This improves upon #2767.
Previously, the code would swallow any panics, which isn't ideal;
panic is intended to be used when a critical error arises,
where the process should fail fast instead of trying to limp along.
This now defers the any recovery (if desired) to the client code.
s.New() will clear all search conditions and search value,when I use Table() to set a table name. Then FirstOrCreate() will use struct name as my database table name,so It doesn't work.
The map is modified in different places in the code which results in race conditions
on execution.
This commit locks the map with read-write lock when it is modified
* Edit DB.clone(), DB.Dialect(), and Scope.Dialect() preserve transactions.
* Adds a test case for tables creations and autoMigrate in the same transaction.
Exporting sqlCommon as SQLCommon.
This allows passing alternate implementations of the database connection, or wrapping the connection with middleware. This change didn't change any usages of the database variables. All usages were already only using the functions defined in SQLCommon.
This does cause a breaking change in Dialect, since *sql.DB was referenced in the interface.
gorm.Errors, which usefully implements `error` for an `[]error` as
returned by `DB.GetError()` was already exported, but because it used a
private field `errors`, it was not able to be created due to the
compile-time error:
implicit assignment of unexported field 'errors' in gorm.Errors literal
The trivial solution would be to export the `errors` field on
`gorm.Errors`, but this led to the issue that the common pattern of
checking `err != nil` failed because a struct{error: nil} != nil.
We can take advantage of type aliasing here to make Errors an []error,
which can in fact be nil and would pass `err != nil` on the happy path.
* Remove `(Errors) GetErrors()`, as it's less useful when Errors is an
[]error which can be iterated over. While this is technically a
breaking change, we never expose an Errors and its difficult to build
one (it can be done with the existing `(Errors) Add(error)`), but
awkwardly. This removal can be reverted without issue and we can make
it an identity method, but it seemed an opportune time to reduce API
surface area on something that likely isn't used.
* Remove errorsInterface, as it's not useful without `(Errors)
GetErrors()`
* Change `(*Errors) Add(error)` => `(Errors) Add(error...) Errors`
because we can't modify even a *Errors when it's a type alias. This is
more idiomatic as it follows the pattern of `slice = append(slice,
element)` Go developers are familiar with.
Charles Strahan