Reproduced the reported ~50% intermittent ancestry failure (3-4 of 8 runs) and
captured it: NamespacedDirectory.roots returned [dir4, dir1] vs [dir1, dir4].
The session report blamed 'several dependent saves that occasionally race' —
wrong. Both records share position 0 and run_tree ordered by position alone,
so the tie order was whatever the index scan produced that run. No race, no
CAS involvement: an ORDER BY tie.
- run_tree now tie-breaks by created_at then META(c).id, exactly like
run_docs already did for finder queries.
- The two client-side tree sorts (descendants, build_tree) get the same
tie-break: Ruby's sort_by is not stable, so [depth, position] alone could
reorder same-position siblings between runs.
ancestry_spec re-run repeatedly post-fix: no failures.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- spec/support/adapter_contract.rb shipped shared examples ('a simply_couch
adapter') that no spec ever included — the parity proof they document never
ran. store_contract_spec.rb now includes them against the live Store.
- concurrency_spec.rb: the suite had zero multi-threaded examples despite the
gem's core claims being CAS optimistic concurrency and Redis-replacement
primitives. Adds: counter increments under 8 concurrent writers, lock
granted to exactly one contender, with_lock mutual exclusion, stale-handle
release after TTL expiry (must not unlock the new holder), same-attribute
CAS conflict surfacing (no lost update), different-attribute auto-merge, and
a 3-thread disjoint-attribute write storm.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Counter#increment now seeds a missing key at the delta (Redis INCR/INCRBY
semantics). Couchbase's native increment seeds at 'initial' and ignores the
delta on creation, so the first INCR returned 0 — a drop-in replacement that
silently undercounts by one is exactly the kind of bug nobody ever finds.
Explicit initial: still overrides. Decrement keeps initial: 0 and its
unsigned clamp-at-zero divergence from Redis is now documented and tested.
- Lock#release requires the CAS handle: passing nil used to delete the lock
document unconditionally, so any caller could unlock a lock it never held.
force_release! is the explicit administrative override.
- Lock#acquire validates a positive ttl — a nil/zero ttl would create a lock
that survives its holder's crash forever.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- database.view now raises SimplyBased::UnsupportedView when a declared view
carries CouchDB-era options the N1QL translation silently ignored before:
:conditions (a JS fragment) and :emit_value, plus CouchDB list functions.
Silently dropping :conditions was the worst of these — a
'view :completed, key: :name, conditions: ...' query returned *unfiltered*
results with no hint anything was wrong. Covered by unsupported_view_spec.
- All column/property names interpolated into N1QL are now routed through
quote_ident, which rejects backticks/backslashes. These names come from the
model DSL rather than user input, but a smuggled backtick must never become
query text. Class names in index predicates go through a proper string
literal quoting helper for the same reason.
- ensure_index: failures are still swallowed (an index build in progress must
not take down queries) but now emit a warning naming the index, because the
silent version left every subsequent query on a primary-index scan with no
trace. The ensured-indexes bookkeeping is mutex-guarded for multithreaded
servers (Puma).
- bulk_load / bulk_destroy now actually batch (collection.get_multi /
remove_multi — one KV round trip). The README already advertised
'bulk (batched KV)' but the code looped one get/remove per id.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Live-run finding: the full suite against Couchbase was 423 examples with 17
failures — all in embedding_spec. The session report claimed embedded specs
were 'excluded by design', but the exclusion filter targeted a :embedding tag
no example carries, so the file ran (and failed) all along.
Three fixes:
- HasManyEmbedded::Property#build wired parent_object before installing the
collection ivar on the parent; parent_object= reaches back into the parent's
collection getter, which re-entered build -> SystemStackError on any access
to embedded documents loaded from the database. Install the ivar first.
- EmbeddedClass.all / .count referenced raw JS views (all_documents /
all_documents_for_count) that the N1QL port silently drops, so .all returned
[] and .count raised NoMethodError. Rebuilt them by enumerating the parent
documents (same work as the old view scan, client-side; an UNNEST query can
replace it later).
- Only reverse lookups (N1QL on a foreign key inside an embedded array) remain
unsupported: tagged :embedded_in_reverse_lookup and excluded in spec_helper
with an explanatory comment, replacing the broken :embedding filter line.
embedding_spec.rb: 16 examples, 0 failures against live Couchbase 8.0.1 CE.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
bterkuile
