fix(transfer): stop quoting simple identifiers when migrating to GaussDB/OpenGauss - #6283

Pushed a fix for this. Summary of what changed and what's still open:

Keyword catalog gap — fixed. Checked the raw HTML of Huawei's official GaussDB(DWS) keyword reference (https://support.huaweicloud.com/intl/en-us/sqlreference-dws/dws_06_0007.html) directly rather than relying on a rendered/summarized view, and confirmed COMPACT is listed as Reserved (functions and types allowed) — exactly the case you flagged. Diffing that table against our existing Postgres reserved-word list turned up 24 more words GaussDB reserves that Postgres doesn't (AUTHID, BUCKETS, COMPACT, DELTAMERGE, FENCED, HDFSDIRECTORY, HOT, INTERNAL, LESS, MAXVALUE, MINUS, MODIFY, NLSSORT, PERFORMANCE, PLAN, PROCEDURE, RECYCLEBIN, REJECT, SYSDATE, TIMECAPSULE, TSTAG, TSTIME, TSFIELD, WARMUP).

Added is_gaussdb_only_reserved_identifier with that set and consulted it alongside the Postgres list, scoped to GaussDB/OpenGauss only, in both quoting call sites:

• quote_transfer_identifier (the transfer DDL/INSERT path this PR touches)
• quote_gaussdb_jdbc_identifier (pre-existing helper for live JDBC identifier quoting — it turned out to also serve plain Postgres connections, so it now takes database_type explicitly so the extra GaussDB words don't leak into real Postgres quoting decisions)

Added regression tests for target-only reserved words (compact, buckets, sysdate, minus, modify, plan, procedure) on both GaussDB and OpenGauss, plus a test confirming the same words are correctly not quoted on a real Postgres target — so the dialect split is exercised both ways, not just the GaussDB side.

Live-instance validation — not done, flagging honestly. I don't have access to a real GaussDB/openGauss instance. I tried spinning one up locally via docker pull --platform linux/amd64 opengauss/opengauss:5.0.0 (openGauss only ships amd64 images) on Apple Silicon, but the pull never completed in a reasonable time under emulation, so I couldn't get a live instance to execute DDL against.

One more caveat on the keyword source itself: the table I diffed against is Huawei's GaussDB(DWS) (the MPP/columnar variant) keyword reference — the core GaussDB/openGauss keyword doc pages are client-side rendered and I couldn't get a raw table out of them to cross-check directly. A few of the 24 added words (TSTAG, TSTIME, TSFIELD, HDFSDIRECTORY) are explicitly flagged in that same doc as used only by the hybrid-data-warehouse feature, so they may never actually appear as column names outside DWS — kept them in anyway since over-quoting is harmless (a strict superset of what's needed), same reasoning as the original PR's is_postgres_reserved_identifier reuse.

If you have access to a GaussDB or openGauss instance (or can point me at one), I'd appreciate either running the added test cases' DDL directly against it, or sharing connection details so I can validate myself before this merges.

I have a writable openGauss 5.0.0 instance available, so I ran the verification myself rather than asking you to find one.

Method: queried the instance's own pg_get_keywords() (the authoritative source both openGauss docs and I pointed to) for the 24 words this PR's is_gaussdb_only_reserved_identifier treats as GaussDB-only reserved, and cross-checked with an actual CREATE TABLE using each word unquoted.

Result: of the 24 words, 9 are not actually reserved on this real instance:

The remaining 15 (authid, buckets, compact, deltamerge, hdfsdirectory, less, maxvalue, minus, modify, performance, procedure, recyclebin, reject, sysdate, timecapsule) are genuinely reserved (reserved or reserved (can be function or type name)), matching the DWS doc. I also confirmed end-to-end: an unquoted CREATE TABLE ... (compact int) fails with ERROR: syntax error at or near "compact" on this instance, confirming compact does need quoting.

Why this matters for the PR: this confirms the concern from the previous review — the DWS keyword catalog is not a safe superset for core GaussDB/openGauss. Right now is_gaussdb_only_reserved_identifier will force-quote fenced, internal, plan, tsfield, tstag, tstime, hot, nlssort, and warmup even though none of them need it on this instance — reintroducing the exact bug this PR sets out to fix (quoting locks in case, breaking later unquoted references) for any table using one of those as a column name.

Suggest trimming the set down to the 15 confirmed-reserved words, or better, deriving it from pg_get_keywords()/SELECT word FROM pg_get_keywords() WHERE catcode <> 'U' at connect time (or a version-gated static list sourced from that catalog) rather than a hand-diffed DWS doc snapshot.

Happy to share the exact test queries I used if useful.

Pushed a fix acting on the instance-verification results from my last comment.

is_gaussdb_only_reserved_identifier is trimmed from 24 to the 15 words confirmed reserved on that real openGauss 5.0.0 instance (authid, buckets, compact, deltamerge, hdfsdirectory, less, maxvalue, minus, modify, performance, procedure, recyclebin, reject, sysdate, timecapsule). Dropped the 9 words that instance's pg_get_keywords() showed are not actually reserved (hot/nlssort/warmup aren't keywords at all; fenced/internal/plan/tsfield/tstag/tstime are unreserved) — keeping them would have reintroduced this PR's original bug (over-quoting locks in case).

Updated the doc comment on is_gaussdb_only_reserved_identifier to point at the real-instance verification instead of the DWS doc diff, and extended quotes_gaussdb_only_reserved_words_not_shared_with_postgres to assert on all 15 confirmed words plus a negative check that the 9 dropped words are not quoted.

cargo test -p dbx-core --lib sql_dialect (100 tests), cargo fmt, and cargo clippy -p dbx-core --lib -- -D warnings all pass clean locally.

Live-instance DDL validation for this specific diff (as opposed to the standalone keyword checks I ran) is still outstanding — happy to run it if you can point me at an instance, otherwise let me know if the trimmed list itself needs anything else before merge.

Pushed a fix acting on the review feedback about the hand-picked list.

Instead of spot-checking individual words again, diffed is_gaussdb_only_reserved_identifier's existing 15-word list against the full pg_get_keywords() result set from the same writable openGauss 5.0.0 instance (653 rows) rather than checking words one at a time. Found 7 more genuinely reserved words the hand-picked list had missed: csn, excluded, groupparent, nocycle, rownum, shrink, verify — all confirmed reserved (R) or reserved, can be function or type name (T) on that instance.

Extended quotes_gaussdb_only_reserved_words_not_shared_with_postgres to cover all 22 words (and the negative Postgres check for the 3 newly added ones that also happen to look like plain words there). Also added table/schema-name coverage to the existing CREATE TABLE regression test (previously only checked column names), and a new gaussdb_insert_quotes_target_only_reserved_words test since the INSERT path shares quote_transfer_identifier but had no dedicated regression coverage of its own.

cargo test -p dbx-core --lib sql_dialect (100 tests) and cargo test -p dbx-core --lib transfer:: (201 tests) both pass clean locally.

Live-instance DDL/INSERT validation for this specific diff is still outstanding, same caveat as before.

Thanks for the detailed catch on maxvalue — pushed a fix that makes the keyword decision version-aware, using the approach you suggested.

What changed

• db::postgres::gaussdb_reserved_keywords runs SELECT word FROM pg_get_keywords() WHERE catcode IN ('R','T') against the target connection once, and AppState::gaussdb_reserved_keywords caches the result per pool_key. When available, this live, per-server catalog is authoritative and replaces (never unions with) the static hand-diffed list — so a word that's reserved on 5.0 but not on a newer version is quoted differently per target, instead of being permanently locked to the 5.0 snapshot. Falls back to the static list when no live catalog is available (e.g. insufficient privileges, an engine without pg_get_keywords()), matching the pre-existing behavior in that case.
• Only a deterministic failure (the query ran and errored or came back empty) is cached — a transient one (connection checkout error, timeout) is retried on the next table rather than pinning the connection to "unavailable" for its whole lifetime.
• Threaded through both the CREATE TABLE DDL and the INSERT/UPSERT write paths, for both the SQL-source and MongoDB-source transfer flows, including the Overwrite-mode TRUNCATE/DELETE step (this one I initially missed in my own self-review — it was still using the static list while CREATE TABLE next to it already used the live catalog, which would have made the two disagree on the same table name; fixed and covered by a regression test).

Regression coverage

Added tests that inject a synthetic 5.0-shaped keyword set (maxvalue present) and a current-shaped one (maxvalue absent) and assert the CREATE TABLE, INSERT, and TRUNCATE paths all quote maxvalue accordingly — plus a None/no-live-catalog case confirming the static-list fallback is unchanged.

Real-target validation

Ran this end-to-end against two real openGauss instances:

• openGauss 5.0.0 — pg_get_keywords() reports maxvalue as catcode='R' (reserved), matching the 5.0 kwlist.h. The generated DDL quotes it; created the table and confirmed an unquoted reference is correctly rejected by the parser there (it's genuinely reserved), while a quoted reference resolves.
• openGauss 7.0.0-RC1 — pg_get_keywords() reports maxvalue as catcode='U' (unreserved). The generated DDL leaves it unquoted; created the table and confirmed an unquoted reference resolves normally — the exact case-locking failure mode from [Bug] 数据传输时，源表字段及表名带引号或者反引号时，迁移到目标库时字段及表名能否去掉引号及反引号 #6205 no longer reproduces on this version.

Ready for another look whenever you have a chance.