dalgo2sql

dalgo2sql

SQL adapter for DALgo - a Database Abstraction Layer in Go.

Our approach to development

We build with our own tooling:

• SpecScore — specify requirements as SpecScore.md artifacts
• SpecStudio — author & manage specs across their lifecycle
• inGitDB — store structured data in Git where applicable
• DALgo — data access layer for Go
• cover100.dev — drive toward 100% test coverage
• DataTug — query & explore data

Status

Usage

go get github.com/dal-go/dalgo2sql

SQLite structured aggregation

With DbOptions.StructuredQueryDialect set to sqlite, structured DALgo queries render GROUP BY, COUNT, SUM, AVG, MIN, MAX, DISTINCT aggregates, HAVING, alias rewrites, result ordering and pagination natively. SUM/AVG normalize numeric inputs to REAL for parity with DALgo's generic float64 fallback. FIRST/LAST are deliberately not advertised until DALgo models aggregate-local ordering; using unspecified SQLite row order would not be deterministic. DALgo's planner therefore refuses a query that uses them for this adapter (see the PostgreSQL section), and nothing runs it.

Exact SQLite decimal SQL functions

Applications that issue SQLite SQL directly can opt in to exact decimal arithmetic on the modernc.org/sqlite driver. Register the functions before opening the first connection:

if err := dalgo2sql.RegisterSQLiteDecimalFunctions(); err != nil {
	return err
}
db, err := sql.Open("sqlite", "app.sqlite")

Store decimal input in a TEXT-affinity column such as DECIMAL_TEXT(38, 6). The functions accept decimal strings, bytes, and exact integers; they reject floats, exponent notation, malformed input, and inputs or results over 38 significant digits or scale 38. Arithmetic returns canonical decimal text. Addition and multiplication never round: an over-precision result is an error. Division, decimal_round, and decimal_avg use round-half-to-even, with scales from 0 through 38.

SELECT decimal_add('0.1', '0.2');             -- '0.3'
SELECT decimal_mul(quantity, unit_price);     -- exact decimal text
SELECT decimal_div('1', '6', 4);              -- '0.1667'
SELECT decimal_sum(amount), decimal_avg(amount, 2)
FROM invoice_lines;
SELECT amount FROM invoice_lines ORDER BY amount COLLATE DECIMAL;

Scalar functions return NULL if any argument is NULL; decimal_cmp otherwise returns -1, 0, or 1. Aggregates ignore NULL and return NULL when every input is NULL or no input rows exist. The DECIMAL collation orders valid decimal text numerically; malformed text is placed after valid numbers and ordered lexically. Constrain a decimal column to valid decimal text when relying on numeric ordering. These opt-in functions do not change ordinary SQLite NUMERIC or DALgo aggregate normalization.

Which path compiles a structured query

A structured DALgo query reaches SQL by one of four paths. They are tried in this order, and which one runs is decided by how the source was opened, not by the query.

1. A native compiler (DbOptions.NativeStructuredQueryCompiler). When one is set it writes the whole statement, and StructuredQueryDialect is not consulted. It is for trusted adapter dialects that bring their own compiler.
2. SQLite (DbOptions.StructuredQueryDialect: "sqlite") uses the SQLite structured emitter, compileStructuredSQL, described above. It is built around SQLite's dynamic typing and has its own path; the typed compiler does not touch it.
3. PostgreSQL, MySQL, Oracle and SQL Server (DbOptions.StructuredQueryDialect: "postgres", "mysql", "oracle" or "sqlserver") use the typed SQL compiler, not the legacy emitter. They are described in the next sections.
4. No dialect, no native compiler uses the legacy text emitter, described after that. A PostgreSQL, MySQL, Oracle or SQL Server source should not stay on it; open it with its dialect.

Any other non-empty dialect is refused with an error when a structured query is read.

PostgreSQL

With StructuredQueryDialect: "postgres" a structured read runs in two steps on one connection: one catalog query that reads every column of each source the query names (their types, NOT NULL flags and collations), then the one SELECT the typed compiler writes from that answer. compileTypedSQL(query, dialect, facts) is the compiler and newPostgresDialect(mode) is PostgreSQL's spelling; the exact SQL and arguments of every supported shape are in testdata/postgres.

• No value is written into the statement. Every constant is a bound argument and every name goes through the dialect's quoting (" doubled, NUL and names over 63 bytes refused). Pass Go integers for whole numbers: a whole number that arrives as a float64 binds as numeric, which is correct but cannot use an index on an integer column. A constant that does not match its column (a number against text) is a server error, not an empty result. A float32 or float64 compared with a column the catalog types as real (float4), in a comparison or in an IN list, is bound as ?::real from its decimal text, so the comparison runs in the column's own type and a stored 0.1 is found by 0.1; as a numeric it is not the same number as the float4 nearest to 0.1, and an equality finds nothing. double precision and numeric columns keep the numeric binding, and so does a float a real cannot hold (past about 3.4e38 or below the smallest normal real, about 1.2e-38), which the server would refuse as out of range, a column whose type the catalog gives by the name of a domain over real, and a float that is not compared with a column (arithmetic, HAVING on an aggregate).
• Identifier case is DbOptions.IdentifierCase: IdentifierCaseExact (the default) writes names as the query spells them, for databases created with quoted mixed-case names; IdentifierCaseFoldLower writes them lower-cased, for databases created by dalgo2postgres. The result keeps the names the query asked for in both modes (Total comes back as Total, an unaliased COUNT(*) as COUNT(*)). A select list whose names differ only in case is refused under IdentifierCaseFoldLower. Do not put a field mask or access check on such a mount unless the names it compares are folded first. A configured primary key is matched to the result's column by the same fold, so a key configured as ID keys the records of a column the server returns as id. A query that runs in DALgo's own engine instead of the compiler (a subquery, a join the adapter declines) does not fold: under IdentifierCaseFoldLower it must spell table and column names as the catalog stores them, in lower case, until DALgo's engine folds names itself. The compiled path accepts any case.
• A table that is not there fails with *TableNotFoundError (it matches ErrTableNotFound), which names the table and the nearest name that exists (SuggestedSchema and SuggestedName; both empty when none is near): table "album" not found; did you mean "Album"? Table names are case-sensitive. Under IdentifierCaseFoldLower only a name the mount can write is suggested. A sequence, a name that resolves to nothing and a relation with no readable column are all "not found"; a source is never compiled without its catalog facts.
• Native aggregation and joins. The adapter reports GROUP BY, HAVING, ORDER BY, COUNT, SUM and AVG with their DISTINCT forms, MIN and MAX as native (QueryCapabilities), so DALgo runs them on the server. A query with a subquery stays in DALgo's generic engine. FIRST and LAST are refused by DALgo's planner for this adapter: they need a provider-declared stable input order, which the adapter does not declare (neither dialect promises a group-key order or a stable row order), so a query that uses one fails with the planner's error and runs nowhere, in the generic engine included. A join is accepted (CanExecuteJoin) when each ON pair has the same type category (numbers with numbers, text with text) or the same type, on the database handle as in a transaction; it is declined when the types differ or a key's type has no usable equality (json, xml, geometric types, oid and the reg* types), when two text keys have different collations of their own (PostgreSQL cannot choose between them and refuses the comparison, SQLSTATE 42P22; a collation of its own against the database's default, or the same collation, is accepted), when the catalog does not hold a key, or when the compiler cannot write the query. JoinFields lists a table's columns from the catalog, under the names the catalog has, whatever the identifier case.
• Arithmetic is on double precision: +, -, * and / read both operands as double precision, as DALgo's generic engine does, so an integer overflow cannot differ between engines; division by zero is NULL. SUM and AVG are cast to double precision. NULLs sort first ascending and last descending, as in DALgo, and the NULLS clause is left out for a NOT NULL column.
• A query the compiler cannot run faithfully in one statement is refused with an error matching dal.ErrNotSupported, among them a result column whose alias, or whose expression text, is over 63 bytes. DALgo falls back to its generic engine only where it has a fallback: a join, which it asks the adapter to accept first (CanExecuteJoin compiles the whole query, and DALgo runs the generic join when it declines), and a query with a subquery. For a query over one source the refusal is the read's error: a 64-byte alias fails the read, and the remedy is a shorter alias. It is never handed to the legacy emitter.
• A join keys each record by the base row (the first source of FROM), as DALgo's generic join does, whichever columns the select list names. The records reader adds the base's configured primary key to the statement as a column of its own, qualified by the base's alias, and reads the key from it by position, so a column of a joined table that carries the key's name never keys a record. A select-all over joins, which no column can be added to, is keyed from the base's own column of that name (the catalog's list of the base's columns, a statement of its own on SQLite); a key the base has no column for leaves the records with the placeholder key, as a select-all over one source does. The same holds for the SQLite dialect. A select list that gives two expressions of a join one output name (a.id and r.id) is refused with join_field: duplicate output name, the error DALgo's generic engine gives for the same query, with the SQLite dialect too.
• A record is keyed by the position of its key column, never by a name two columns share. Where the reader knows the column it takes that position: the catalog's key of a select-all (its position among the source's columns, which a select-all returns in table order), and the first item of a select list that is the key's own field of the base source (a field of a joined source that has the key's name is not). The catalog is asked in a statement before the one that reads the rows, so a result that does not hold the key's column at that position, because the source's columns changed between the two, fails the read (the columns of the source changed during the read) instead of keying the records from another column. Where only a name says which column is the key (a declared key on a select-all, a select-all over joins) the column stored under the name the statement writes is taken (on a fold-lower mount, the key's name folded), else a column that spells the name as it was declared, else the first column that is the name folded. A select list over one source that gives one output name to two different expressions (a and b AS a) is refused before any statement, as it is for a join, with every dialect and compiler (columns[1]: duplicate output name "a"); one expression written twice is one column asked twice and is left alone. The names a wildcard lists are not known before the statement, so they are not compared.
• Records are keyed by the source's primary key. The records reader keys each record by the primary key column of the recordset declared for the source in DbOptions.Recordsets (else DbOptions.PrimaryKey). With no key configured, the key is the source's primary key as the catalog reports it, when that is exactly one column. The one catalog query that serves the compiler serves the key, so a read still sends one. A source with no primary key, with a composite one, a view, a materialized view and a foreign table (the catalog reports none for these), and a grouped or aggregated query, key their rows by ordinal: the position of the row in the result, from 0, as decimal text. A recordset that is declared with no single-column primary key keeps the literal ID __dalgo_record_id, as does a read with the SQLite dialect, the legacy emitter or a native compiler of the caller's and no key configured: they have no catalog to ask. On a fold-lower mount a declared recordset is looked up by the query's spelling of the source, then by that spelling folded as the catalog lookup folds it, so a recordset declared as album is found by a query that says Album or ALBUM.
• A stream error is never the end of a result. Both readers return the error the server raised in the middle of a result (a timeout, an overflow at one row) from Next, instead of ErrNoMoreRecords; a read whose context ended returns the context's error, wherever in the read it ended (errors.Is(err, context.Canceled) or context.DeadlineExceeded), also when the driver reports its connection broken (driver.ErrBadConn) or database/sql reports it closed (sql.ErrConnDone) because of it; the connection goes back to the pool in every case.
• The TugQL money option is refused (ErrNotSupported) whatever the select list: the records reader reads the option from the caller's query before it adds the key column to it.
• No protected-write factory. NewDatabase returns the plain adapter for PostgreSQL; the protected read/write profile exists for SQLite only.

The catalog query and the statement need the same connection, so the database handle takes one connection for the life of a structured read and gives it back when the reader is closed, when it is read to its end, or when the context of the read ends, as the pool does for a read of its own.

MySQL, Oracle and SQL Server

StructuredQueryDialect: "mysql" (MySQL 8.0.17 and later, verified against 8.4), "oracle" (Oracle Database 23ai, verified against Oracle Database Free 23) and "sqlserver" (SQL Server 2022, through go-mssqldb) run a structured read the way the PostgreSQL dialect does: one catalog query on a pinned connection, then the statement the typed compiler writes from its answer, every value bound. IdentifierCase must be empty or exact. What differs from PostgreSQL (the exact SQL of every shape is in testdata/mysql and testdata/oracle and testdata/sqlserver, and dialect_mysql.go, dialect_oracle.go and dialect_sqlserver.go document the rest):

• Names are written exactly as the query spells them, in backticks (MySQL) or double quotes (Oracle; SQL Server, which reads them so with QUOTED_IDENTIFIER ON, as go-mssqldb's login sets it). An Oracle table created without quotes is stored in upper case and must be spelt so; a not-found error suggests the name that differs only in case. MySQL refuses a name over 64 characters, outside the Basic Multilingual Plane or ending with a space; Oracle one over 128 bytes or holding a double quote; SQL Server one over 128 UTF-16 code units. Every field is written qualified with its source (`Album`.`Title`), because MySQL compares an ORDER BY name with the select-list outputs without regard to case and Oracle 23ai reads select aliases in GROUP BY and HAVING; a qualified name is always the column. Oracle writes a table alias without AS.
• Placeholders are ? (MySQL), :1, :2, ... (Oracle, as go-ora binds them) and @p1, @p2, ... (SQL Server), whatever DbOptions.Placeholder says. LIMIT and OFFSET are bound (LIMIT ? OFFSET ?; OFFSET :1 ROWS FETCH NEXT :2 ROWS ONLY; OFFSET @p1 ROWS FETCH NEXT @p2 ROWS ONLY, after ORDER BY (SELECT NULL) when the query orders by nothing, as SQL Server requires).
• NULL order: MySQL and SQL Server already sort NULL first ascending and last descending, as DALgo does, so no NULLS clause is written; Oracle writes NULLS FIRST/NULLS LAST unless the catalog says the column is NOT NULL.
• Arithmetic and aggregates are on DOUBLE (MySQL), BINARY_DOUBLE (Oracle) and FLOAT (SQL Server), with a zero divisor read as NULL. SUM and AVG cast their argument on MySQL and SQL Server (MySQL's AVG of an INT is a DECIMAL rounded to four places, SQL Server's is an int) and their result on Oracle.
• Result values: go-ora delivers every Oracle NUMBER as text; the readers turn it into an int64 when the column's scale is 0 or unconstrained (COUNT included) and the value is whole, else into a float64 (decimal text with ExactNumericValues). MySQL and SQL Server deliver integers as int64 and DECIMAL as text, as before.
• Refused (dal.ErrNotSupported), as with PostgreSQL: FIRST and LAST, window functions, subqueries, cursors, the money option, RIGHT and FULL joins; a NaN or infinite float constant (none of the three engines has one); on Oracle and SQL Server a GROUP BY expression that carries a constant (neither has GROUP BY by position); on Oracle an empty string or byte constant, which Oracle reads as NULL, so != "" and NOT IN ("") would find no row. SQL Server refuses a statement that binds over 2100 values, and its COUNT fails past 2^31 - 1 rows.
• Text follows the engine's collation. MySQL (default utf8mb4_0900_ai_ci, which ignores case and accents) and SQL Server (default SQL_Latin1_General_CP1_CI_AS, which ignores case) compare text by the column's collation, so =, IN, ORDER BY, GROUP BY, DISTINCT, MIN and MAX treat 'a' and 'A' alike, where DALgo's generic engine compares bytes. Oracle's default is binary.
• Catalog: information_schema.COLUMNS for the named schema or DATABASE(); ALL_TAB_COLUMNS for the named schema or the session's CURRENT_SCHEMA; sys.columns of the relation OBJECT_ID resolves from the statement's own name (default schema, then dbo). A synonym (Oracle, SQL Server) and a SQL Server #temp table are not seen, so a read of one is refused as table-not-found. A MySQL TEMPORARY table is not seen either: alone it is not found, but one that shadows a base table of the same name is read by the statement while the facts describe the base table (its columns, NOT NULL and primary key), so do not read through a connection that holds one.

Key reads and writes (Get, Set, Insert, ...) do not use these dialects: their names must be plain identifiers, as with no dialect.

Legacy text path

A source opened with no StructuredQueryDialect and no NativeStructuredQueryCompiler renders structured queries through the legacy text emitter. A source leaves this path by setting a dialect: sqlite for SQLite, postgres for PostgreSQL, mysql for MySQL, oracle for Oracle and sqlserver for SQL Server. A source opened without one stays on the legacy path until its consumer sets it.

The emitter pastes names and values straight into the statement, so it fails closed. What it refuses, by where it is found (each refusal is an error wrapping dal.ErrNotSupported, and no SQL is executed):

• A name (collection, schema, alias, field, column alias, function name) that is not a plain identifier.
• A string constant, anywhere (a plain value, an IN list, inside a slice), that holds a backslash, a bracket or a control character.
• A string inside a slice passed as one constant (dal.Constant holding a []string or []any, which encoding/json renders) that also holds a double quote, <, >, &, U+2028, U+2029 or invalid UTF-8, which json writes as an escape or replaces with U+FFFD. A plain value and an IN list may hold these characters.
• A constant of a kind whose text cannot be checked: a []byte (or any byte slice, nested ones included), a named slice type, a non-finite number (NaN, Infinity, -Infinity, float32 included), a time outside the range JSON can write, a time inside an IN list, and any Go type that is not a scalar or an unnamed slice of scalars.
• An aggregate other than COUNT, SUM, AVG, MIN and MAX, and a keys-only query that would be ordered by a field that is a reserved word (such as a primary key named order), because the emitter writes the name unquoted and the server rejects it. The reserved words are the union across engines (PostgreSQL, MySQL, SQL Server and SQLite), so a word that one of them accepts bare, such as index or user on PostgreSQL, is refused too: the text must run on each.

NUMERIC result values

Since dalgo2sql v0.21.0 the readers turn the text a driver delivers for a column typed NUMERIC (pgx delivers PostgreSQL's NUMERIC as text) into a float64 when it is decimal text, or exactly NaN, Infinity or -Infinity. Text that is not a number stays a string, and a float64-typed recordset column refuses it with an error. The records reader converts every such column; the recordset reader converts a float64-typed column, and with lib/pq, which delivers NUMERIC as []byte, it keeps []byte.

Compared with dalgo2sql v0.20, SQLite results are unchanged in the recordset reader. In the records reader a bare NUMERIC column changes in two cases: a BLOB holding decimal text, and the texts NaN, Infinity and -Infinity, become float64. A DECIMAL column is not NUMERIC and keeps its text.

Callers that must preserve a database decimal can opt into DbOptions.ExactNumericValues. With that option, NUMERIC values remain decimal text in record, recordset, and map-backed Get/GetMulti reads; they are never normalized through float64. Key reads into struct data are refused because a struct field can force a lossy conversion. A driver that already returned a float32 or float64 is refused because its original decimal cannot be recovered. The zero value keeps the legacy behavior. PostgreSQL's NaN, Infinity and -Infinity spellings remain text so a downstream provider can accept or reject them without inventing a numeric value.

PostgreSQL BYTEA normally follows the legacy records-reader behavior and is converted from []byte to string. Callers that need binary fidelity can set DbOptions.PreserveBinaryValues; in map-backed records reads this keeps only columns whose database type is BYTEA as copied []byte values. Text columns, other database types, and the zero-value behavior are unchanged. Map-backed Get and GetMulti reads honor the same option. Recordset reads already keep byte-slice columns as bytes.

SQLite native result values

SQLite's affinity is not its storage class: an INTEGER column can hold a text or a BLOB cell, a TEXT column can hold a BLOB, and a NUMERIC column can hold an integer, a real, text, a BLOB or NULL. The readers normally apply the affinity and text normalization described above, so a value can come back as another class (or, in a recordset, be truncated to the type inferred from the first row).

Callers that need the stored value itself can set DbOptions.PreserveNativeValues. It is supported only with StructuredQueryDialect: "sqlite"; another dialect, an empty one included, is refused before any statement, so the option never claims native semantics of an engine this package has not reviewed. It cannot be combined with ExactNumericValues, whose decimal text is not a native value; that pair is refused before any statement too. PreserveBinaryValues is a PostgreSQL option and has no effect on SQLite, whose BLOBs are already native []byte.

With the option on:

• Map-backed records reads (Get, GetMulti, and the records reader) store the driver's own int64, float64, string, []byte or nil (NULL). A []byte is copied before the driver can reuse its buffer. A read into struct data is refused before any statement, as ExactNumericValues refuses it, because a field type could force the conversion the option exists to avoid.
• Recordset reads build every column as recordset.Column[any]: ValueType() is interface{} (never nil), Name() and DbType() are the result's own, and a NULL cell is nil. One column can hold any of the five storage classes in different rows of the same result; the reader does not infer the column type from the first row.
• The compiled SQL, the declared column metadata, and the keys and IDKind of read records are unchanged: a key is still derived from the same value as before. The zero value of the option keeps the legacy behavior exactly.

The option is a read option. It does not change writes, Exists, or the statements of a structured query.

NULL in a recordset

The recordset readers hold each result column in a column that marks its NULL cells. A NULL is nil from the column's GetValue and Values() and from GetValueByIndex, GetValueByName and Data() of a row, for every column type the readers build (text, integers, floats, booleans, times, bytes, the sql.Null* types and the columns a SQLite view gives no type). It used to be the zero value of the column's Go type (0, false, an empty string, the zero time), which a cell can also hold, so a consumer could not tell a NULL from a stored zero; the records reader has always returned nil. A consumer that relied on the zero value now gets nil and tests for it.

A column keeps its Go type: ValueType() and DbType() answer as before, so a consumer that aligns numbers by ValueType() still does, and DefaultValue() is still the zero value. IsBitmap() is false for a boolean column, whose values are held in a slice. A value that does not fit its column is an error of Next that names the column and the Go types, never the cell, where it was a panic. The integer columns of sql.NullInt16 and sql.NullInt32 take the int64 a driver delivers when the column holds it, and the 0 and 1 of a SQLite BOOLEAN column are false and true. SQLite takes the type of a column from the first row of the result: a result whose first row holds a NULL has no column types, and its columns are text columns, as before.

Reading a record: the target and the key

A read (Get, GetMulti) fills a pointer to a struct with exported fields, or a map with string keys whose elements hold any value (a map by value must be made; a pointer to a nil map is filled). Every other target is an error wrapping dal.ErrNotSupported, returned before any statement: a nil record, a record with no data, a struct given by value, a nil pointer, a struct with a field that is not exported, a nil map, a map whose keys are not strings or whose elements cannot hold every column value, data that is neither a struct nor a map (a scalar, a slice, a pointer to either, a pointer to a pointer, a pointer to an interface), and a batch that mixes map and struct data. Every operation that takes a key returns an error wrapping dal.ErrNotSupported for a nil key.

End2end - is a separate module

For end-to-end testing a SQLite driver is used. To avoid bringing a dependency to SQLite into the consumers of dalgo2sql, the end2end tests are in a separate module.

This is an unusual approach, as usually you would want to bring dependency to underlying driver with a dalgo adapter. But this is not a case for this adapter as database/sql that is referenced by dalgo2sql is an abstraction layer and consumer is free to choose the underlying driver.

Reading rows into structs

Both the records reader (queries that read into a record) and single-record Get accept a pointer to a struct as the record data. A column reaches a field by its db tag, or by the Go field name without a tag; db:"-" and unexported fields are never filled. Fields promoted from embedded structs count, and a nil embedded struct pointer is allocated when a column needs it (not when its type is unexported, which is an error).

The fields are listed breadth-first, as scany does: every exported field is reachable by its own name, including promoted fields that share a Go name but carry different tags (two embedded structs that each have an ID tagged user_id and account_id, or an outer ID tagged id beside an embedded ID tagged legacy_id). An embedded struct tagged db:"-" is skipped with all its fields. A column is matched to a field by exact name first, then without regard to case, then without regard to case and underscores, so AreaSqKm, areasqkm and area_sq_km all reach a field AreaSqKm. Where two fields spell the same name, the shallower one wins, and at one depth the first in declaration order. Two columns of one row that reach the same field are an error, never a silent overwrite.

An exact spelling beats depth, as it beats a looser spelling: with an untagged ID in the outer struct and an embedded field tagged db:"id", the column id reaches the tagged embedded field, not the outer one (scany gave it to the outer ID, whose snake_case name is also id); the column ID still reaches the outer field.

A struct field that is named (not embedded) is one column, named after its db tag or its Go name, whatever its type. scany's db:"" on a named struct field, which maps the nested struct's fields without a prefix, is not supported: those fields are not reached, and a column that names one of them has no field.

One difference in precedence from scany: the exact spelling is tried first, so with an untagged Name declared before a field tagged db:"name", the column name reaches the tagged field (scany gave it to Name, whose snake_case name is also name); the column Name still reaches the untagged one.

Records reader. A column without a field is skipped (the identity column of a record is not a field of its data). NULL is nil in a pointer, an interface or a []byte field, calls Scan(nil) on a sql.Scanner, and is an error naming the column for any other field, as it is for Get (it used to store the zero value). Values follow what a map target gets, with these differences by field type:

• a sql.Scanner receives the driver's value unchanged, so a decimal type sees the exact NUMERIC text and a JSON type sees []byte;
• a string field takes the driver's text, including the exact NUMERIC text;
• a []byte or json.RawMessage field takes a copy of the driver's bytes;
• an integer field takes the integer the driver's text spells, so a NUMERIC such as 9007199254740993 (what PostgreSQL returns for SUM of a bigint) keeps every digit and the whole int64 and uint64 range is reachable; a whole NUMERIC with a scale (9007199254740993.00, what SUM over numeric(p,2) returns) is read from the part before the point, so it keeps every digit too; a fraction is an error, judged by the text (so one beyond 2^53 is one too, never the float64's rounding), and so is a whole number outside the field's range (-9223372036854775809 for an int64 is an error, never the nearest bound);
• float and bool fields take the normalised value (NUMERIC becomes float64); time.Time and any fields take it as is ([]byte becomes string for an any field, as in a map). Every field is checked for overflow.

Get. Only the column matcher differs from the scany library this path used before; values are still scanned by database/sql into the field addresses, so its conversions (an integer cell into a string field, []byte into json.RawMessage or a sql.Scanner, NULL being an error for a non-pointer field) are unchanged. scany stays in charge of a sql.Scanner target with one column and of a row that has a dotted column name no field matches: scany accepts dotted names for nested struct fields (address.city), and for such a row it matches every other column by its own spelling too. Any other column without a field is an error that names the column and the struct type (column "typo": no corresponding field in T). What now differs: names match without regard to case and underscores (scany matches db tags exactly and Go names as snake_case), so a SQLite column Name reaches the field Name; column names scany rejected can match; two columns that reach one field are an error; and an exact name wins over a looser one (see the precedence note above).

Keys-only query order

A keys-only query (SelectKeysOnly) with no ORDER BY is ordered ascending by the primary key (from DbOptions.Recordsets, else DbOptions.PrimaryKey, else, with the PostgreSQL dialect, the primary key the catalog reports), so the order is defined and repeatable on each database. Text keys sort by that database's collation (the SQLite compiler uses BINARY; the legacy emitter, PostgreSQL and MySQL use the column's own collation). A query that names an order keeps it. Two cases keep the statement they always had: a query with joins (the unqualified key is refused by the SQLite compiler and ambiguous elsewhere), and, on the legacy text emitter, a primary key that is not a plain identifier.

License

Free to use and open source under MIT License.