Upgrading to 9.26
9.26 is a bug-fix release with one new capability. Ten changes, eight of them fixes for schemas that Weasel was getting wrong — usually silently, which is what makes several of these worth reading even if nothing looks broken from where you are sitting.
Two behaviours change in ways you can observe
DbObjectName.Schema and .Name now hold the undelimited name, and whitespace inside a SQL Server function body is now significant. Both are described below. Emitted DDL is unchanged in every other respect.
The silent ones
These four produced no error. If any of them describes your schema, the database was not what Weasel reported it to be.
A delimited name meant the object was created once and never migrated again
#499. A schema or table name can reach the model already delimited — DbObjectName.Parse splits a qualified name and keeps the parts as written, and before 9.25 delimiting a name by hand was the only way to use one that needed it. Nothing normalized it afterwards, and introspection binds that name against catalogs that hold the bare one.
The delimited spelling never matched, so the object read as absent, and the delta was Create on every run — written as CREATE TABLE IF NOT EXISTS, which succeeds against the table that is already there and does nothing:
--- declared as "MixedCase".things ---
delta after adding a column = Create
added_later present in the database? NOEvery change after the first was discarded, with no error, and ApplyAllAsync returned normally. On SQL Server it surfaced loudly instead — There is already an object named 'x' in the database — because the CREATE SCHEMA guard missed the bracketed name the same way.
Names are now normalized on the way into the model, which is the change to DbObjectName.Schema/.Name flagged above. If you compare either property against a literal that includes quotes or brackets, that comparison changes meaning. Names that arrive bare are byte-identical, and so is all emitted DDL.
No index was read in any schema named pg…
#504. The index introspection query carried NOT nspname LIKE 'pg%' beside the nspname = :schema it was already filtered by. It could never have excluded anything — except a user schema starting with those two letters, which it excluded completely.
pgcontrol, pgqueues, pgdata: no index was read back at all. The primary key still arrived, so the table looked correct and only its non-PK indexes were missing. Every declared index was then reported Missing rather than Different, so the patch was a bare create index with no drop — 42P07 on the second startup, which aborted everything ordered after it. The schema never converged, so db-assert could not pass either.
An invalid index reported no drift
#503. Weasel did not read pg_index.indisvalid, and pg_get_indexdef renders an invalid index exactly like a valid one.
An invalid index is ignored by the planner. The object exists, \d shows it, Weasel said the schema matched configuration, and every query meant to use it silently did a sequential scan. PostgreSQL leaves an index in that state whenever CREATE INDEX CONCURRENTLY fails partway. Such an index is now reported as drift and rebuilt.
A SQL Server column with a default could never be dropped
#505. SQL Server refuses DROP COLUMN while a default constraint references the column, so a column declared with DefaultValue(...) could be added by a migration but never removed by one — the patch was generated, it just always failed. Weasel now finds the constraint (SQL Server names it itself, so it has to be looked up) and drops it first.
Migrations that failed outright
More than about a thousand tables on SQL Server
#496, and #497 for the public MigrateAsync extensions. Introspection put every object's query into one command; SQL Server refuses a request carrying more than 2100 parameters, and a table's query binds two. Past roughly 1050 tables the migration failed before any comparison happened. Queries are now batched under the dialect's limit.
A least-privilege role could not migrate into its own schema
#495 and #498. PostgreSQL checks CREATE on the database before it evaluates CREATE SCHEMA's own IF NOT EXISTS, so a role granted USAGE, CREATE on one schema and nothing else was refused with 42501 permission denied for database — on a schema that had been there for months. Since that statement opens the script, the whole migration aborted with nothing applied. There is now an existence pre-check, which is what SqlServerMigrator has always done.
A name containing a dot could not be parsed
#501. Qualified names were split on every ., so a legal delimited name containing one — "my.schema".things, [my.table], or the PK_dbo.__MigrationHistory that EF6 gives its own history table — threw at parse time. The split is now delimiter-aware.
Providers expose their identifier rules through a new IDatabaseProvider.Rules property. It is virtual and defaults to null, so a provider implemented outside Weasel keeps compiling and keeps its current splitting.
New: a non-blocking index on a partitioned table
#494. IndexDefinition.IsConcurrent emits CREATE INDEX CONCURRENTLY, which PostgreSQL refuses on a partitioned parent — so adding an index to a partitioned table meant an ACCESS EXCLUSIVE lock for the whole build, a write outage rather than a migration.
Weasel now emits the supported three-step sequence: the parent index ON ONLY, each partition's index built concurrently, and an ALTER INDEX ... ATTACH PARTITION for each. Nothing changes unless an index is both concurrent and on a partitioned table.
SQL Server function bodies: whitespace is now significant
#496. CanonicizeSql collapsed all whitespace and rewrote a list of PostgreSQL-isms that could never appear in T-SQL. It now normalizes line endings — the same source file is CRLF on one machine and LF on another, and both apply to one database — and the CREATE [OR ALTER] FUNCTION preamble, which is the only thing sys.sql_modules does not return as written.
Reformatting a declared function body is now a change. It costs one drop-and-recreate, after which it stays clean. Nothing is lost; the comparison simply matches what the catalog actually stores.