v0.106.0

0.106.0 (2026-10-11)

⚠ BREAKING CHANGES

• isthmus: Unsupported unary preferences are rejected. Export
  names SILENT for integer negate/abs and NAN for FP64 asin/acos.
  Consumers without those policies must reject these plans. Undeclared
  values fail even with a supported fallback.
• isthmus: Unsupported floating-point preferences are rejected.
  Exported arithmetic calls gain rounding=TIE_TO_EVEN; divide also gains
  on_domain_error=NAN. Consumers without those policies must reject these
  plans. Undeclared values fail even with a supported fallback.
• isthmus: Unsupported decimal preferences are rejected. Export
  names SILENT explicitly, so consumers supporting only ERROR must reject
  these plans. Undeclared values fail even with a supported fallback.
• isthmus: Unsupported integer preferences are rejected. Export
  adds single-value arithmetic options, except overflow on plain i8/i16
  operators. Consumers without those policies must reject these plans.
  Undeclared values fail even with a supported fallback.
• isthmus: Default SQL-to-Substrait conversion rejects
  CONTAINS_SUBSTR, and default Substrait-to-Calcite conversion rejects
  contains. Provide a semantics-preserving custom mapping if these
  functions are needed.
• isthmus: Unsupported preferences, including INITCAP charsets,
  fail import. Explicit export options may require consumers to reject the
  invocation.
• isthmus: UPDATE conversion now throws
  UnsupportedOperationException for input pipelines it cannot preserve,
  including target scans with incompatible row types or correlation
  bindings that cannot be transferred safely. Callers must handle the
  unsupported conversion instead of relying on the previous output, which
  could discard selection conditions, retain unresolved outer references,
  or change how often nondeterministic expressions run.
• Mandatory read and post-join filters are now enforced
  by Calcite and Spark conversion. Post-join field references are decoded
  against the join's direct output before emit, rather than concatenated
  inputs. Regenerate plans that reference removed semi/anti input columns
  or use concatenated-input indices for right-oriented joins; outer-join
  references now retain their nullable output types.
• core: All five helpers (min, max, avg, sum, and sum0) now
  derive output types from the supplied extension catalog, so custom
  catalog return declarations can change the result schema. With the
  default catalog, SUM over i8/i16/i32 or fp32 now reports the wider type,
  and floating-point SUM0 reports fp64 instead of i64. Callers that depend
  on the previous output schemas must use the returned types. The
  sum0(Rel, int) overload now emits sum0 with its required result type
  instead of sum with a nullable result. The Spark module currently lacks
  a sum0 mapping, so plans using this overload cannot be converted to
  Spark until that mapping is added.
• core: CustomComparison.of(int), getCustomFunctionReference(),
  the generated customFunctionReference(int) builder method, and
  withCustomFunctionReference(int) are replaced by
  of(ScalarFunctionVariant), getDeclaration(),
  declaration(ScalarFunctionVariant), and
  withDeclaration(ScalarFunctionVariant). Supply the comparator
  declaration from your extension collection instead of a plan-local
  integer anchor. Input comparator anchors must resolve through the plan's
  extension declarations to a scalar function in the configured extension
  collection. Earlier output could omit or misassign comparison
  declarations; affected plans must be repaired or regenerated. Invalid
  comparator arity and non-boolean returns now fail during construction.
• spark: Spark 3.4 conversion rejects partition predicates on
  differently cased overlapping file and partition columns when Spark can
  prune using partition values that differ from its output values. Full
  reads and safe projections remain supported. Resolve the schema
  collision before converting affected filters.
• core: ExpandFieldType.CONSTANT_FIELD is renamed to
  CONSISTENT_FIELD, and building a SupportedRelation for an extension
  relation without message types now throws IllegalArgumentException.
• isthmus: Plain CREATE TABLE AS SELECT now emits ERROR_IF_EXISTS
  and CREATE TABLE IF NOT EXISTS emits IGNORE_IF_EXISTS instead of
  REPLACE_IF_EXISTS. Older Isthmus readers that accept only
  replacement-mode CTAS cannot read these plans. CREATE OR REPLACE TABLE
  IF NOT EXISTS is now rejected. Callers requiring replacement must
  request it explicitly.
• spark: Spark write conversion now rejects partitioned or
  bucketed V1 file/Hive commands, non-append filesystem save modes,
  incoming file UPDATE, MODIFIED_RECORDS output, and legacy filesystem
  INSERT plans with non-append create modes. Execute unsupported writes
  directly in Spark. Replay legacy APPEND_IF_EXISTS payloads only when
  their targets are known to be unpartitioned and unbucketed.

Features

• core: bump substrait to v0.103.0 (#1351) (2a8c9b0)
• core: bump substrait to v0.103.1 (#1387) (81a9dee), closes #1307

Bug Fixes

• core: correct SUM and SUM0 helper results (#1262) (8d91988)
• core: make Util.memoize thread-safe (#1406) (9042fea), closes #1389
• core: preserve custom comparison function identity (#1267) (c1e3767)
• core: preserve extended aggregate sorts and options (#1264) (fc6588c)
• core: retain reference scope during dereference (#1268) (ca8315b)
• isthmus: honor decimal arithmetic overflow options (#1393) (c9be8a4)
• isthmus: honor floating-point arithmetic options (#1394) (73232c3)
• isthmus: honor signed integer arithmetic options (#1392) (bc9ad04)
• isthmus: preserve CTAS creation policies (#1266) (d05cd6b)
• isthmus: preserve supported string function options (#1390) (0d3f1a1)
• isthmus: preserve UPDATE predicates (#1259) (eb8d3f7)
• isthmus: preserve update targets in nested schemas (#1265) (c130b20)
• isthmus: quote dynamic function identifiers when needed (#1403) (5277678)
• isthmus: remove the lossy contains mapping (#1391) (aeba8e0)
• isthmus: restore bidirectional sqrt mapping (#1396) (2df13ea)
• isthmus: validate unary arithmetic option preferences (#1395) (c3def2c)
• preserve mandatory read and post-join filters (#1260) (b7fcde5)
• spark: preserve aggregate outputs in projections (#1263) (2ce688f)
• spark: preserve partition values in file scans (#1270) (c51fd20)
• spark: reject lossy write conversions (#1269) (72e902e)