Function committed_rows_and_checkpoint_survive_a_reopen
pub async fn committed_rows_and_checkpoint_survive_a_reopen<F: ProjectionFixture>(
open: impl AsyncFn() -> F,
) -> RuleOutcomeExpand description
A committed read-model row and its checkpoint are both still there after every handle is dropped and the fixture reopens its storage.
PS-1’s durable, read across a reopen. Every other rule in this family observes a commit through a second handle onto the same running store, which is what PS-1’s coupling needs and is exactly what a store that never reaches its medium survives: the process still holds the rows and the checkpoint, so every fresh handle sees them. This rule is the one that looks after the process’s hold has been let go.
Rejects: InMemoryDatabaseStore — an adapter whose committed state
lives in process memory, an in-memory database shared behind one handle
or a write-back cache never flushed, so a reopen finds the checkpoint
NeverRun — and VolatileReadModelStore, whose checkpoint is durable and
whose read model sits in a process-local cache, so a reopen finds a
checkpoint that claims rows the read model no longer has. The second is
the sharper of the two: it is PS-1’s coupling broken by a restart, and it
passes commit_is_atomic_with_the_read_model, because before the reopen
both halves are present.
§Why it is capability-gated
A reopen needs the fixture’s co-operation — which file to close, which
pool to drain — and a store with no medium outside the process has
nothing a reopen could read back. So the rule is gated on
REOPEN, which CF-17 governs, and a
fixture that declines it gets this rule as a reported skip carrying its
own reason. MemoryProjectionFixture declines it, so a reference run
prints that skip.
The writer is dropped before the reopen, on purpose: a handle that carried the state across would make the rule pass for the wrong reason. The checkpoint is compared against the position the commit was given, held in a binding, never a literal (CF-6).