IndexedDbVFS rollback does not work unless cache_size=0 about wa-sqlite HOT 1 CLOSED

The problem is in the new "no journal" journal, which is a hack to take advantage of the fact that we can always store all database changes from SQLite transaction as a single atomic write to IndexedDB. The VFS skips passing all the journaled blocks back and just cancels the write.

This is working fine for the database as stored in IndexedDB. But it means that SQLite's internal cache of the database isn't rolled back properly. In general it isn't safe for SQLite to keep using that cached data because once the transaction lock is released some other connection could modify the database. But SQLite has a trick where every change to the database increments a counter in the database header. When SQLite gets a new lock, it reads that counter and if it hasn't changed then the cached data should still be valid. Except that now it isn't because the rollback hack allowed the two to diverge.

The solution should be just to increment that counter. We'll see if that works.

from wa-sqlite.