|
As per expected by SQLite, create new connections so they have
independent transaction, and serialize its use via a channel with a
single consumer.
== Other changes
=== Remove `db` attribute from `queueT` type
Now that `initDB()` can create many database connections (A.K.A. opaque
pointers to the `sqlite3*` object), it makes more sense to hand to it
the responsability to create and destroy these databases. So now a
`queries.close()` includes what previously was `db.Close()`, and needing
to do it was the only reason that made us keep the `db` attribute.
The arguments to `initDB()` were adjusted to reflect that, as it no
longer is given an initialized database handle, but only the database
path, and does the creation of the handles by itself.
=== Shoehorn multi-statement queries into single-statement ones
In order to avoid creating one `execSerialized()` function per attribute
of the `queryT` queries, we try to leverage SQLite's built-in
transactionality per individual statement. So compound queries that were
previously done in multiple statements wrapped with a `inTx()` that
could be shoe-horned into a single `SELECT` were rewritten to avoid
having more of this application-level serialization.
The topic/consumer owner validation was also changed: now it is a
trigger on the relevant tables that `ABORT` the implicit/invisible
transaction and makes the `INSERT` fail.
In order to make this possible, the `"%s_owners"` table now has an extra
column: `instance_id`. Despite being extra data, this meta-information
isn't duplicated from anywhere else, and it is an actual useful
information for operators to leverage.
So now the trigger is responsible for stopping the transaction from
going forward without adding an explicit `inTx()` around it, and the
writes can mostly be shrunk to single-statement queries.
Fortunately, all this is enough to fix the `new-instance-takeover`
functional test.
|
