Comprehensive SQLite reference for implementing an MSSQL-compatible server on top of SQLite. Covers SQL syntax, type affinity and STRICT tables, DDL (CREATE/ALTER/DROP TABLE/INDEX/VIEW/TRIGGER, FKs, generated columns), DML (INSERT/UPDATE/DELETE/UPSERT/REPLACE, ON CONFLICT, RETURNING), SELECT (joins, CTEs, window functions, set ops, row values, EXPLAIN), built-in scalar/aggregate/date/math/printf functions, PRAGMAs (journal_mode, foreign_keys, busy_timeout, defer_foreign_keys, etc.), transactions and SAVEPOINTs, WAL mode and locking semantics, indexes and the query planner (partial/expression indexes, ANALYZE, EXPLAIN QUERY PLAN), JSON1/FTS5/RTree/virtual-table extensions, and limits/quirks that diverge from T-SQL. Use when designing the SQL-translation layer, schema bootstrap, transaction model, or compatibility behaviors of the server.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Comprehensive SQLite reference for implementing an MSSQL-compatible server on top of SQLite. Covers SQL syntax, type affinity and STRICT tables, DDL (CREATE/ALTER/DROP TABLE/INDEX/VIEW/TRIGGER, FKs, generated columns), DML (INSERT/UPDATE/DELETE/UPSERT/REPLACE, ON CONFLICT, RETURNING), SELECT (joins, CTEs, window functions, set ops, row values, EXPLAIN), built-in scalar/aggregate/date/math/printf functions, PRAGMAs (journal_mode, foreign_keys, busy_timeout, defer_foreign_keys, etc.), transactions and SAVEPOINTs, WAL mode and locking semantics, indexes and the query planner (partial/expression indexes, ANALYZE, EXPLAIN QUERY PLAN), JSON1/FTS5/RTree/virtual-table extensions, and limits/quirks that diverge from T-SQL. Use when designing the SQL-translation layer, schema bootstrap, transaction model, or compatibility behaviors of the server.
SQLite Reference
Complete reference for SQLite 3.53.x, scoped to building a SQLite-backed MSSQL-compatible server. Synthesized from the official SQLite documentation snapshot (sqlite-doc-3530100).
syntax.md — Lexical structure (identifiers, all four quoting styles incl. [brackets], reserved keywords), literals (string, blob X'...', numeric, NULL, TRUE/FALSE, CURRENT_*), bind parameters (?/?NNN/:name/@name/$name), operators and precedence, expression forms (CAST, COLLATE, CASE, IN/EXISTS, subqueries, COALESCE/NULLIF/IFNULL/IIF), three-valued NULL logic, name resolution and T-SQL three/four-part name mapping
data-types.md — Storage classes (NULL/INTEGER/REAL/TEXT/BLOB), the 5-rule affinity-determination algorithm, insert/comparison/arithmetic coercion, INTEGER PRIMARY KEY rowid alias quirks, STRICT tables (ANY type, SQLITE_CONSTRAINT_DATATYPE), date/time storage conventions, BOOLEAN representation, REAL precision (binary64), collations (BINARY/NOCASE/RTRIM), NULL semantics incl. UNIQUE divergence from MSSQL, T-SQL→SQLite type map
ddl.md — CREATE TABLE (column-def, constraints, WITHOUT ROWID, STRICT, IF NOT EXISTS, CTAS), AUTOINCREMENT vs rowid alias, ALTER TABLE supported ops + 12-step rebuild workaround, DROP, CREATE INDEX (UNIQUE/partial/expression/COLLATE/DESC), CREATE VIEW + INSTEAD OF triggers, CREATE TRIGGER (BEFORE/AFTER/INSTEAD OF, NEW/OLD, RAISE), CREATE VIRTUAL TABLE, foreign keys (enable via PRAGMA, actions, RESTRICT vs NO ACTION timing, DEFERRABLE), sqlite_schema introspection, ON CONFLICT actions (ROLLBACK/ABORT/FAIL/IGNORE/REPLACE)
dml.md — INSERT (VALUES multi-row, SELECT, DEFAULT VALUES, OR <action>), UPDATE (row-value SET, multi-table via FROM with PostgreSQL semantics — contrasted with T-SQL), DELETE (truncate optimization), UPSERT (ON CONFLICT (target) DO NOTHING/DO UPDATE SET excluded.col), REPLACE (incl. trigger/FK side effects), RETURNING (scope, ordering, restrictions), INDEXED BY / NOT INDEXED, conflict-action precedence
queries.md — SELECT grammar, DISTINCT/ALL, table-star, JOIN types (CROSS/INNER/LEFT/RIGHT/FULL/NATURAL/USING/ON, RIGHT+FULL since 3.39), GROUP BY / HAVING with SQLite's bare-column relaxation, ORDER BY (NULLS FIRST/LAST, COLLATE precedence), LIMIT/OFFSET both forms, compound SELECTs (UNION/INTERSECT/EXCEPT), subqueries, row values, VALUES as row source, CTEs (recursive, MATERIALIZED/NOT MATERIALIZED), window functions (full frame-spec incl. ROWS/RANGE/GROUPS, EXCLUDE, named windows), EXPLAIN vs EXPLAIN QUERY PLAN, table-valued functions
functions.md — Built-in scalar functions (abs, coalesce, concat, format/printf, glob, hex/unhex, iif, instr, length, lower/upper, quote, random/randomblob, replace, round, substr, trim/ltrim/rtrim, typeof, unicode, zeroblob, etc.), date/time (date/time/datetime/julianday/unixepoch/strftime/timediff + all modifiers), math (built-in default in node:sqlite), aggregates (count/sum/total/avg/min/max/group_concat/string_agg with FILTER/DISTINCT/ORDER BY), printf format specifiers, T-SQL gap callouts
pragmas.md — Concurrency/durability (journal_mode=WAL, synchronous, wal_autocheckpoint, busy_timeout, locking_mode), FK/triggers/constraints (foreign_keys, defer_foreign_keys, recursive_triggers, trusted_schema), memory/caching (cache_size, page_size, mmap_size, temp_store), schema introspection (table_list/table_info/index_list/foreign_key_list, pragma_* table-valued functions for information_schema emulation), planner (optimize, analysis_limit, automatic_index), debug/safety, plus connection bootstrap recipe for an MSSQL-compat server
transactions.md — BEGIN [DEFERRED|IMMEDIATE|EXCLUSIVE], COMMIT/ROLLBACK, SAVEPOINT/RELEASE/ROLLBACK TO nesting, autocommit, isolation model (snapshot per-connection in WAL, serializable in rollback mode), locking-state lifecycle (UNLOCKED→SHARED→RESERVED→PENDING→EXCLUSIVE), WAL reader/writer concurrency, checkpoint modes (PASSIVE/FULL/RESTART/TRUNCATE), SQLITE_BUSY/SQLITE_BUSY_SNAPSHOT handling, deferred FK behavior, server-design recommendations
COLLATE NOCASE on char/text columns for MSSQL's case-insensitive
default collation — comparisons and UNIQUE constraints follow the
column collation; LIKE is ASCII-case-insensitive anyway.
Application-enforced character widths because SQLite TEXT affinity does
not enforce varchar lengths. Transpiled explicit conversions call a
truncating character UDF; DML/default/computed and bulk storage call a
rejecting UDF before binding. Both use Windows-1252 bytes for non-Unicode
families and UTF-16LE code units for Unicode families.
Ordinary typed columns for IDENTITY storage — allocation belongs to the
engine's database registry, not SQLite rowid/AUTOINCREMENT. UDF expressions
reserve generated values during INSERT/MERGE evaluation so failed statements
leave gaps; TRUNCATE resets engine state transactionally.
Native functions: concat/concat_ws (NULL-as-empty matches
T-SQL CONCAT), iif, math functions (ceiling, power, ln,
log10, atan2, …), window functions, group_concat(x, sep) for
STRING_AGG, char, unicode, instr, substr, partial indexes for
filtered indexes, UPDATE ... FROM for T-SQL update-joins.
@name bind parameters so T-SQL variables pass through unrenamed.
Application-defined functions (mssqlite_*) for everything else —
date arithmetic with MSSQL boundary semantics, PATINDEX, dynamic +.
Expressions in DEFAULT clauses — DEFAULT (strftime(...)) for
DEFAULT GETDATE().
Dotted object names inside quoted identifiers ("sys.tables",
"app.users") to flatten MSSQL schemas without query interception,
and the temp schema for #temp tables.
Each SQL database owns a primary DatabaseSync and independent catalog.
Every primary attaches the other stores under encoded aliases, so
cross-database DML shares one SQLite transaction. Cross-file crash atomicity
requires disk-backed main with rollback journaling; it is not guaranteed
for in-memory main or WAL. SQLite's ten-attachment limit caps mssqlite at
eleven logical databases. Cross-database DDL switches to the target primary
and is rejected during user transactions.
What's NOT covered
The C API (sqlite3_*) — use node-sqlite for the JS-facing surface.
File-format internals, recovery, atomic-commit byte-level details, the source code organization.
Tcl bindings, the CLI shell, build/compile instructions.
Historical version notes and release logs.
Differential backlog
Native SQLite shortcuts that still leak through the T-SQL compatibility
boundary are tracked in TODO.md. The current SQL Server
2025 differentials cover text comparison/collation and result metadata; the
OPENJSON path, scalar JSON, and typed SELECT INTO contracts from that audit are
implemented.
Table variables use unique tables in the SQLite temp schema. Their names
stay out of the catalog; the engine keeps the T-SQL column definitions for
result metadata and drops each backing table at batch/procedure scope exit.
Current divergence: their DML shares the surrounding SQLite transaction,
while SQL Server table-variable updates survive user transaction rollback.
Built-in T-SQL table-valued functions lean on JSON1: STRING_SPLIT feeds
an adapter-produced JSON array to json_each, while OPENJSON feeds
source-spanned default-schema rows or exact WITH-row slices to json_each.
Column adapter UDFs apply SQL Server BIN2-like lax/strict paths before SQLite
casts the declared output types. Node's bundled SQLite does not expose the optional
generate_series virtual table, so GENERATE_SERIES renders as a
streaming recursive CTE with SQL Server's direction-sensitive default step.
Do not use JSON1 scalar extraction for JSON_VALUE/JSON_QUERY compatibility:
json_extract returns SQLite numeric types and normalizes object/array text.
mssqlite instead parses source-spanned nodes, returning lexical scalar text
or the exact selected fragment while enforcing SQL Server lax/strict errors
and JSON_VALUE's 4000 UTF-16-unit ceiling. OPENJSON shares that reader for
root and WITH-column paths while retaining json_each as its row source.
SQLite has no general LATERAL keyword, but eponymous virtual-table function
arguments and scalar subqueries can reference earlier FROM sources.
STRING_SPLIT/OPENJSON/GENERATE_SERIES APPLY use the former. Derived/VALUES
APPLY uses the latter to aggregate a typed, tagged JSON rowset under a
100,001-row input cap, rejects counts above 100,000, and expands through
json_each; pack/unpack UDFs retain BLOB and bigint cells losslessly.
T-SQL VALUES table sources wrap native SQLite VALUES in an outer SELECT that
renames generated columnN fields. Engine resolution first infers SQL Server
common types and validates the row/alias shape; transpilation explicitly
coerces each cell before rendering. This avoids SQLite's compound-SELECT term
limit and works in FROM, joins, CTE queries, uncorrelated APPLY, and MERGE.
SQLite has no PIVOT/UNPIVOT operators. mssqlite renders PIVOT as grouped
aggregates over CASE WHEN pivot_key = value THEN aggregate_value END.
UNPIVOT becomes a MATERIALIZED CTE plus one NULL-filtered UNION ALL
term per source column, ensuring volatile source expressions run once.
SQLite's dynamic affinity can carry mixed numeric values in direct
rendering, but engine execution rejects differing known declared types to
retain SQL Server's stable result metadata.
SQLite NUMERIC affinity and native numeric aggregates are not used for
DECIMAL/NUMERIC: columns have TEXT affinity and contain canonical fixed-scale
strings. Scalar and aggregate UDFs perform scaled-BigInt operations; a
sortable decimal key UDF replaces lexical TEXT ordering.
Opaque special types keep identity outside SQLite affinity: xml is Unicode
TEXT, CLR hierarchyid/geometry/geography values are native serialized
BLOBs, and sql_variant is a BLOB containing its TDS base-type envelope.
Target-column DML injects cast UDFs before binding, so a number or string
cannot be affinity-coerced into an indistinguishable value. Variant casts
unpack the envelope; special-type operators and methods are rejected before
SQLite sees them.
SQLite has no offset-preserving temporal storage class. mssqlite stores
datetimeoffset(n) as canonical fixed-scale TEXT containing local civil
time and ±HH:MM, and uses mssqlite_datetimeoffset_key (UTC day plus
100ns ticks) for comparison, IN/BETWEEN, ORDER BY, UNIQUE constraints, and
explicit expression indexes. Inserts, updates, defaults, and typed variables
are passed through mssqlite_datetimeoffset_cast so affinity cannot discard
the offset or fractional precision. Do not use SQLite datetime() or
julianday() for this type: they normalize/format with lower precision and
cannot retain the original offset.
SQLite date() and strftime() may normalize impossible civil dates and
therefore are not conversion validators. Ordinary T-SQL date/time casts and
declared storage use mssqlite_temporal_cast / mssqlite_implicit_temporal,
which round-trip year/month/day through proleptic civil-day math, validate
time ranges, raise 241, and return NULL for TRY conversions. FROMPARTS
constructors use separate NULL-propagating UDFs and raise 289 for invalid
component ranges.
T-SQL computed columns lower to GENERATED ALWAYS AS: VIRTUAL for ordinary
definitions and STORED for PERSISTED. CREATE supports both; SQLite permits
ALTER TABLE ADD only for VIRTUAL generated columns. Expressions may reference
columns in the same row and deterministic scalar functions, but not
subqueries, aggregates, window functions, or nondeterministic callbacks.
Exact decimal and checked integer generated expressions use separately
registered deterministic UDF names so SQLite accepts them in schema DDL.
SQLite cannot alter a column declaration in place. ALTER COLUMN captures the
physical CREATE TABLE plus dependent indexes/triggers/views, creates a
replacement table with the changed member, inserts through the target T-SQL
storage cast, swaps names, recreates dependencies, updates sys.columns, and
runs foreign_key_check inside one savepoint. Incoming foreign keys require
temporarily disabling enforcement outside the savepoint; the operation is
rejected inside a user transaction when that pragma transition is required.
Node's node:sqlite API cannot register custom SQLite collations. mssqlite
therefore declares BINARY/NOCASE as a baseline and renders supported SQL
Server collations through a deterministic normalization-key UDF. Predicates,
IN/BETWEEN, joins, ORDER/GROUP BY, DISTINCT aggregates/projections,
UNION/EXCEPT/INTERSECT, expression indexes, and supplemental UNIQUE indexes
all use the identical effective-or-default key, preserving CI/CS, AI/AS and
BIN2 behavior. Accent-insensitive keys use Unicode NFD with combining marks
removed; only trailing U+0020 spaces are ignored as SQL comparison padding.
The original value is projected from grouped/set wrappers, so normalization
does not leak into client rows. LIKE remains on its dedicated matcher. Native
SQLite foreign keys cannot call a UDF, so text-bearing foreign keys omit the
native clause and install persistent triggers for child validation and parent
NO ACTION/CASCADE/SET NULL/SET DEFAULT behavior over the same key.
SQLite UNIQUE indexes treat NULLs as distinct, unlike SQL Server. mssqlite
expands each nullable logical key k to (k IS NULL), ifnull(k, 0) in a
unique expression index. The flag separates real values from the arbitrary
sentinel, while k may itself be a collation, exact-decimal, or temporal key
expression. CREATE TABLE constraints keep their native definition plus a
reserved supplemental index (2627); explicit unique indexes use the expanded
key directly (2601). Partial-index WHERE predicates remain unchanged.
SQLite has no ROLLUP, CUBE, GROUPING SETS, or GROUPING function. mssqlite
expands them into one ordinary GROUP BY query per grouping set joined by
UNION ALL, replacing omitted keys with NULL and GROUPING calls with branch
constants. A MATERIALIZED CTE evaluates simple sources once; duplicate
explicit sets remain separate branches as SQL Server requires.
FOR JSON uses SQLite JSON1 construction rather than string concatenation:
per-row json_object/json_patch preserves escaping and optional NULLs,
json_group_array produces the outer array, and json() marks JSON_QUERY
or nested FOR JSON fragments after subquery boundaries. Empty PATH/AUTO
results coalesce to []; WITHOUT_ARRAY_WRAPPER uses comma concatenation
and, like SQL Server, is only valid JSON for a single row.
Persisted scalar user functions are not SQLite schema functions: the engine
registers varargs callbacks by final name and interprets the stored T-SQL
body in an isolated scope. Inline TVFs substitute argument expressions into
their stored SELECT AST and become derived sources; correlated simple forms
lower to ordinary equality joins because SQLite lacks general LATERAL.
SQL Server DML triggers are not mapped to SQLite CREATE TRIGGER: SQLite
fires once per row and exposes only scalar OLD/NEW values. The engine wraps
each triggering statement in a savepoint, captures full affected rowsets
with RETURNING plus an UPDATE pre-image snapshot, materializes temp
inserted/deleted tables, and interprets the stored T-SQL body once.
INSTEAD OF operations build intended images without mutating the base table.
Rowversion is the deliberate internal exception to that rule. Its physical
BLOB column remains nullable so an AFTER INSERT trigger can replace the
transient NULL; an AFTER UPDATE trigger detects unchanged OLD/NEW versions
for MERGE, cascades, and other indirect writes. Normal engine DML supplies
the value directly so SQLite RETURNING sees it. Both triggers call a
nondeterministic UDF backed by a synchronous server-wide unsigned counter;
persisted decimal-text state flushes only outside user transactions, so a
SQLite rollback does not reuse an allocated version.
SQLite has no schema sequence generator. mssqlite persists definitions and
allocation state as catalog rows, advances a synchronous database-wide BigInt
registry from a nondeterministic UDF, and flushes it only outside SQLite user
transactions so SQL Server-style rollback-independent consumption survives.
SQLite returns NULL for division/modulo by zero and promotes overflowing
integer scalar arithmetic. mssqlite therefore renders checked arithmetic UDFs
that evaluate operands once and raise SQL Server 8134/8115 (or NULL only under
ARITHABORT OFF + ANSI_WARNINGS OFF). Custom SUM and integer AVG aggregates
check int/bigint accumulator widths instead of waiting for SQLite's
signed-64-bit overflow; AVG divides exact sum/count state toward zero and
supplies an inverse callback for window use rather than SQLite's REAL avg().
SQLite compares mixed storage classes by its own class ordering and numeric
affinity may convert invalid text to zero. mssqlite therefore infers the
higher-precedence T-SQL type before rendering predicates, CASE/IN/BETWEEN,
compound SELECTs, and VALUES, then calls strict conversion UDFs. Target DML
and MERGE assignments use the same category conversions before storage.
SQLite LIKE lacks T-SQL bracket classes and Unicode collation behavior.
mssqlite therefore compiles every LIKE pattern in a UDF after applying the
effective SQL collation, including ranges, negated classes, ESCAPE, and the
trailing-source-space alternative required by SQL Server.
SQLite length/substr and JavaScript iteration use code points where SQL
Server's non-SC collations use UTF-16 units. Dedicated UDFs use JS code-unit
indexing. Lone-surrogate results cross SQLite as raw UTF-16LE BLOBs because
SQLite TEXT replaces them, then the engine decodes them after metadata is
known so TDS retains the exact two-byte unit.
Current engine deviation from the bootstrap recipe above: single shared
connection per server with plain BEGIN (sync API, single process) —
revisit WAL + IMMEDIATE if a multi-connection engine lands.
