Function head_is_the_highest_visible_position
pub async fn head_is_the_highest_visible_position<F: Fixture>(
open: impl AsyncFn() -> F,
) -> RuleOutcomeExpand description
head() covers the whole store, not the part some default query matches.
ES-30, in the portable form ADR-0013 §4 amends the clause to: append a
batch holding at least one event a narrower default query would not
match, read the whole store with Query::all(), and assert head() is
not below the highest position that read yielded. The relation does
not assert read-your-own-writes, and
head_not_below_the_visible_log carries the whole argument for why. The
anchor does read back the untagged append, so the rule settles the
frontier first (CF-42).
§The seeding is the rule
A rule that appended one uniform batch could not fail: whatever query a
scoped head is written against, one batch of identical events either
matches it entirely or not at all, and in both cases the scoped answer
and the true answer agree. So the highest event here is untagged and
carries a second type, and that is what a scoped head has to miss.
§What it rejects
DefaultQueryHeadStore: SELECT max(e.position) FROM event e JOIN tag t ON t.event_id = e.id — the head statement written against the same
joined view the read path is built around, because there is one view and
reusing it is the obvious move. An untagged event at the top of the log
is invisible to it, so the store under-reports its head to every
projection runner and every paginating caller, and the events past the
under-reported head are exactly the ones nothing else in the deployment
tags.
ES-30’s other two rejected implementations are not this rule’s and were
never going to be. A cached last-written position returns the right
answer against a single handle and is head_advances_across_two_handles’
to reject; the backwards().limit(1) workaround as a universal answer is
hung on E2E-13 by ES-30’s own text.