Thread

Re: Row pattern recognition

Henson Choi <assam258@gmail.com> — 2026-05-07T12:41:46Z

Hi Tatsuo, hi Jian,

Attached are 10 incremental patches on top of v47, covering the
follow-up plan ([A]..[D]) I sketched on 2026-05-05, Jian's two
coding-style items (trailing commas, outer parentheses; gram.y
folded into 0007), one comment-only patch (0009), and one trivial
cleanup (0010).  Packaging into the next version is at your
discretion.

Patch summary:

  - 0001 Add DEFINE non-volatile baseline to rpr_integration B9
      Adds a STABLE/IMMUTABLE baseline ahead of the volatile
      case so the rejection in 0002 is visible against a working
      reference.

  - 0002 Unify RPR DEFINE walkers and reject volatile callees
      [A].  Collapses the four DEFINE walkers (parser / planner
      / executor) into one phase-tagged traversal and, on that
      base, rejects volatile / NextValueExpr at parse-analysis.
      STABLE / IMMUTABLE accepted.  The offset runtime-constant
      check is folded in; B9 flips to error-case and its XXX is
      dropped.

  - 0003 Cover RPR empty-match path with EXPLAIN tests; fix
         stale XXX comments
      [B].  Adds rpr_explain coverage that surfaces "NFA: N
      matched (len 0/0/0.0)" so the NFA-found-but-empty-frame
      path is regression-visible; replaces stale rpr_nfa XXXs.

  - 0004 Reclassify DEFINE qualifier check and reword
         diagnostic to "expression"
      [D].  Tri-classifies (pattern var / range var / fall-
      through) so unknown qualifiers fall through to the
      standard "column does not exist" diagnostic.  Reworded to
      "expression" since the quoted token may include
      indirection on composite types (e.g. (A.items).amount).

  - 0005 Sync stale comments on DEFINE/PATTERN handling
      validateRPRPatternVarCount() rejects DEFINE names not in
      PATTERN, but comments / SELECT doc described it as
      "collection" / "filtered during planning".  Wording
      aligned; XXX note's "column references" bumped to
      "expressions" to match 0004.  Comment + doc only.

  - 0006 Add trailing commas to RPR enum definitions
      Per Jian.  RPRNavKind, RPRNavOffsetKind, RPRPatternNodeType.

  - 0007 Remove optional outer parentheses from ereport() calls
         in RPR files
      Per Jian + Tatsuo's gram.y patch.  19 sites in
      parse_rpr.c / optimizer/plan/rpr.c plus the gram.y sites.

  - 0008 Add high-water mark tracking to NFA visited bitmap
         reset
      [C].  Resets only the touched span (O(span_words) instead
      of O(numElements/64)), at the cost of two int16 comparisons
      per visit.  Semantics unchanged; happy to drop if you
      prefer the simpler bulk reset.

  - 0009 Document DEFINE subquery rejection as intentional
         over-rejection
      Comment-only.  Records that ISO/IEC 9075-2 §4.18.4 /
      6.17.4 (R010 / R020) permits a non-RPR, non-pattern-var-
      correlated subquery in DEFINE, and that our blanket
      rejection is deliberate over-rejection.  The case
      distinction (walk subquery Query for nested RPR; match
      ColumnRef qualifiers against ancestor p_rpr_pattern_vars)
      is doable with existing infrastructure and left as future
      work, not blocked on any other feature.

  - 0010 Remove duplicate #include in nodeWindowAgg.c
      "common/int.h" was included twice and the first occurrence
      was misplaced between catalog/* headers.  Drops the
      misplaced first; the second sits at the correct
      alphabetical position.

Still deferred:
  - B7 Recursive CTE XXX: pending community input on the
    §4.18.5 / 6.17.5 interpretation.

Please let me know if any of 0001-0010 looks misjudged or if you
would prefer a different slicing.

Best regards,
Henson