The plan-wide acts (tier removal, default change, reorder, tier add and
delete) are now sent against each other and against a conferral, each
interleaving chosen by a lock a side transaction holds and releases once
the requests wait on it. dbtest.AwaitLockWait proves a request is queued,
and PostgreSQL grants a row lock to its waiters in the order they queued,
so no test sleeps. The kit gains Backend.Waiters, how many backends must
wait at once, with a test of its own; its first caller is these tests.
TestTierRemovalAgainstDefaultChange runs a removal and a default change
on one pool in each order: queued on the pool's row, and one after the
other. Every case ends with no live position on a deleted tier.
TestDefaultChangeTwiceSerializesOnThePool sends one default change twice
and shows one request grandfathering the pool and the other finding the
legacy grant the first wrote. TestPlanActsRefuseWhileARuleChangeHoldsTheRendezvous
holds the exclusive rendezvous while all five acts run, and pins each
act's contention answer: the removal's and the organization's "Another
entitlement set change is in progress. Try again." and the generic
sentence of the reorder, the add and the delete. It waits out the 5 s
lock timeout once. TestRenumberingCommitsDuringAConferral commits a
reorder, a tier delete and a tier removal while a default change's
conferral waits between its rank reads; the files pin the transition as
it is recorded today (upgrade 0 to 1 for the reorder where a transfer
belongs; downgrade 1 to 0 for the delete and the removal), because a
renumbering takes the rendezvous in shared mode only. Pinned with #187:
TestReorderEnactsAgainstTheRankZeroItSaved sends two reorders whose
enactments queue on one pool; the first enacts against the rank 0 it
saved after the second has put the other product back at rank 0, so it
grandfathers the holder onto a legacy grant of the product that is rank 0
again, and the second finds the holder on that grant and leaves it.
No output changed: the commit adds tests, their golden files and one field
of the test kit, and touches no production file. The names and results of
./internal/server and ./internal/testkit/..., run after a database reset at
the parent and at this commit, differ by the 13 new tests and subtests, all
passing; the new tests passed three runs in a row and 20 runs of the four
that do not wait out the timeout. The edit to the plan-wide acts' golden
test is its JSON encoder, split so the new tests encode their files the same
way; its 44 files are unchanged. The kit's wait.go and wait_test.go and the
AwaitLockWait bullet of docs/testing.md change beside it. 14 breaks, each
caught by its new test (new side only): 12 edit production code unchanged
from the parent and 2 edit this commit's wait.go. The section's tests take
about 6 s, 5 s of it the one wait.
Unprobed:
- Interleavings the tests do not choose, such as a removal that arrives
between the default change's save and its pool lock; the cases cover the
two queue orders on the pool and the two serial orders.
- A conferral other than the default change's during a renumbering (a
reconcile, a grant act, the floor); it is the same database function,
which the reorder case reaches.
- The default change's own stale candidate: a renumbering between its save
and its enactment. It uses the same fixed product the reorder case
reaches; no case pauses the default change there.
- The two new failure messages of AwaitLockWait for a count above one;
they are text, read only when a wait fails.
cgalo5758
6f65c744fe