Repository navigation
Turso v0.8.1 parity: corpus refresh, 0.8 planner and bindings, MVCC and spill fixes - #75
Merged
Marc-André Moreau (mamoreau-devolutions) merged 27 commits intoOct 1, 2026
Conversation
Refresh conformance/sqlite-sqltests from the pinned turso-src v0.8.1 (8549c1659): 33 new files and 41 refreshed files since v0.8.0-pre.7. The 113 newly exposed case-level differences are recorded in the expected- failures ledger pending closure; the inventory's live count follows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A BEGIN CONCURRENT transaction could overwrite a peer's update committed after its snapshot, or resurrect a row the peer deleted: base-row deletion markers are separate tombstone versions, invisible to the older snapshot, and the no-visible-version path only checked in-flight peers. Treat any begin/end stamp committed after the writer's snapshot as a write-write conflict (Turso 0f7f30eac). Concurrent inserts also promoted every rowid through the store-global allocator, silently rewriting explicit rowids and INTEGER PRIMARY KEY values after a peer delete. Only rowids the statement allocated itself are promoted now; an explicit key is kept and a clash with a concurrent writer's live version is a write-write conflict. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rences
Port the Turso v0.8.1 behavior the refreshed sqltest corpus exposed:
- json_group_object/jsonb_group_object skip rows whose label is SQL NULL
(517ec809f "json: skip NULL labels in object aggregates").
- jsonb(NULL) returns SQL NULL (5775d5337 "json: preserve SQL NULL in jsonb()").
- json_each/json_tree bind a selectable rowid (qualified, unqualified, and in
INSERT ... SELECT), numbered from zero like SQLite and core/json/vtab.rs;
generate_series reports the generated value as its rowid (core/series.rs)
(16a02b139 "core/translate: bind rowid on virtual tables").
- Date/time functions read a BLOB value or modifier as UTF-8 text
(b2512eb71 "core/functions: read a BLOB date/time argument as text").
- unixepoch() and strftime('%s') floor to whole seconds, so 1 ms before the
epoch is -1 (535f684f1 "core/functions: fix datetime rounding").
- Scalar min() keeps the last of tied arguments while max() keeps the first,
matching SQLite's minmaxFunc (78b1ac633, merged as d81560556).
- generate_series yields no rows when any supplied bound is NULL.
- The simple-count INDEXED BY exemption covers COUNT(col) over a provably
non-NULL column, as Turso's detect_simple_aggregate does.
The integrity_check/memory.sqltest attached-corrupt-database case stays in the
ledger as an intentional managed-ATTACH policy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nts, coerce FTS queries Error-code fidelity (Microsoft.Data.Sqlite parity): - EmbeddedSqlException.SqliteErrorCode now yields the (extended) SQLite result code for every classic failure class: explicit where the raise site knows it (primary-key/rowid conflicts 1555/2579, RAISE 1811, datatype mismatch 20, sequence exhaustion 13) and otherwise derived from SQLite's own error text (UNIQUE 2067, NOT NULL 1299, CHECK 275, FOREIGN KEY 787, STRICT 3091, readonly 8, busy 5, full 13, too big 18, corrupt 11, ...). Wrappers that re-raise the same failure keep the wrapped code. - The SQLite facade reports the primary code in SqliteErrorCode and the "SQLite Error N" text, and the extended code in SqliteExtendedErrorCode, on the embedded, remote (Hrana error.code names such as SQLITE_CONSTRAINT_UNIQUE) and SqliteRemoteException paths. - A garbage plain header (bad magic, invalid page size, too few usable bytes, too-short file) opened without a key or codec reports SQLITE_NOTADB (26) "file is not a database" like SQLite and upstream 6480cd77e; keyed/codec opens keep the encrypted-or-not-a-database phrase. - SqliteErrorCodeDifferentialTests pins the codes against e_sqlite3 for in-memory, file-backed and MVCC databases. Explicit checkpoints (upstream a95d01284 / 7ad5d9de4): PRAGMA wal_checkpoint is refused with SQLITE_BUSY "cannot checkpoint while another statement is active - SQL statements in progress" while another statement on the same connection is active or suspended; open blob handles do not block it. FTS (upstream d485aa6d8): integer, real and blob fts_match/fts_score queries are searched by their text form on the indexed and scalar paths instead of raising "requires a text query". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stats, sync status, token provider Close the remaining gaps against turso-src v0.8.1 bindings/dotnet: - Connection strings (AhtolaConnectionStringBuilder and the SQLite facade, so EF Core UseAhtola reaches them): Sync Client Name, Sync Long Poll Timeout, Bootstrap If Empty, Partial Bootstrap Prefix/Query, Partial Sync Segment Size/Prefetch, Remote Encryption Cipher/Key, Push Operations Threshold, Pull Bytes Threshold, Force Logical MVCC Pull (only False is supported; True fails closed), Sync Experimental Features (validated, no effect) and the Ahtola-specific Automatic Sync Mode. They map onto AhtolaReplicaOptions and are rejected on local/direct-remote connections, matching Turso's HasAdvancedReplicaOptions. - Explicit sync operations on AhtolaConnection and SqliteConnection: Pull/PullAsync, Push/PushAsync (drains every batch pending at the call), Checkpoint/CheckpointAsync and GetSyncStatistics/GetSyncStatisticsAsync, reusing the managed replica engine. Pull-only and push-only requests coalesce per kind in ManagedReplicaSyncRegistry. Sync results now report real main/revert WAL sizes; statistics report pending journal changes. - Automatic sync: AutomaticSyncStatus and AutomaticSyncStatusChanged (ported from TursoAutomaticSyncCoordinator) make background failures observable. AhtolaAutomaticSyncMode.PullOnly matches Turso's pull-only loop; the default stays PushAndPull. - AuthTokenProvider hook (connection, facade, replica options) resolved per HTTP request, per Hrana WebSocket connect, and per replica sync request. - Hrana replication_index is treated as opaque per PROTOCOL.md 7.2.6: the numeric watermark is still tracked and echoed, other values are ignored instead of failing the batch. - Sync client ids may now carry a validated client-name prefix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uilds PRAGMA auto_vacuum (Turso v0.8.1 translate/pragma.rs without the experimental autovacuum feature, which Ahtola does not have): - The read form reports the mode page 1 declares (largest-root-page and incremental-vacuum header fields), so an auto-vacuum database SQLite created reports FULL (1) / INCREMENTAL (2) instead of a hard-coded 0. - The set form resolves the mode like upstream's requested_mode: names in any case, or a numeric literal valued 0..2 (so 00 and 0x0 are NONE). NONE stays accepted; FULL, INCREMENTAL and unrecognized values keep upstream's --experimental-autovacuum diagnostic. ALTER COLUMN in MVCC mode (upstream validate_indexes_can_be_rewritten): a rewrite that changes stored values (becoming generated, toggling VIRTUAL, an affinity change) or may change a VIRTUAL generated value, and that touches an index observing the column or a generated column derived from it, is rejected with 'cannot ALTER COLUMN "x": rebuilding affected indexes is not supported in MVCC mode'. The two alter_column *-mvcc corpus cases now fail only on the CLI "Parse error:" prefix, like the other prefix-only entries; their ledger summaries are updated accordingly. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ports the Turso v0.8.1 (turso-src 8549c1659) fixes behind 19 ledger entries recorded by the v0.8.1 corpus adoption, plus non-corpus parity: - REPLACE defaults now land before virtual generated columns are recomputed (Turso ea7b14a0a), so CHECK, UNIQUE keys, NOT NULL, STRICT typing and RETURNING see the substituted value; generated-column NOT NULL is enforced after the default pass. - Invalid generated-column clauses report SQLite's messages (be3c173b7): a second generation clause, a clause after DEFAULT, or an unknown type token fail with `error in generated column "<col>"`. - UPSERT honors the INSERT target alias (859c91ade): the alias names the target row, hides the base name, and shadows `excluded`. - UPSERT DO UPDATE keeps columns a BEFORE UPDATE trigger rewrote and recomputes generated columns (a99d061a5). - User INSERT/UPDATE/DELETE on the __turso_internal_ namespace fails with `table X may not be modified` regardless of case (67262f427); managed backup/snapshot replay is exempt like upstream nested statements. - ALTER TABLE RENAME onto a view reports `there is already another table or index with this name` (799e4737c). - PRAGMA temp.synchronous reads OFF and ignores writes, like SQLite. - CREATE TABLE / ADD COLUMN enforce SQLITE_MAX_COLUMN 2000 with SQLite's messages (a81efd85b). - Named parameters tokenize with SQLite's CC_VARALPHA rule: bare `$`, `:`, `@` (or `$::`, unclosed `$a(b`) are `unrecognized token`, and TCL names `$::v`, `$ns::v`, `$arr(e)` are single parameters (22a0a0f6e, a45cd87ff). - PRAGMA count_changes uses SQLite's column names, reports zero-row statements, leaves RETURNING rows alone, and counts only rows an UPSERT inserted. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # src/Ahtola.Tests/Conformance/managed-sqltest-expected-failures.txt
…grity_check
Turso "report STRICT type violations from integrity_check" parity:
- PRAGMA integrity_check/quick_check now checks every non-rowid-alias column
of a STRICT table against its declared INT/INTEGER/REAL/TEXT/BLOB storage
class ("non-TEXT value in t.b"), before its NOT NULL check, using SQLite's
OP_IsType masks (NULL always passes; REAL accepts the integer storage class
SQLite writes for integral reals). ANY and domain/custom types are exempt.
- Stored rows' VIRTUAL generated columns are recomputed on read with affinity
only, like a SQLite/Turso column read, instead of enforcing NOT NULL and the
STRICT storage class at load. A row that violates them stays readable and
integrity_check reports it ("NULL value in t.b", "non-INT value in t.b")
rather than the database failing to open.
- The sqltest harness now builds the four database/integrity_strict_*.db
fixtures like upstream's generate_strict_fixture (managed write path plus a
same-length first-row serial-type swap), so the parity_strict_* corpus files
run instead of being skipped, and all pass.
Ledger: removed the two parity_gencol_not_null_violation entries, which now
pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Turso 74eb80c83 / SQLite walTryBeginRead: a reader whose WAL is fully backfilled normally takes read-mark 0 and ignores the WAL, but a checkpoint holds read-mark 0 exclusively for as long as it runs. Instead of retrying until the checkpoint finishes, SqliteWalReadSnapshotCoordinator now pins the committed boundary with an existing mark already at mxFrame or an idle mark it advances to mxFrame. Every frame is already in the database file, so the mark only keeps a later checkpoint from restarting the WAL underneath the reader; the usual confirmation (mark value, boundary, commit frame) still applies. An empty WAL has no commit frame to pin and still waits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er cases Correctness: - Parser: ON/USING join constraints on comma joins and CROSS JOIN (`a, b ON ...`, `a CROSS JOIN b USING (...)`, NATURAL CROSS JOIN), matching SQLite's joinop/on_using grammar. JoinTableSource carries a Cross flag; CROSS JOIN is a join-order barrier (Turso JoinInfo::no_reorder) and never flips its hash-build orientation, so its left operand stays the outer loop. - FULL OUTER JOIN with correlated subqueries referencing the joined tables is no longer rejected: Turso v0.8.1 unnests these after the FULL JOIN, and the managed evaluator already evaluates them per joined row, null-padded rows included. - The transient equality hash probe no longer prunes rows a SQLite scan would still evaluate: an equality is only probeable while no conjunct before it can raise. A positive single-table EXISTS keeps pruning, matching SQLite's existsToJoin automatic-index semantics. Fixes NOT EXISTS hiding 'malformed JSON' errors. Planner/EQP (from the executed plan): - Partial-index implication accepts either disjunct of an OR predicate (Turso query_term_implies_predicate), and a covering partial index can be scanned without a key constraint. - A correlated IN the semi-join rewrite declines is described as its real per-row CORRELATED LIST SUBQUERY scan. - FORMAT=JSON reports the ported Turso scan estimate for an unfiltered table scan (sqlite_stat1 row count or the 1,000,000-row default). Ledger: 17 v0.8.1 entries closed; 5 now-parsing join/memory entries record their new plan-pattern differences. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # src/Ahtola.Tests/Conformance/managed-sqltest-expected-failures.txt
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e Turso cost model
Port Turso v0.8.1 f7aac1fc5 ("optimizer: allow unnesting correlated
subqueries with non-equality comparisons"): a correlated EXISTS, NOT
EXISTS or positive IN may now move any direct inner/outer comparison
(=, <>, <, <=, >, >=, IS, IS NOT) into the semi/anti join condition
(unnest.rs is_inner_outer_comparison). The fallible/non-deterministic
inner-WHERE gates are unchanged.
The inner access of a semi/anti join is now chosen by a port of the
v0.8.1 join planner (choose_best_btree_candidate, the temporary-index
alternative of find_best_access_method_for_btree, and the LeftAnti hash
join of join_lhs_and_rhs / try_hash_join_access_method with its
should_not_use_hash_join and probe-index gates), costed with exact ports
of cost.rs / estimate_hash_join_cost (TursoCostModel). Execution and
EXPLAIN QUERY PLAN both consume the same SemiAntiInnerAccess:
- declared index searches (equality prefix plus one range),
- automatic (ephemeral) index searches,
- a new left anti hash join that hashes the outer rows, probes the inner
table once and emits unmatched build rows in order.
The compiled automatic-index semi-join lowering declines when the cost
model prefers another access, so EQP and execution stay one plan.
Closes unnest-correlated.sqltest correlated-comparison-uses-joins,
correlated-in-inequality-uses-indexed-semi-join,
correlated-in-with-only-inequality-uses-indexed-semi-join and
not-exists-can-use-hash-anti-join.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…R, and semi/anti JSON estimates - Port Turso v0.8.1 in_filters_implied_by_or_branches (70f75beb7, lift_common_subexpressions.rs): an OR whose every branch pins a column to copyable literals implies column IN (values); when several columns of one table qualify only the sole leading-index column is used, and none when a compound index starts with two of them. The planner consumes the implied filter while choosing an access path; the statement itself is unchanged, so the original OR still filters every returned row. - TryPlanManagedIndexScan now searches an index by an explicit positive IN list or an implied IN filter on its plain ascending leading column (choose_best_in_seek_candidate), reported as Turso's (col=?) seek. - The OR-by-union planner replans compound AND branches with the compound-seek analysis and may use one index in several branches (MULTI-INDEX OR t (txy, txy)), matching consider_multi_index_union. - IndexCoversSelect understands IN lists and BETWEEN. - EXPLAIN QUERY PLAN FORMAT=JSON attaches the ported cost model's estimates (input/output rows, access and running cost, including cost_with_where_work with Turso's extra_steps) to planned semi/anti join nodes; search, multi-index and hash-join ops can now carry estimates. Closes multi_index_or_compound.sqltest compound-or-keeps-correlated-index-columns-together, nested-or-branches-use-one-index-search, nested-and-branches-use-one-index-search and turso/explain-query-plan-json.sqltest exists-is-reported-as-a-semi-join, not-exists-is-reported-as-an-anti-join. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PRAGMA page_count now reports the real size for every database, and a
connection's PRAGMA max_page_count ceiling is enforced with SQLite's
SQLITE_FULL ("database or disk is full", result code 13) and its
transaction semantics.
- :memory: databases keep rows as heap objects, so their page count is
measured by serializing the catalog through the same page builders a
persisted image (VACUUM INTO / full rewrite) uses, via a private
in-memory page planner store. Per-tree counts are cached by row-store
lineage/revision; page_count is a high-water mark with freelist, and
VACUUM compacts it. Values match bundled SQLite for the tested shapes.
- File-backed databases: SqlitePager refuses a write transaction whose
target size exceeds both the ceiling and the committed size (Turso
Pager::allocate_page), before any frame is written. Explicit
transactions plan the persist (pager PlanCommitsOnly probe) after each
statement so the statement fails like SQLite's, not the COMMIT.
- max_page_count is per connection, never persisted, clamps to the
connection-visible page count, and treats 0/negative as a query.
- In an explicit transaction a single-row write without a statement
journal rolls back the whole transaction (sqlite3VdbeHalt); any other
statement only rolls back itself.
- MVCC: autocommit writes (which build pages at commit) and checkpoints
are held to the ceiling; MVCC transaction COMMITs are not (Turso).
- Removes the sql-error-message-prefix database-full ledger entry.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # src/Ahtola.Core/EmbeddedDatabase.cs # src/Ahtola.Tests/Conformance/managed-sqltest-expected-failures.txt
…he executed join tree Port the per-step costing of optimizer/join.rs join_lhs_and_rhs and access_method.rs (find_best_access_method_for_btree, the temporary-index fallback, try_hash_join_access_method, estimate_hash_join_cost, cost_with_where_work) into JoinOrderEnumerator.EvaluateTursoStep: declared index seeks, INTEGER PRIMARY KEY lookups, the ephemeral index over a full scan, left-deep hash joins that build the last placed table (materialized behind a longer prefix up to MAX_MATERIALIZED_BUILD_ROWS), outer-join Poisson output estimates and the cross-product penalty. Two-table LEFT/FULL base-table joins are planned the same way, and VdbeHashJoinRuntime now keeps unmatched rows for either build side so LEFT/FULL can hash their preserved input. EXPLAIN QUERY PLAN for a compiled join now walks the OpenJoinCursor plan it will execute and renders Turso's EqpDetail rows (SCAN / SEARCH ... USING [COVERING] INDEX (a=? AND b=?) / SEARCH ... USING INTEGER PRIMARY KEY / HASH JOIN / MATERIALIZE hash build input / LEFT-JOIN / USE SORTER FOR ORDER BY). Shapes Turso would not produce for a table with INDEXED BY or NOT INDEXED (temporary index, hash join) keep the previous placeholder description. Unit tests that pinned the old single-row EQP, the automatic-index choice or the old build-side heuristic now assert the plan the Turso cost model selects. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nd seek join indexes on constant key prefixes FORMAT=JSON nodes now carry Turso v0.8.1's plan_estimate figures (optimizer/mod.rs) wherever the described access path is one the ported cost model prices exactly: - filtered full-table scans (estimate_scan_cost + cost_with_where_work, with constraint_output_multipliers and the closed-range factor), declining for BETWEEN/LIKE/IN, subqueries, OR-implied IN filters, or a constraint Turso could seek instead; - single-table index searches and ordered index scans (estimate_cost_for_scan_or_seek, with the is_index_ordered exemption from the non-covering full-scan penalty); - the evaluator's LEFT JOIN index plan (the null-supplying seek priced by the same inner-access planner as semi/anti joins, rows_after_join's outer-join Poisson term); - multi-index OR unions (multi_index.rs estimate_multi_index_scan_cost with rowid-only branches); - uncorrelated IN-subquery seeks (choose_best_in_seek_candidate, list rows capped at sqrt(rows)) and the correlated scalar subquery seek priced per outer row; - compound-select arms, FROM-clause coroutine subqueries (scan plus body cost), and recursive CTE scans (1,000,000-row fallback) with their one-row recursive input. Join planning: a `column = literal` term on a join member can now extend an index seek key (constraints.rs constant constraints) when the literal needs no affinity conversion and the column's collation matches the index; the compiled index seek uses the literal verbatim. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…us in AGENTS.md Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…oin costing, EQP estimates Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # src/Ahtola.Tests/Conformance/managed-sqltest-expected-failures.txt
Record what the v0.8.1 refresh delivered, the 33 remaining v0.8.1 differences (evaluator-route planner shapes and documented policies), what stays unported, and known follow-ups. The inventory's live expected-failure count is now 109. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A GROUP BY whose sort spills under the default 2 MB execution budget took 25-40 s at 10k rows: the spill codec issued two or three unbuffered file operations per value written or read. The sorter's spill file now coalesces writes in a buffer that absorbs the codec's reserved length headers, and each run reader reads ahead in blocks. Both buffers scale with the memory limit (none below 512 bytes, so tiny budgets keep their exact accounting) and are charged to the spill and merge infrastructure. The 10k-row GROUP BY drops from ~28 s to ~0.5 s; groupby/default.sqltest from ~15 minutes to ~80 s, so it no longer hits the 30 s per-case cap. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A 3,000 x 3,000 equi-join whose build side spills under the default 2 MB budget took ~32 s: - Resident build partitions filled the budget, so probe-batch admission (which keeps a quarter of the budget free) never grew a batch past one probe, and each probe reloaded its partition. Batch admission now evicts cached partitions to keep growing, and each batch answers groups whose partition is already resident first, so cyclic batches stop evicting a partition just before reusing it. - About half of the loads ran out of memory part-way and restarted after each eviction. Partitions now record their entries' retention as they are written; a load makes room for that lower bound first and skips a load that cannot fit. - Every partition read went through the spill codec one value at a time. Sequential partition scans now read through a transient read-ahead block charged to the statement's memory. The join drops to ~0.4 s (10,000 x 10,000: ~2.4 s); output keeps probe order. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ggregates Evaluator DISTINCT, UNION/INTERSECT/EXCEPT, DISTINCT aggregates and recursive-CTE UNION compared every row with every kept row, whatever the memory budget: count(DISTINCT pad) over 20k rows took ~16 s and a self-UNION ~42 s. They now share a DistinctRowSet that buckets rows by a hash consistent with DISTINCT equality (integer/real numerically, BINARY, NOCASE and RTRIM text by their built-in rules; any other collation hashes text as a constant, so it stays correct). RowsEqual still decides equality within a bucket; UNION keeps the later row of an equal group as the pinned corpus expects. The same queries now take ~0.1-0.25 s. Differential tests against bundled SQLite cover value classes, collations, compound operators, DISTINCT aggregates and recursive UNION. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Marc-André Moreau (mamoreau-devolutions)
merged commit Oct 1, 2026
2e78ba3
into
master
11 checks passed
Marc-André Moreau (mamoreau-devolutions)
deleted the
claude/ahtola-turso-parity-check-b1919f
branch
October 1, 2026 20:18
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.
Brings Ahtola up to the Turso v0.8.1 pin (
turso-src@8549c1659): vendors the v0.8.1 sqltest corpus, closes most of the differences it exposed, ports the 0.8 planner and .NET binding additions, and fixes the MVCC and performance bugs found along the way.Conformance
docs/turso-gap-inventory.jsonis reconciled.docs/turso-remaining-gap-plan.md("v0.8.1 refresh")::memory:journal_mode, CLIParse error:prefix, SQLite schema-text quoting, readonly VACUUM message, file-backed ATTACH, COVERING wording.Correctness
BEGIN CONCURRENTwriter could overwrite a peer's update committed after its snapshot, or resurrect a row the peer had deleted. This is now a write-write conflict (Turso0f7f30eac).CROSS JOIN … ON/USINGand a comma join followed byONnow parse.PRAGMA count_changesis implemented.jsonb(NULL),json_treerowid), date/time (BLOB arguments, pre-epoch rounding), min/max ties andgenerate_serieswith NULL arguments.SqliteExceptionnow carries SQLite primary and extended result codes, locally and remotely: 19/2067/1299/275/1555/787, 13, 26, 5, 8, …max_page_countis enforced on file-backed and:memory:databases, and:memory:gets exact page counts.integrity_checkreports STRICT and generated-column type violations.Turso 0.8 features
Pull,Push,Checkpoint,GetSyncStatistics.PullOnlymode. The default still pushes and pulls.replication_indexhandling.Performance
groupby/default.sqltest(whole file)count(DISTINCT …)Behaviour changes worth reviewing
InvalidDataException.SqliteErrorCodeis no longer always 1.max_page_countis now per-connection, as in SQLite.Not ported (product decisions)
Force Logical MVCC Pull=TrueValidation
SqlitePagerPortableLockCoordinatorTests.DeleteModePagerFaultsWhenPeerSwitchesMainFileToWal, was a cross-process worker hitting its 30 s exit timeout. That class passes 22/22 on rerun../build.ps1 validate-trimpassed on the bindings slice.🤖 Generated with Claude Code