not a database anyone points it at. The page therefore published one external engine more than the product has. `EXTERNAL_DATABASE_TYPES` now defines the split beside `SHIPPED` in `compatibility.ts`, as an exhaustive `Record<DatabaseType, boolean>` so a new type-id cannot join without someone answering the question, and the claim reads 14. The pill stays and is marked `(embedded)`: it is a provider the connection picker offers, so hiding it would put the login page at odds with the app. The eighteen wire-compatible relatives were nowhere. MariaDB, Citus, TimescaleDB, Valkey, Vitess, FerretDB and the rest were published in README.md and the docs compatibility table but on no surface a visitor meets first. `WireCompatibleLine` names all eighteen from `WIRE_COMPATIBLE_ENGINES`, with the count as a `.length`, on both the desktop hero and the mobile block. It carries no tier word: the per-engine tier belongs to the connection dialog's hint, which has room to qualify it, so the line claims a measurement instead of parity. `verifiedEngineCount()` becomes `connectableProductCount()`, and the rename is the fix. It summed the shipped drivers, so it answered 33 while README.md published 32 for the same claim - a count is wrong when its denominator is unstated, not when its digit is stale. Two things found on the way. `tests/components/WireCompatibilityHint.test.tsx` shipped with #426 and was in no run group, so its seven tests had never run in CI once; it gets its own group, because it mocks `@/lib/db/compatibility` and `mock.module` is process-wide. And `TOTAL_GROUPS` read 30 against 32 calls, so every green summary reported a count no run had. Measured on the built app rather than argued: at 1440x900 and 1920x1080 the page does not scroll, and the sign-in card is above the fold at every size tested. At 1280x800 it now scrolls 56px, because that column measured exactly 800px before this change - zero slack - so anything added there costs height. 104px of it was reclaimed by folding the line into the pills' block, `leading-snug`, and a shorter lead sentence. The residual, with its numbers and the candidates for the rest, is U18 in docs/BACKLOG.md.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 0 of #424. Eleven drivers, but a MariaDB user had no way to know that MySQL is the right button. This adds a compatibility registry, a hint in the connection dialog and a docs table - all reading one source - covering the nine engines that were connected to and measured, and nothing else.
What is here
src/lib/db/compatibility.ts- the single source of truth: name, the driver that serves it, the support tier, the version a live probe reported, and the caveats it found.src/components/WireCompatibilityHint.tsx- renders under the driver selector in the connection dialog.docs/providers/README.md#wire-compatible-engines- the per-engine table, with a reproduction command.database-compose.yml- the ten instances the probes ran against, behindprofiles: ["compat"], plus aredisservice that was missing.The measurements
mysqlpostgresredisredisredismongodbpostgrespostgrespostgresEach probe ran through the shipped provider - never a native CLI - and called all fifteen introspection surfaces separately, because a compatible engine answers the connection and then diverges exactly there. PostgreSQL 18 and MySQL 26.7.0 were probed first as baselines, both clean, so every number is a delta against a known-good run.
Three things worth a reviewer's attention
A recorded assumption in #424 was refuted. The issue said CockroachDB's
pg_stat_*surface was the risk. Every pg_stat-driven surface answered; all eight failures are missing size/uptime/tablespace builtins (pg_total_relation_size,pg_size_pretty,pg_postmaster_start_time,pg_tablespace_location). Narrower and more fixable than what we wrote down.The data refuted a boolean, so
tierexists. Citus matches PostgreSQL surface for surface while Materialize answers no introspection call at all. Publishing both as "compatible" is the overclaim the issue forbids, so the dialog prints the qualifier next to the name."Full" can still mean wrong numbers. Citus answers everything, and its answers are silently incorrect for a distributed table: 0 rows / 8 KB against a real 50 rows across 32 shards. A missing panel is honest; a populated wrong one is not - hence the caveat.
Claim discipline, as a test rather than a paragraph
tests/unit/db/compatibility.test.tsfails on an entry without a probed version string, and on a query-only or partial entry with no caveat. #424 says a name counts only after a live probe; that rule is now a CI gate.What is not here
Six names with a plausible local path did not clear gate 4 (YugabyteDB, TimescaleDB - pulls exceeded budget; TiDB - needs a PD + TiKV cluster; SingleStore - licence key; StarRocks, OceanBase - multi-GB images), and every managed-only service has no credentials on this machine. None is ticked or published. TimescaleDB is the tempting one: assuming it works because it is "real PostgreSQL" is exactly what gate 4 refuses.
Verified count moves from 11 to 20 against a Phase 0 ceiling of 47. The README headline still says "ten engines" on purpose - a query-only engine is not an engine with a full interface, so the compatible names are a separate, qualified claim.
Verification
All six local gates pass, plus the coverage gate:
The e2e spec (
e2e/wire-compatibility.spec.ts) covers the hint in a real browser, including the negative case: SQLite has no wire protocol, so it must show no hint at all.Findings in full: #424 (comment)