Skip to content

v1.9.7

Choose a tag to compare

@teaguesterling teaguesterling released this 03 Oct 14:37
· 36 commits to main since this release
4721b31

Preparation for DuckDB 2.0, plus the bugfixes that work surfaced. Seven merged PRs; no new functions and no signature changes.

DuckDB v1.5.6

The stable line pins DuckDB v1.5.6 (069cc9f9b5), up from v1.5.5 — all 14 live references, the extension-ci-tools reusable-workflow pin, artifact names and the CLI download URL (#86).

DuckDB 2.0 readiness

TableFunction was restructured on v2.0-cyanoptera: SimpleNamedParameterFunction was deleted, taking the named_parameters member with it, and the public arguments member went too. Both are ported behind feature probes in duckdb_compat.hpp, leaving the v1.5 path byte-identical (#87, #90).

One detail worth recording for anyone doing the same port. The first arguments shim compiled, linked, and then failed at extension load on every table function:

INTERNAL Error: CompatSetArgumentTypes cannot shrink a signature from 2 to 1 parameters

On 2.0 the named parameters live in the same parameters vector as the positional ones, because WithTypedKwargs appends one VAR_KEYWORD entry. One path argument plus any number of named options therefore reports GetParameterCount() == 2, and setting one positional type reads as a shrink. Use GetPositionalParameterCount(), which counts only STANDARD/POSITIONAL_ONLY and never VAR_KEYWORD — the same set v1.5's arguments = replaced.

Next-line legs are green on linux_amd64, wasm_eh and windows_amd64.

Replacement scan registration (#95, issue #82)

FROM 'file.md' raised Binder Error: No extension found that is capable of reading the file, while read_markdown('file.md') worked. The scan was registered in MarkdownExtension::Load, which the loadable extension's entry point never calls — DUCKDB_CPP_EXTENSION_ENTRY calls LoadInternal only — so the shipped artifact had every table function and no replacement scan. Moved into LoadInternal, which both load paths reach.

Not a 2.0 regression, though the issue assumed one: the entry macro expands identically on v1.5.6 and v2.0-cyanoptera, and the failure reproduces on a stock v1.5.6 CLI. Static linking hid it in local builds and in the CI stable legs, since those run Load(); only a loadable load exercises the entry point.

Verified against a stock v1.5.6 CLI with the built .duckdb_extension loaded — pre-fix both plain and decorated bare-FROM raise the binder error, post-fix both return rows, with an explicit read_markdown() call as the control.

deflist respects the declared encoding (#89, issue #77)

deflist was the one block type that reinterpreted its content as Pandoc JSON regardless of the declared encoding. A block saying encoding='text' whose text happened to parse as [[[term],[defs]],...] was replaced by the extracted text instead of emitted as written. It is now gated on encoding='json', like blockquote, list and table.

The gate matches the producer's contract rather than being our invention — duck_block_utils checks the same condition before treating a deflist as Pandoc — and our reader never emits deflist, since cmark does not parse definition lists.

This changes output if you relied on text-encoded deflist content being reinterpreted. The encoding='json' cases are unaffected, and are pinned as controls so a regression that stopped interpreting deflist altogether would be caught rather than mistaken for the fix.

CI correctness

  • "Test WASM in Browser" reported success while running no browser test, for every DuckDB pin since v1.4.3. The compatibility resolve was a step inside the browser job, so when no published duckdb-wasm matched, the testing steps skipped and the job passed; a step-level skip does not appear in the checks list. Split into a separate job so the browser job is reported as skipped (#91, issue #84). Two shell-injection paths in that workflow closed at the same time.
  • The submodule drift guard exempted every uses: line, including the one that is a DuckDB-line version (duckdb/extension-ci-tools@vX.Y.Z). A bump could update duckdb_version: and ci_tools_version:, miss that pin, and pass every assertion. Exemption narrowed, and verified by mutation rather than a green run: with the pin alone reverted, the patched guard fails naming the line and the old guard passes (#92, issue #88).

Also

Issue #78 closed — the vendored duck_block header's formatter markers landed upstream above #pragma once, so they arrive on the next re-vendor with no local divergence.