Repository navigation
feat(compat): four more wire-compatible engines, measured rather than assumed (#424) - #442
Merged
Merged
Conversation
… assumed (#424) Phase 0 of #424 recorded TimescaleDB, YugabyteDB, TiDB and StarRocks as unmeasurable. Three of those four reasons no longer hold, and one was never true: TiDB was recorded as needing "a PD + TiKV cluster, no single-container path", and `--store=unistore` runs it in one container with neither. TimescaleDB and YugabyteDB were recorded against a failed image pull, which is a property of that day's link and not of the engine. Only StarRocks' reason was accurate, and it expired with the link. All four were probed live through the real providers, every introspection surface called separately against a plain PostgreSQL 18 and MySQL 8 baseline, then re-run in a browser against the compose services added here. TimescaleDB 2.29.2 / PG 17.11 via postgres full YugabyteDB 2.25.2.0 via postgres full TiDB v8.5.1 via mysql full StarRocks 3.3.22 via mysql partial (11 of 15 surfaces) Every caveat is measured. The ones worth naming here are the ones that mislead rather than fail: a TimescaleDB hypertable reports 0 rows and 24 KB for 5000 rows and 2.4 MB, because the statistics describe the empty parent and the data lives in 31 chunks; YugabyteDB reports 0 rows and 0 bytes until someone runs ANALYZE; and StarRocks' version() returns a fictitious 5.1.0, so the product displays "MySQL 5.1" for a 3.3.22 server. Confirmed in the UI, not just the harness: the TimescaleDB explorer shows 61 objects for 2 user tables, with metrics at 0 beside plain_rows at 100. Two caveats exist only because gate 5 ran: TiDB and StarRocks both reject EXPLAIN FORMAT='json', so the Explain panel fails while the query itself runs. Gate 7 passes on TiDB (grounded on the real orders.country, ran unmodified, answered TR), YugabyteDB and StarRocks. It does NOT pass on TimescaleDB, and that is recorded as B52 rather than glossed: the composed PostgreSQL grounding capture is one row per column against a 200-row budget, and the extension's own catalogs answer 473 of 478 rows before the user creates anything. Verified extension-caused, not path-caused -- plain PostgreSQL 18 and YugabyteDB both captured their schemas under the same least-privilege role, and granting that role the internal schemas changed nothing. Verified count 23 -> 27 (14 shipped drivers + 13 probed relatives). Two stale numbers in WireCompatibilityHint corrected on the way: the dialog offers fourteen drivers, not eleven, and the caveat-count rhetoric is now count-free so it stops ageing.
|
cevheri
added a commit
that referenced
this pull request
Aug 20, 2026
…two thirteens colliding (#424) (#444) Phase 0's wrap-up asks that the published engine claims agree. They did not, and the sharpest case was one this repo created for itself: taking the wire-compatible relatives from nine to thirteen in #442 put a second, disjoint "thirteen" next to the thirteen shipped drivers, in both README.md and DOCKERHUB.md, with only the word "more" separating them. Both now say which thirteen they mean, and the README states the total the two make - thirteen drivers reaching twenty-six named engines. DOCKERHUB carried the same fault twice over: "Thirteen engines" sat above a table with fourteen rows, the fourteenth being the embedded LibreDB store, and the relatives table below it stated no number at all while rendering thirteen engines in ten rows. Both are now explicit about their denominator and about the merged rows. Genuinely stale copy, each verified against the code rather than assumed: - The two chart READMEs still named SEVEN engines while their own sibling Chart.yaml named thirteen. That is the string ArtifactHub shows, so it undersold the product to exactly the reader who checks. Fixed to Chart.yaml's own enumeration so the two cannot drift apart again. - The operator CSV keywords named seven engines, leaving Couchbase, ClickHouse, Druid, Elasticsearch, OpenSearch and Trino invisible to OperatorHub search. Added to operator/config and propagated with `make -C operator bundle`; the bundle diff is the six keywords and nothing else. - deploy/rancher/E2E_VALIDATION_TASK.md told a human tester to expect Chart.yaml to name seven engines against an app-readme of ten. Chart.yaml names thirteen. The instruction now describes the mismatch that actually exists. - The winget locale contradicted itself inside one file: ShortDescription said thirteen engines, the body enumerated seven as though complete. Deliberately NOT changed: the Rancher catalog prose still says ten engines. That wording is version-scoped and frozen on purpose - the catalog page describes a released version and waits for a tag to carry Elasticsearch, OpenSearch and Trino - and its own accuracy gate says so. A sweep that reads "ten" as staleness would undo a decision. Two relatives are now named in the marketplace listings, under a rule the owner settled: a relative is named only where its caveats are cosmetic. So TiDB and FerretDB join MariaDB and the three Redis forks in the flatpak/flatpark metainfo and the Azure description, and Citus, TimescaleDB and YugabyteDB stay out despite being full-tier, because they report row counts that are wrong rather than missing. StarRocks, CockroachDB, Materialize and RisingWave stay out on their tiers. Chart version 0.1.40 -> 0.1.41: charts/libredb-studio/README.md is a packaged file and 0.1.40 is already published, so #167 requires the bump. appVersion stays 0.12.0. The operator's chart copy moves with it, being byte-identical under two guards. B53 records what this PR deliberately did not fix: ten sites say the composed grounding path leaves "the other nine engines" to their providers, which is stale, but the replacement is not the obvious one. Nine was fourteen type-ids ago; the external reading gives eleven and the original basis gives twelve, the sites never say which they count, and B49 means one of the twelve cannot be grounded at all. Picking a digit silently would repeat the fault this PR exists to fix.
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 (wire-compatible labelling) — no driver code, only verification and honest
publishing. Four engines the epic recorded as unmeasurable have now been probed live.
Why these four were open, and why they are not any more
--store=unistoreruns it in one container with no PD and no TiKV.A deferral records a reason, and the reason ages at a different rate than the decision. Three of
these four had expired.
What was measured
Probed through the real providers via
createDatabaseProvider, every introspection surface calledseparately, against a plain PostgreSQL 18 and MySQL 8 baseline so every number is a delta.
postgrespostgresmysqlmysqlversion()reports 5.1.0)The caveats that matter are the ones that mislead rather than fail:
2.4 MB. The statistics describe the empty parent; the data lives in 31 chunks. Same class as the
Citus caveat already recorded. The explorer also shows 61 objects for 2 user tables.
ANALYZE. Measured on a 20000-row table: 0 before,20000 rows and 7878 kB after. A full database looks empty.
version()returns a fictitious5.1.0, so the product displays "MySQL 5.1"for a StarRocks 3.3.22 server. Overview, health, sessions and monitoring all unavailable.
InnoDBentry on a server that has no InnoDB.Gate 5 found what the harness could not
TiDB and StarRocks both reject
EXPLAIN FORMAT='json', so the Explain panel fails while the queryitself runs normally. The provider harness never calls explain; only driving the editor in a browser
surfaced it. Both now carry the caveat.
Gate 7, including the one that fails
Grounded plan runs, each against the live container behind gate 4:
orders.country, ran unmodified, answeredTR.public.orders.orders.The composed PostgreSQL grounding capture is one row per COLUMN against
maxResultRows: 200andrefuses rather than truncates. A stock TimescaleDB answers 478 rows from
information_schema.columns, of which 473 belong to the extension (timescaledb_information163,_timescaledb_internal146,_timescaledb_catalog139,_timescaledb_config17,timescaledb_experimental8) and 5 to the user. So the agent is unusable there before the usercreates a table.
Verified extension-caused rather than path-caused, with three controls: plain PostgreSQL 18 captured
3 tables, fingerprint ctx_0d63under the same least-privilege role; YugabyteDB captured its schematoo; and granting the agent role USAGE and SELECT on the three internal schemas changed nothing. The
existing code comment estimates this limit as "roughly 25 tables of eight columns", which frames it
as a big-schema problem — B52 records the measurement that reframes it as a catalog-heavy-extension
problem.
Counts and incidental fixes
Verified count 23 → 27 (14 shipped drivers + 13 probed relatives); the README's relatives line
moves from "Nine more engines" to "Thirteen more engines". Two stale numbers in
WireCompatibilityHintcorrected on the way: the dialog offers fourteen drivers rather thaneleven, and the caveat-count rhetoric ("sixteen caveat lines", true of neither the old registry nor
the new one) is now count-free so it stops ageing.
database-compose.ymlgains all four underprofiles: ["compat"], and every browser check in thisdescription was run against those compose services rather than hand-started containers.
Gates
format·lint(0 errors) ·typecheck·knip·test(all 30 groups) ·build— all green.Coverage
39181/39181 lines (100.00%). The repo's bidirectional agent-documentation guard caughtB52 missing from
docs/AGENT.mdand it is now cited in both files.Closes nothing; ticks four boxes in #424.