Complete M7.3 fenced region ownership

ADR-0060 and schema v2 add database-time region leases and a fenced head
advance. Acquire/renew/release/advance lock the lease row, check holder,
epoch and expiry with clock_timestamp() after the lock wait, and the
advance compares the exact predecessor. Holders are DB-generated tokens;
epochs never wrap or repeat. The runtime group role gets EXECUTE on five
SECURITY DEFINER functions only - no table, provisioning or CREATE rights.
Migration history admits clean->v2, v1->v2 and a v2 no-op, and refuses
privileged runtime roles.

Host side: portable LeaseAuthority fences from send-time deadlines on
CLOCK_BOOTTIME, keeps the earlier deadline on unknown renewal and stays
fenced until a newer epoch. Selected timings: 20s lease, 5s renew, 2s
margin, 2s renew and 5s commit deadlines. Linux evo::postgres wraps the
functions with typed LeaseStoreError results.

Validated: Windows CTest 55/55; WSL Release, clang-tidy and ASan/UBSan
private-PostgreSQL suites (contenders, expiry, paused/queued stale
writers, unknown renewal, exhaustion, bypass denial, upgrade path).
Three SQL mutations were each caught. Docker/Gitea CI, remote TLS and
cluster runs were not executed. No director, host or deployment added.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
dingo
2026-09-28 11:14:32 -04:00
co-authored by Claude Opus 5.5
parent 598b3938e6
commit 2496649a9b
22 changed files with 1501 additions and 29 deletions
+40
View File
@@ -2041,3 +2041,43 @@ claim lineage recovery. No retained-history or World0 requirement was relaxed.
The adapter is separate from the headless kernel. Region fencing, publication,
ambiguous durable commit reconciliation and crash recovery are later M7 gates.
Only clean-to-v1 and exact-v1 restart are supported migration paths at this gate.
## M7.3 fenced region ownership - 2026-09-28
- Windows MSVC Release configure/build/CTest passed 55/55, including the new
portable `region-lease-contracts` test. Repository clang-format 19 check passed.
- WSL Ubuntu (GCC 14.2.0; clang 19.1.1; PostgreSQL/libpq 18.6 source build):
GCC14 Release, clang19 analysis (clang-tidy warnings-as-errors) and clang19
ASan/UBSan builds of `evo-postgres-tests` each passed `tools/ci/postgres-check.sh`
against a fresh private cluster. The first analysis build flagged two
unchecked optional accesses and two unreserved vectors. These were fixed and
all three configurations were rerun after formatting.
- Real-database coverage, as the runtime login rather than SET ROLE:
- Denied (42501): table read/write, provisioning, the internal check function,
and object creation. A hostile search_path is ignored.
- Busy, forged-holder, stale-epoch and released-lease rejections; epoch
non-reuse.
- 1..64 tick delta; predecessor mismatch; exact-retry detection via the head.
- 1 s database-time expiry and successor acquisition; paused stale writer fenced
locally and by the database.
- A writer queued 1.3 s on the lease row lock is rejected after expiry.
- A 100 ms renewal deadline under lock (unknown) fences locally before a
successor acquires.
- 64 two-thread acquisition rounds with exactly one winner and consecutive
epochs; 32 racing head advances with exactly one winner.
- Epoch and sequence exhaustion.
- Migrations tested: clean->v2, repeated v2 no-op, v1->v2 upgrade, and failed
0001 rollback. Superuser and missing runtime roles are refused with no schema
left. Unknown v3, changed v2 checksum and changed v1 checksum are refused.
- Mutation sensitivity (scratch copies only): replacing post-lock
`clock_timestamp()` with `now()`, removing the lease check from head advance,
and disabling the busy check each made the suite fail at the expected test.
- Local Unix-socket latency, 200 samples per operation in each final run:
- p50 about 0.61-0.74 ms and p99 about 0.8-1.15 ms for acquire, renew,
advance and release.
- Max up to 2.8 ms under analysis/ASan builds.
These are local numbers, not remote TLS or cluster measurements.
- Not executed: Docker/Gitea CI, remote TLS endpoints, Kubernetes Secrets, real pod
kill, real network partition (a lock wait plus client deadline stands in for
it), VM freeze and TSan. The contention threads each own a separate
connection; libpq is uninstrumented.