Conversation
…irror of the build, with the AST readers taking the new parser's shapes (T-467) DuckDB 2.0.0 is scheduled for 2026-10-21. The v2.0-cyanoptera preview path at artifacts.duckdb.org redirects to an expiring Actions blob and advances with every nightly, and only some builds have extensions published for them, so the driver is pinned to alpha38195 — the build with httpfs, json, ducklake, aws and postgres for every platform — mirrored as the prerelease duckdb-v2.0.0-alpha38195 on this repository. What the preview changed and what took it: - A bare NULL select column arrives over Arrow as the null type, with no type, data or length, which adbc 0.12.1 cannot decode; Engine.Result reads it as nils, as many as the batch's other columns hold. - The PEG parser's json_serialize_sql: a CAST's target is an unbound TYPE node (cast_type.type_info.expr, or type_expr in later builds) carrying the name as written; a CONSTANT is a literal of kind and text in later builds; a FUNCTION's children are arguments; LIMIT n% is a LIMIT_MODIFIER with limit_type PERCENTAGE. Engine.Ast gains cast_type/1, constant/1 and arguments/1 reading both shapes, and the materialized gates, the pruner, SingleTable, and the decomposer read through them. - EXPLAIN renders box-drawn, mixed-case plans, and EXPLAIN ANALYZE over the C API answers "profiling is disabled" unless enable_profiling is set; the runner sets it with the job's other per-connection statements, ahead of the lockdown that seals the configuration. - epoch_ms of a value past the timestamp range is an INTERNAL error in this build, which invalidates the database instead of yielding NULL under TRY; the writer test for that case is tagged out (:duckdb_preview_internal_error) until the pin moves past alpha41396, where it is fixed. Catalogs this build creates are DuckLake 1.1-dev1, which the 1.5.3 extension refuses; a 1.0 catalog attaches unchanged and 1.5.3 reads it back, so an existing deployment keeps its rollback. The deployment doc carries the note. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBmgYoYciJzbcxNYA9Z4d6
…talog format apart from the engine version Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBmgYoYciJzbcxNYA9Z4d6
… connection, so job engines and partial workers read the lake on DuckDB 2.0 (T-467) The integration and kind-cluster jobs failed on every query with "TransactionContext Error: Failed to query snapshot ... Current transaction is aborted (please ROLLBACK)", on SQLite and Postgres metadata alike. The DuckLake extension built for the 2.0 preview runs three SETs on its metadata connection at every transaction — pg_experimental_filter_pushdown off for Postgres, sqlite_disable_multithreaded_scans on for SQLite, and current_transaction_invalidation_policy right after BEGIN. A locked configuration refuses them, and the refused SET after the BEGIN invalidates the transaction, so the snapshot query that follows fails. Bisected to `SET lock_configuration = true` alone, on the CLI as well as through ADBC; 1.5.3's extension predates the SETs. DuckDB's allowed_configs lists options exempt from the lock, so both lockdowns (Runner and PartialWorker) now set it to exactly those three before locking, through Catalog.DuckLake.allowed_configs_statement/0. None of the three reaches outside the engine, and the list locks with the rest. The two DuckLake query-path tests that read EXPLAIN ANALYZE enable profiling on their engine first, as the runner does for an analyze job. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBmgYoYciJzbcxNYA9Z4d6
… anywhere in a cast target is refused, an all-NULL batch is an error, the lockdowns share one ordered builder, and the overflow test is skipped by version
Findings from the review of the 2.0 preview branch, each fixed at its root:
- The decomposer accepted `LIMIT n%` — 2.0 folds it into LIMIT_MODIFIER with
limit_type PERCENTAGE — and rendered it as `LIMIT n`; gate_modifiers refuses
it as :percent_limit. A DECIMAL limit's serialized value is DuckDB's unscaled
integer, so `LIMIT 1.5` read as 15; Ast.integer/1 accepts only the integer
types, and SingleTable and the decomposer read limits through it.
- The zoned-type gate read only a cast's top-level type, so a TIMESTAMPTZ
nested in a STRUCT or list target passed; it now reads every TYPE node in a
cast target through Ast.type_name/1.
- A batch of nothing but untyped NULL columns carries no row count and read as
empty; Result.from_adbc/2 answers {:error, :unsized_batch}, from_adbc/1
raises, and a connection reports it as an error.
- Ast.cast_type/1 answered "UNBOUND" for an unbound cast without a TYPE node;
the TYPE heads guard their name and the id clause excludes UNBOUND. The
readers are exercised against the pinned parser in ast_test, so a pin that
moves to a spelling they do not read fails there.
- The Runner and PartialWorker lockdown lists were near-clones whose ordering
(allowed_configs before lock_configuration) held by position; both now call
QueryService.Lockdown.statements/3, and Identifier.sql_list/1 replaces four
copies of the list renderer. The runner's explain goes back to one query with
a prefix.
- The writer's overflow test is skipped by pinned version rather than excluded
by a static tag, so it runs again when the pin moves. The deployment note
says what the fault does in production, and names the driver mirror as the
trust root, with each asset's SHA-256 in the configuration doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NBmgYoYciJzbcxNYA9Z4d6
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Tracker: T-467 (groundwork for T-395).
Pins the adbc driver to DuckDB
2.0.0-alpha38195, thev2.0-cyanopterapreview build that has httpfs, json, ducklake, aws and postgres published for every platform (the newer alpha41396 has none), and adapts the engine to what 2.0 changed:NULLselect column now arrives as Arrow's null type with no type, data or length, which adbc 0.12.1 cannot decode.Engine.Resultreads it as nils, as many as the batch's other columns hold.Engine.Astgainscast_type/1,constant/1andarguments/1, which read both the 1.5 shape and 2.0's (unboundTYPEnodes for a cast's target,literal {kind, text}constants in later builds, named functionarguments,limit_type: PERCENTAGE). The materialized gates, the pruner,SingleTableand the decomposer read through them.EXPLAIN ANALYZEover the C API needsenable_profiling, which the runner now sets with the job's per-connection statements before the lockdown seals the configuration.SETs on its metadata connection at every transaction, whichlock_configurationrefused, invalidating the transaction and failing every query on a job engine or partial worker (SQLite and Postgres metadata alike). Both lockdowns now setallowed_configsto exactly those three options before locking; none reaches outside the engine and the list locks with the rest.epoch_mspast the timestamp range is an INTERNAL error in alpha38195 that invalidates the database instead of yielding NULL underTRY; alpha41396 has the fix. The writer test for that case carries:duckdb_preview_internal_errorand is excluded by default.Catalogs this build creates are DuckLake format
1.1-dev1, which the DuckLake extension built for DuckDB 1.5.3 refuses to attach; a format1.0catalog attaches unchanged under the preview and a 1.5.3 node reads it back afterwards, verified with SQLite metadata. Three versions are involved: the DuckDB engine, the DuckLake extension built for it, and the catalog format stamped in the metadata database.docs/deployment.mdcarries the upgrade note.The exact build is mirrored as the prerelease
duckdb-v2.0.0-alpha38195on this repo, since DuckDB's preview URL redirects to an expiring blob and moves nightly.CI: every job green, including the integration and kind cluster suites. Locally:
mix precommit(1847 tests),mix ciandmix dialyzergreen, and the previously failing integration files verified against a local Postgres 17.11.🤖 Generated with Claude Code
https://claude.ai/code/session_01NBmgYoYciJzbcxNYA9Z4d6