MVCC checkpoints are always TRUNCATE, plus they block all other transactions.
This guarantees that never need to let transactions read from the SQLite WAL.
In MVCC, the checkpoint procedure is roughly as follows:
- Take the blocking_checkpoint_lock
- Write everything in the logical log to the pager, and from there commit to the SQLite WAL.
- Immediately TRUNCATE checkpoint the WAL into the database file.
- Release the blocking_checkpoint_lock.
Currently header changes are tracked through pager by reading page 1.
MVCC has it's own layer to track changes during txn so this commit makes
it so that headers are tracked by each txn separately.
On commit we update the _global_ header which is used to update
`database_size` because pager commits require it to be up to date. This
also makes it _simpler_ to keep track of header updates and update
pager's header accordingly.
this is only used for returning LimboResult::Busy, and we already
have LimboError::Busy, so it only adds confusion.
Moreover, the current busy handler was not handling LimboError::Busy,
because it's returned as an error, not as Ok. So this may fix the
"busy handler not working" issue in the perf thrpt benchmark.
We start a pager read transaction at the beginning of the MV transaction, because
any reads we do from the database file and WAL must uphold snapshot isolation.
However, we must end and immediately restart the read transaction before committing.
This is because other transactions may have committed writes to the DB file or WAL,
and our pager must read in those changes when applying our writes; otherwise we would overwrite
the changes from the previous committed transactions.
Note that this would be incredibly unsafe in the regular transaction model, but in MVCC we trust
the MV-store to uphold the guarantee that no write-write conflicts happened.
We must iterate the row versions in reverse order because the
versions are in order of oldest to newest, and we must commit
the newest version applied by the active transaction.
In insert_version_raw(), we correctly iterate the versions backwards
because we want to find the newest version that is still older than
the one we are inserting.
However, the order of `.enumerate()` and `.rev()` was wrong, so the
insertion position was calculated based on the position in the
_reversed_ iterator, not the original iterator.
Based on #3126Closes#3029Closes#3030Closes#3065Closes#3083Closes#3084Closes#3085
simple reason why mvcc update didn't work: it didn't try to update.
Closes#3127
1. commit state machine was assuming that begin_write_tx() cannot
fail, but it can fail if there is another tx that is not using
BEGIN CONCURRENT.
2. if a brand new non-CONCURRENT transaction attempts to start
exclusive transaction but fails with Busy, we must end the read
pager read tx it just started, because otherwise the next time
it attempts to do something it will panic with:
"cannot start a new read tx without ending an existing one"
