fix: verify server snapshots and recover whole migration sessions (#3)

This commit is contained in:
Mikhail Chusavitin committed 2026-10-26 22:48:07 +03:00
1 parent 3d4101ecc0
commit 8569871a17
29 files changed
+1923 -133

No files matched your search

+22 -2
View File
@@ -441,6 +441,25 @@ MariaDB:
`UPTO` as applied **without running them**, for a server whose schema is already current;
- migrations use their own connection without read/write timeouts (`openMigrationDB`),
not the app's 3s-timeout sync connection — a long backfill must not be cut mid-file;
- server migration commands finish before runtime connection workers start;
- `-migrate` requires a DBA maintenance window with every other writer stopped,
directly granted global ALL PRIVILEGES, and no other connection using the schema.
Do not expand runtime-user privileges; use the existing developer DBA workflow;
- the runner holds a schema advisory lock and one connection for the entire session.
It automatically writes a full consistent snapshot before changes and per-table
snapshots before every file; each archive is restored into a temporary schema
and compared before any corresponding original-schema change. Snapshot or
verification failure stops the run. See [06-backup.md](06-backup.md);
- on a statement/verify failure, per-step snapshots are restored in reverse order,
including migration history; a full-session snapshot is the fallback. A persistent
marker refuses retry after interruption until `qfs -migrate-restore <archive.zip>`
restores and verifies the original state. Connection/recovery failure retains the
marker and archives and emits an explicit restore diagnostic;
- recovery supports InnoDB base tables and their triggers. Views, routines,
events and cross-schema FKs block migrations before changes; new migrations must
stay within supported table SQL. Existing triggers are captured with their
definer, body, order and SQL/charset context, then restored after all data; Verification requires CREATE/DROP DATABASE.
The runner refuses NO_BACKSLASH_ESCAPES sessions;
- each file, its `-- verify: <description> | <SQL>` queries and history insertion run
on one database session, preserving session variables;
- every verify query and the history insert use a fresh GORM statement
@@ -452,8 +471,9 @@ MariaDB:
- only duplicate column/key names and already-existing objects are tolerated.
FK creation errors (including errno 150/121) remain fatal;
- each file must be safe to re-run after a partial failure (`ADD COLUMN IF NOT EXISTS`,
idempotent `UPDATE ... WHERE <not yet done>`). The current statement splitter assumes
semicolons occur only between statements; do not add semicolons inside literals;
idempotent `UPDATE ... WHERE <not yet done>`). The statement parser preserves
semicolons in literals/quoted identifiers and strips ordinary comments. Malformed
quotes, executable comments and DELIMITER directives are rejected before execution;
- there is no GORM `AutoMigrate` against MariaDB: it retypes production columns to match the
running binary's models. On 2026-09-24 a stale v1.18 binary run with the old AutoMigrate
`-migrate` did exactly that, which is why it was removed.