Function batch_reads_reflect_pending_writes
pub async fn batch_reads_reflect_pending_writes<F: ProjectionFixture>(
open: impl AsyncFn() -> F,
) -> RuleOutcomeExpand description
An open batch can be read through, and sees its own pending writes.
PS-12. A projection that maintains a counter, a running total or any
other read-modify-write does get then set inside the batch, and a
store whose batch get answers from committed state loses every write
the batch has not committed yet — silently, and worst exactly when the
chunk is large.
Rejects: CommittedReadBatchStore — a batch get that goes to the
store. It is the natural shape rather than a contrivance: one round
trip on the connection the batch is already holding is what an adapter
author writes, and it is right for every projection that never reads what
it wrote.
The assertion is on the value written, not on is_some(): a store
that answered Some(0) for every key would pass a presence check while
losing the write.
§What a skip here does not cover, stated rather than left to be inferred
Declaring READS_THROUGH_BATCH
false is a conformant answer — PS-12’s second arm permits a batch with
no read path at all — and this rule then reports a skip. What that skip
does not say is that the adapter is safe for read-modify-write
projections. It says the opposite: such an adapter has no read path for a
projection’s apply to use, which is the situation §4.4 describes going
wrong, and rebuild_is_chunk_size_invariant is unverified for it for
exactly the same reason.