A second corpus: the statements a real Tableau emitted, replayable without Tableau

[fupelaqu]

A second corpus: what a real Tableau emitted, replayable without Tableau

Until now, "does Tableau work?" could only be answered by installing Tableau. This adds the statements a
real Tableau Desktop 2026 for macOS actually sent, as a tracked test resource with its own replay
suite, so a clean checkout can tell you when Tableau support regresses.

Captured 2026-09-13 over the JDBC driver 0.3.3-SNAPSHOT (core 0.23.0-SNAPSHOT) against real
Elasticsearch 8.18.3, with p6spy at the JDBC boundary. 127 statements across six sessions → 26
distinct shapes. Normalisation is applied to the grouping key only; every committed statement is
verbatim.

This is a SECOND corpus, deliberately

epic-19-bi-corpus.csv is untouched, and so are epic-21-attribution.csv, baseline-pre-epic21.csv
and CorpusReplaySpec (byte-identical; CorpusReplaySpec stays 12/12 with its summary line
unchanged). That corpus is a fixed baseline: the published 56/99 and 56/75 and the Epic 23 series
are anchored to it, and appending would silently move denominators the epic already published. The new
file has no baseline, no series and no denominators.

What the rows are

owner	rows	meaning
works	14	preview, aggregate, filter, sort, SELECT 1, SELECT * … LIMIT 1
issue:328	5	the HAVING (COUNT(1) > 0) family — must parse; this is the regression guard for the fix that unblocked Tableau's data-source preview
epic22a_derived_table	3	the connect-time SELECT … FROM (SELECT 1 AS \COL`) AS `SUBQUERY`` — rejected, Epic 22 owns it
capability_open	2	the temp-table probe pair
rejected_by_design	2	GROUP BY 2 against a one-item SELECT list — an out-of-range ordinal, correctly rejected

21 of the 26 parse; 0 throw; 0 come back as an internal parser error. 14 shapes are absent from the
Epic 19 corpus.

Three things the capture corrected:

• The MySQL dialect's temp-table probe is a plain CREATE TABLE, not CREATE TEMPORARY TABLE. It
  parses and is refused at execution by the index-name validator. The parse-time refusal added in Tableau's temp-table probe is refused at the wrong layer, for a misleading reason, and one line from creating the indices for real #326
  covers the SQL-92 TEMPORARY spelling, which this dialect never emits. The suite guards the label and
  the verdict; refusing these would be a product decision, not a bug fix.
• The connect-time derived table is the most frequent shape in the whole capture — 17 of 127, once
  per connection.
• GROUP BY 2 against a one-item SELECT list is Tableau's ordinal capability probe's negative
  control, issued immediately after GROUP BY 1. Our rejection, naming the valid range, is the
  correct answer.

Gate integrity

Four id sets and the parse total are pinned in compiled code, asserted two-way against the CSV.
Falsified with 9 mutations, 9 predicted REDs, restored by captured bytes. The one that justifies
pinning the total: a statement that regresses while its expected cell is edited to absorb the
change passes the verdict gate — only the pinned count catches it.

The suite reuses CapturedSqlProbe.parseCsv, CorpusReplay.replayAll and CorpusReplay.Outcome. No
second codec, no second replay loop.

What this does NOT tell you

It is parse-only by construction — the sql module has no Elasticsearch client — so it proves these
shapes are accepted, never that they answer correctly. The README names which rows already have
oracles in GroupByCompletenessSpec and which have none; four works rows have none at all. That is a
clean pickup for a later story and the honest next step for any claim about Tableau.

It also records what Tableau emitted for those interactions on that version. It is not a substitute
for running Tableau, and the README says so.

⚠️ Recorded and deliberately not explained: this capture has Tableau emitting GROUP BY 1 where the
August corpus has GROUP BY \t`.`category`` for the same interaction. A control run rules out the
Tableau customization file; the cause is unestablished, so both spellings are recorded and neither is
given a story.

Verification

TableauLiveReplaySpec 11/11 · CorpusReplaySpec 12/12 · sql/test 1104 · + compile · the CI lint
line green · scalafmtAll with files staged first. Tests on 2.13 only, as CI runs them. No ES legs —
parse-only by construction.

Content scan over all 26 statements: zero hits for host, credential or user patterns. The only
identifiers present are the synthetic fixture's tables and columns, Tableau's own generated names, three
country literals, and the test cluster name.

🤖 Generated with Claude Code