* Consistently lock in ensureVersionTable and do not call ensureVersionTable from Drop across all database implementations
* Add test for dropping postgres databases
* Fix failing database tests
* Fix CockroachDb test, lock table should be created before versionTable
* Add Initialize() to Driver interface, and add integration tests for Drop() between database implementations and migrate
* Remove Initialize, document breaking behaviour of Drop
* Revert introduction of Initialize method
* Removed Initialize in Stub as well
* Remove call to non-existent Initialize and make sure to close re-initialized database connections
* Revert changes to TestDrop in database/testing
* Split Test and TestMigrate into different test entrypoints
* Remove unused import in migrate_testing
* Remove erroneous code to fix tests
* Add stub source imports to database tests
* Add Stub source to migrate tests
* Use example migrations for tests
* Add file driver to database tests
* Align database directory layout
* Add file source driver to Cassandra
* Review changes
* Minor syntactic change for cleaner diff
I believe this closes#297 as well.
I have been working on adding testing of migrations and it requires acquiring
the lock in mysql multiple times to go up and down. After nailing this down to
GET_LOCK returning a failure for every subsequent GET_LOCK call after the
first, I decided it was time to rtfm and lo and behold:
https://dev.mysql.com/doc/refman/5.7/en/miscellaneous-functions.html#function_release-lock
RELEASE_LOCK will not work if called from a different thread than GET_LOCK.
The simplest solution using the golang database/sql pkg appears to be to just
get a single conn to use for every operation. since migrations are happening
sequentially, I don't think this will be a performance hit (possible
improvement). now calling Lock() and Unlock() multiple times will work;
prior to this patch, every call to RELEASE_LOCK would fail. this required
minimal changes to use the *sql.Conn methods instead of the *sql.DB methods.
other changes:
* upped time out to 10s on GET_LOCK, 1s timeout can too easily leave us in a
state where we think we have the lock but it has timed out (during the
operation).
* fixes SetVersion which was not using the tx it was intending to, and fixed a
bug where the transaction could have been left open since Rollback or Commit
may never have been called.
I added a test but it does not seem to come up in the previous patch, at least
easily, I tried some shenanigans. At least, this behavior should be fixed with
this patch and hopefully the test / comment is advisory enough.
thank you for maintaining this library, it has been relatively easy to
integrate with and most stuff works (pg works great :)
