An adversarial fable red-team of the multi-backend PR surfaced 19 findings; the
verified, non-regressing subset is fixed here (two were confirmed intentional and
left as-is with clarified comments).
Dolt-parity (high):
- sqlkit tx AddLabel/RemoveLabel/CloseIssue recorded no audit event, unlike both
Dolt references, so bd label add/remove and bd batch / mol squash close lost
their EventLabel*/EventClosed rows on Postgres/MySQL/SQLite. Delegate to the
same issueops with-event helpers the embedded oracle uses.
- Heartbeat left updated_at stale on Postgres/SQLite (no ON UPDATE trigger;
MySQL/Dolt auto-bump), so bd stale diverged. Stamp updated_at on the heartbeat
UPDATE (claim already stamps it, so no double-assign).
Correctness / hardening:
- postgres/mysql/sqlite stampSchemaVersion: idempotent insert (ON CONFLICT /
INSERT IGNORE / INSERT OR IGNORE) + re-read, so concurrent first provisions do
not fail one opener with a duplicate-key error.
- sqlkit tx SetConfig now syncs custom_statuses/custom_types like the oracle.
- pgdialect INSTR->POSITION leaves the call loud when both args carry a
placeholder, preventing a silent parameter swap.
- creds command runner uses cmd.exe on Windows (sh is POSIX-only).
- Dolt gateway credential command is gated on the Dolt backend, not just
shared-server mode (was failing closed for SQL workspaces).
- Scrub pg/mysql URL passwords from the bd.args telemetry span.
Conformance rigor:
- E2E normalize() maps zero/epoch timestamps to <ZERO-TS> so a lost timestamp no
longer reads as byte-identical to a real one.
- E2E logs its candidate set and honors BEADS_CONFORMANCE_REQUIRE so a
sqlite-only run cannot masquerade as full coverage.
Docs/hygiene:
- Scrub dev-machine paths (/data/tmp/pg-recon, proposal_2.md, "proof-wedge") from
the Postgres schema DDL comment; clarify the SQLite timestamp comment.
Deliberately not changed (would regress): sqlkit UpdateIssue uses the without-event
variant to MATCH dolt-server (only embedded differs); sub-second timestamp
precision is an intentional per-backend waiver (whole-second round-trip is the
portable contract) — forcing round-down would discard PG/SQLite fidelity.
Verified: embedded + postgres + mysql + sqlite + server-dolt TestConformance and
the pg/mysql/sqlite E2E differential all pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Maintainer fixup for the choose-your-own-storage-backend PR, amended across the
pre-review CI-repair and the review-loop apply-fixes iterations so contributor
commits stay intact under one maintainer commit.
CI repair (pre-review):
- default.nix: bump vendorHash for the added Postgres/MySQL driver deps.
- conformance.yml: provision the Postgres service DB with a C-locale cluster so
text ORDER BY sorts by code point, matching Dolt's utf8mb4_0900_bin oracle.
Review fixes — blockers:
- uow/doltserver_tx: on a non-serialization DOLT_COMMIT failure, ROLLBACK on the
pinned session before releasing it, and poison (pool-discard via driver.ErrBadConn)
the connection when the rollback also fails, so a later borrower can't inherit an
open transaction and implicitly commit orphaned writes. Satisfies the existing
doltserver_tx_test rollback/poison cases.
- sqlitedialect: translate GREATEST/LEAST to SQLite scalar max/min so the
ReconcileChildCounters upsert (GREATEST(last_child, ?)) no longer aborts creation
or import of hierarchical dotted issue IDs (e.g. x-1.2) on the SQLite backend.
Review fixes — majors:
- conformance: add a portable dotted hierarchical child-ID scenario (x-1, x-1.1,
x-1.2, x-1.2.1) exercising the ReconcileChildCounters GREATEST path on every
backend — the coverage hole that let the SQLite blocker ship green. Verified
against the embedded-Dolt oracle and the real SQLite driver.
- postgres: warn at provision/open when the database collation is not code-point
ordered (C/POSIX/C.*), since text ORDER BY otherwise diverges from Dolt silently.
- sqlite: pin the pool to one connection and add busy_timeout + WAL to the DSN so
overlapping operations wait instead of surfacing a raw "database is locked"
(matches the Dolt backend's single-connection posture). Adds a concurrency guard.
- cmd/bd: gate explicit `bd dolt commit` / `bd vc commit` with a typed
storage.ErrUnsupported on non-Dolt backends instead of the no-op Commit printing
"Committed."; internal opportunistic auto-commit is untouched.
Review fixes — attempt 2 (stale init backend-flag test contract):
- cmd/bd/init_test, init_embedded_test: the init backend-flag tests still asserted
the pre-#3151 one-backend contract (--backend=sqlite must fail "SQLite backend
has been removed"; --backend=postgres is an "unknown backend"), so
`go test ./cmd/bd -run '^TestInitBackendFlag$'` failed in a clean environment now
that sqlite/postgres/mysql are supported pluggable backends. Rewrite
TestInitBackendFlag and the cgo TestEmbeddedInit sibling (rejected_backends →
sql_backend_flags) for the new contract: --backend=sqlite initializes and
persists backend=sqlite; --backend=postgres is recognized and fails only on
missing connection config (not "unknown backend"); a genuinely unknown backend
is still rejected. Pin BEADS_DIR and scrub BEADS_*/BD_* so the subtests are
hermetic against an ambient .beads or Dolt server.
Cleanups (non-gating):
- cmd/bd/init: extract checkExistingSQLBackend so the added SQL-backend branches
don't inflate checkExistingBeadsDataAt's complexity, and load config once.
- configfile: correct the stale "Deprecated: always dolt" Backend field comment.
Review fixes -- attempt 3 (SQLite datetime parity blocker + conformance/rewrite
hardening of the storage-parity gate):
- sqlite/dsn: add _time_format=datetime. Without it modernc binds time.Time via
t.String() ("2026-07-09 12:34:56.123 +0000 UTC"), a shape SQLite's date functions
cannot parse, so the counts mega-query's DATE_FORMAT(created_at,...) (strftime on
SQLite) returns NULL and deps_json exposes a zero created_at while Dolt emits the
real timestamp. datetime stores whole-second "2026-07-09 12:34:56", matching
Dolt/MySQL datetime(0). Verified end-to-end against the real modernc driver.
- conformance: add ReadyWorkDepCreatedAtParity -- imports a dependency with a known
created_at (the bd import / CreateIssuesWithFullOptions bound-time.Time path) and
asserts the counts-rendered deps_json created_at equals it, closing the blind spot
that let the SQLite blocker ship green (the counts tests only checked DependencyCount).
Proven to fail on the pre-fix DSN and pass on both the fixed SQLite backend and the
embedded-Dolt oracle.
- conformance: add ClaimReadyIssueConcurrentExclusivity -- races 24 workers over 8
ready issues and asserts each is claimed exactly once, exercising the SQL backends'
multi-writer claim property that was previously only asserted, never raced. Gated to
the multi-connection dialects (postgres/mysql) via DialectName; SQLite (single-conn
pool) and embedded-Dolt skip.
- conformance.sh: run the embedded-Dolt reference conformance with
BEADS_TEST_EMBEDDED_DOLT=1 and fail loudly if it skips. It self-skips without the env
var, so the parity gate was reporting success while never running the reference oracle.
- mysqldialect: drive the is_blocked-recompute error-1093 rewrite test from the real
issueops templates (new issueops.SelfReferentialRecomputeTemplates) instead of a
hand-copied fixture, so template drift fails a non-gated test rather than only the
env-gated MySQL conformance suite. All four mark/unmark x issue/wisp templates
verified rewritten into the materialized derived-table form.
Note: the review re-flagged TestInitBackendFlag as failing, but it passes in a clean run
(plain and gms_pure_go, matching the scorecard's own rerun); the only cmd/bd failure in
this sandbox is an unrelated pre-existing embedded-Dolt/CGO test not in this PR's diff.
Review fixes -- attempt 4 (conformance-gate self-brick + all-wisps inline-dependency
drop blockers; Postgres error-log password-leak major):
- conformance.sh: anchor the embedded-Dolt reference skip/pass guards to top-level
result lines (^--- SKIP/PASS: TestConformance). The attempt-3 loud-fail guard used
an unanchored grep, so the indented, legitimate subtest skip of
ClaimReadyIssueConcurrentExclusivity (embedded-Dolt is single-writer) matched it and
false-failed the entire storage-parity gate on every run even when the top-level
suite passed. Whole-suite-skip protection is preserved.
- sqlkit/portable + embeddeddolt/create_issue: the all-wisps arm looped the
single-issue issueops.CreateIssueInTx, which runs no batch dependency-persist pass,
so an all-wisps batch carrying inline dependencies (e.g. importing no-history beads
with inter-bead edges -- export DOES include no-history beads, GH#2619) silently
dropped every edge (not even reported via OnSkippedDependency) and, one transaction
per issue, forfeited batch atomicity. Both now mark ephemeral, then run the whole
batch in ONE transaction through issueops.CreateIssuesInTx, matching the server-mode
DoltStore all-wisps arm. The embedded-Dolt reference oracle had the same latent drop,
so the oracle and the new SQL backends were corrected together to keep conformance
parity (the finding compared sqlkit against the server-mode DoltStore, which already
persisted these edges).
- conformance: add CreateAllWispsInlineDependencies -- an all-wisp batch with one valid
inter-wisp edge and one unresolvable edge; asserts the valid edge persists and the
unresolvable one is reported via OnSkippedDependency. Fails on the pre-fix per-issue
loop, passes on the batch path; verified on SQLite and the embedded-Dolt oracle, with
a full embedded-Dolt conformance re-run clean.
- pgdialect: scrub query-param/sslpassword secrets from pgx ParseConfig errors in Open
and OpenRaw (new ScrubDSNError). pgx's ParseConfigError redacts only URL userinfo, so
a postgres://u@h/db?password=SECRET DSN (the shape bd init connects with) leaked the
cleartext secret through the wrapped parse/TLS-config error into logs. Connect-time
errors were verified not to carry the password. Adds redact_error_test coverage for
the query-param, sslpassword, userinfo, and libpq-keyword shapes plus the real
Open/OpenRaw parse-failure path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rebase of PR #4601 (feat/choose-your-own-backend) onto current origin/main,
squashed to a single commit because the original 73-commit history built and then
removed the uowStore derisk spike incrementally, which does not replay cleanly
over the 125 commits main has since advanced (notably #4679, the ready-work
counts fix, which merged to main and is now inherited rather than re-applied).
The full pre-rebase history remains on the original branch tip for reference.
Contents (multi-backend feature + the six review follow-ups):
- bd init --backend=postgres|mysql|sqlite provisioning behind the Dolt-parity
seam; the three SQL backends share *sqlkit.Store over pluggable dialects and
carry a generated typed-unsupported shell for the non-core surface.
- Passwords in init DSNs are redacted before persisting to metadata.json (driver
parser + re-parse verification), with a negative-test matrix proving no secret
survives (pgdialect + mysql).
- The uowStore derisk spike and its report artifacts are removed; the shell
generator now lives in internal/storage/unsupportedgen (neutral name).
- Compile-time guard `var _ storage.DoltStorage = (*Store)(nil)` on each backend
plus a regen-idempotence test keep the generated skip lists honest.
- sqlkit exposes CountReadyWork so every SQL backend satisfies
storage.ReadyWorkCounter and inherits the shared page-pushed ready-counts path.
- Commercial/product names scrubbed from the feature's own code for vendor
neutrality.
Rebase reconciliations: adopted main's metadata-load hardening in the store
factories alongside the backend dispatch; combined the config interface
(GetMetadata + ReconcileVersion) and the conformance registrations (ReadyCounts +
Claim/lease); took main's versions of the shared-infra concurrency (uow
doltserver_tx) and delete-cascade (domain/issue_delete) evolutions; implemented
the interface's new UnclaimIssue on sqlkit (delegating to issueops) and added it
to the skip lists.
Verified: go build (cgo + CGO_ENABLED=0) and go vet clean; unit tests
(sqlbuild/issueops/pgdialect/mysql/unsupportedgen) and the SQLite conformance
suite green (RunAll, incl. ReadyCounts equivalence and Claim/lease on the real
sqlkit backend).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>