* fix(store-factory): honor shared-server mode in prime's store resolution
newDoltStoreFromConfig and its read-only sibling
openNonMutatingStoreFromConfig checked only cfg.IsDoltServerMode(),
which deliberately does not read dolt.shared-server from config.yaml
(to avoid a circular import with the doltserver package). A workspace
with config.yaml tracked but metadata.json gitignored as machine-local
state — the common shape of a linked git worktree — left
configfile.Load returning (nil, nil), so both paths silently fell
through to the embedded backend and created a phantom database named
"beads" instead of reaching the real shared server.
cmd/bd/main.go's own top-level command resolution already compensates
for this exact gap (GH#3817), but bd prime resolves its store through
a separate path (formatMemoriesForPrime -> ensureStoreActiveForPrime
-> ensureStoreActiveWithContext -> newDoltStoreFromConfig) that never
got the same fix. Centralized the compensation into
effectiveServerMode so the two internal paths and main.go's own
resolution cannot drift from each other again.
Fixes#6551.
* fix(store-factory): route read-only shared mode
* fix(store-factory): reach the !cgo twins and scope shared-server per workspace
Maintainer review fixes for PR #6589 (GH#6551). Addresses the three majors in
the review synthesis, plus the two non-gating items, and corrects two false
claims in the PR's own new prose.
1. major — the !cgo twins kept the unfixed gate, and "centralizing the check
here keeps them from drifting again" was structurally false.
effectiveServerMode sat behind //go:build cgo, so store_factory_nocgo.go's
own copies of newDoltStoreFromConfig / newReadOnlyStoreFromConfig could not
call it: the released CGO_ENABLED=0 binaries (.goreleaser.yml ships three)
still ignored shared-server in the exact GH#6551 shape, and there the
fallthrough is the "reinstall with embedded support" refusal rather than a
phantom database — actively misleading, in the one build where a shared
server is the only usable backend. Moved the helper to a new untagged file,
store_factory_server_mode.go, and called it at all four sites. The
centralization claim is now true rather than reworded, and the new file's
header records why it must stay untagged.
2. major — TestPrimeCfgNilHonorsSharedServer was vacuous. Two separate causes,
both fixed:
- the beads dir had no files at all, so hasBeadsProjectFiles was false,
FindBeadsDir() returned "" and ensureStoreActiveWithContext refused with
ErrNoBeadsDatabase before the factory was reached. It now gets the
config.yaml the reported shape actually has, which is simultaneously
discoverable and still metadata.json-less.
- it exported BEADS_DOLT_SHARED_SERVER=1, which the OLD gate already
honored, so the A/B stayed green on unfixed code even after the first
fix. The env var is now deliberately empty: the config.yaml layer is the
whole of the base gap. Renamed to TestPrimeHonorsSharedServerConfigYaml,
because the "cfg == nil" mechanism the old name asserts does not happen at
this call site — normalizeLoadedConfig substitutes DefaultConfig() before
the gate, so the gate's `cfg != nil` conjunct never short-circuited here.
The same false mechanism is corrected in TestEffectiveServerMode's comment.
Also added the positive assertion the test lacked: it now requires the
captured output to name the port it pinned, so a refusal before store
resolution can no longer pass. Confirmed by A/B against the reverted gate.
3. major — cross-workspace opens resolved activation from the launcher's
process-global config, contradicting the invariant three lines above the
gate. doltserver.IsSharedServerMode() reads process-global state, but
routed.go:214, create.go:408 and init_contributor.go:330 pass a foreign
beadsDir, so a launcher enabling shared-server in its own config.yaml could
retarget a workspace that explicitly asked for embedded — a connect failure
or a same-named database holding someone else's rows.
effectiveServerMode now takes beadsDir and resolves the yaml layer against
it via config.WorkspaceYamlValueStrictWithLocal (config.yaml only in the
first iteration; see item 4 below). The env arm stays machine-global
(configfile.IsDoltServerMode already honors it above an explicit dolt_mode,
so that is base-equivalent), and the process-wide merge is still consulted
when the target declares no dolt_mode of its own — which preserves both the
GH#6551 headline fix and base behavior, since IsDoltServerMode falls through
to config.yaml only when c.DoltMode == "". Purely workspace-scoping the read
as prescribed would have dropped the user-global layer, which viper merges,
and regressed the fix it was meant to protect.
Shared-server mode still wins over a stale metadata.json dolt_mode=embedded
for the bound workspace (main.go:1710 / shouldUseExternalDoltStatus,
GH#2946); TestEffectiveServerModeResolvesYamlPerWorkspace pins both halves
of that pair, and each half was mutation-proved to fail independently.
Non-gating, taken because they were adjacent:
- TestEffectiveServerMode's false-case assertions are now hermetic
(config.ResetForTesting + HOME/BEADS_DIR scoping), matching its sibling.
- the new assertion no longer matches on internal/storage/dolt's connection
prose; it uses errors.Is(err, syscall.ECONNREFUSED) and the port the test
itself pins, both owned outside the package this PR does not own.
Required by fix 1 rather than optional: routing the !cgo factories through
effectiveServerMode makes them consult ambient config, which reds the two
pre-existing TestNocgoNew*FromConfig flag-suggestion tests on any machine with
dolt.shared-server enabled. Both now isolate those signals first, verified by
running them with BEADS_DOLT_SHARED_SERVER=1.
Tests: TestNocgoFactoriesHonorSharedServer is new and is the missing behavioral
half of fix 1 — the cgo-tagged tests cannot observe the !cgo twins and the
pure-Go CI job only checks compilation. Each of its two arms was proved by
reverting one !cgo call site at a time.
Validated: gofmt, git diff --check, go build and go vet under both CGO_ENABLED
settings, and every touched test individually in both lanes under a private
temp root (the shared /tmp suite-root sweeper races sibling agent sessions; see
the pre-review CI note).
Iteration 2 of the maintainer review found one new major, introduced by fix 3
itself, plus four non-gating items. All five are addressed here.
4. major -- the workspace-scoped read dropped the config.local.yaml layer, so
the four factories resolved dolt.shared-server differently from main.go,
bootstrap, doctor and migrate-dolt-mode: the same resolver-divergence class
GH#6551 is an instance of, newly introduced by scoping the read. Initialize
merges .beads/config.local.yaml LAST (the file-loading block at the end of
Initialize), i.e. at the
highest priority, and it is the documented place for machine-specific
settings that must not be committed (docs/reference/configuration.md).
dolt.shared-server is exactly that kind of state while .beads/config.yaml is
git-tracked, so "the repo tracks it, this machine opts out" is the ordinary
use of the escape hatch -- this repo is itself that shape. The divergence ran
both ways and the enabling direction was the silent one: config.yaml false
(what `bd dolt shared-server off` persists) plus a local true fell through to
embeddeddolt.Open and re-created the GH#6551 phantom database with no error.
Added config.WorkspaceYamlValueStrictWithLocal beside WorkspaceYamlValueStrict
so the layering lives in the package that owns it, and pointed the helper at
it. Each layer keeps the strict contract -- absent file or absent key falls
through, unreadable/unparseable/non-scalar is an error -- which matches
Initialize, since it fails outright on a config.local.yaml it cannot merge.
config.local.yaml's name is now a config constant used by both readers, so
the merged and workspace-scoped paths cannot drift onto different files.
TestEffectiveServerModeHonorsWorkspaceLocalYaml pins both directions and uses
the merged config.GetBool as its oracle, so the assertion cannot drift from
the layer order it pins; reverting to the config.yaml-only reader fails both
arms independently. The config package's own contract test additionally pins
the order, the fallthrough when the local file is silent on the key, and the
error on a malformed local file; inverting the two layers fails it.
The prescription asked for cast.ToBool parsing. Used strconv.ParseBool
instead: spf13/cast is an indirect dependency, ParseBool is exactly what
cast.ToBool calls for a string (so "1"/"t"/"T"/"TRUE" resolve here as they do
everywhere else, which was the point), and it matches migrationSharedTopology,
the in-tree precedent for parsing this key from this reader. An unparseable
value is viper's false too, and now says so.
Corrected the parity claim the finding named: for the bound workspace these
are the same two files the process-wide merge reads, which is true only now.
Non-gating items from iteration 2, taken because the iteration was already
forced and all four live in code this fixup owns:
- the read-only factory's positive discriminator was a near-tautology: the
pinned port was "1", OR'd with errors.Is(ECONNREFUSED), so the string half
matched any errno, timestamp or port containing that digit and held up
wherever the dial error was not ECONNREFUSED. Pinned 59998 instead, five
distinctive digits like the prime sibling. Verified both halves now fire
independently, so neither is carrying the other.
- the new !cgo behavioral pin ran in no CI lane: the pure-Go cmd/bd test step is
a name allowlist and omitted it, so the exact single-call-site regression this
commit mutation-simulated would have passed every job. Added the three Nocgo
test names to that -run group in both pr.yml and main.yml. The three pass
under -tags gms_pure_go, CGO_ENABLED=0, -short in 0.4s with no server, and
reverting store_factory_nocgo.go's first call site now reds that lane --
failing only the mutated arm, so per-site discrimination is preserved. (This
also retires the "the pure-Go CI job only checks compilation" note above: that
job does run a named subset, and now it runs these.)
- the unreadable-config.yaml arm degraded silently and its comment claimed the
opposite: it said the configfile.Load error path "refuses to degrade to the
embedded default", but returning false IS the degrade to embedded here. Kept
the conservative false (main.go degrades identically on the same corrupt
input, so it is parity, and a target whose own config is unreadable must not
be retargeted by the LAUNCHER's config), added a debug.Logf so the discarded
parse error is recoverable, and corrected the comment to describe what the
code does.
- TestNocgoFactoriesHonorSharedServer's discriminator was negative-only, the
one-sidedness class this commit fixed twice elsewhere: any future pre-gate
refusal would have passed it without the shared-server arm running. Added the
positive half the siblings use, against its own distinctive pinned port.
Iteration 2 validation: gofmt, git diff --check, go build ./... and go vet under
both CGO_ENABLED settings, the full internal/config suite, the cmd/bd
store-factory and prime shared-server tests (including the 30s flagship binary
test), and the exact pure-Go CI step command with the widened allowlist. Every
fix above was mutation-proved, each arm separately.
Iteration 3 of the maintainer review returned approve with no blocker or major
finding: one minor and five nits. All but one are addressed here, amended into
this same fixup rather than stacked.
5. minor -- the layered reader honored a lone config.local.yaml, so the fixed
divergence survived at the edge of the fix. Initialize hangs its local merge
off the project config.yaml it actually found (primaryConfigPath, assigned
only where an os.Stat on a config.yaml succeeded), so a .beads holding
config.local.yaml ALONE is never merged: config.GetBool answers from defaults
while the four patched factories answered from the local file. Same
resolver-divergence class as item 4, one-directional instead of two, and loud
(a split-brain connect failure) rather than a phantom database.
Took the "require the pair" arm rather than only narrowing the prose. A
reader whose entire contract is "resolve the way Initialize's merged reader
does" should satisfy that contract instead of documenting where it does not,
and the permissive read left it the single resolver in the tree that honors a
file nothing else reads -- the divergence pointed the other way. Two lines in
WorkspaceYamlValueStrictWithLocal; a stat error other than not-exist drops the
layer too, matching Initialize, whose own probes for a project config.yaml
fail the same way, and the tracked read then surfaces that error instead of
swallowing it.
Added the missing table arm (local only, no tracked config.yaml) and mutation-
proved the guard in both directions: neutralizing it reds that arm alone, with
the divergent ("true", true) answer, while all seven incumbent arms stay green;
always dropping the local layer reds the two override arms and the malformed-
local arm while the new one stays green. The guard fires on exactly the
population it was written for.
Both parity claims now state Initialize's condition as well as its order
(store_factory_server_mode.go, yaml_config.go).
Nits, taken because amending this fixup costs nothing extra:
- the !cgo test's own comment still told the reader the pure-Go job "only checks
that the package compiles" -- in the commit that allowlisted that test by name
in both workflows. It reads as license to drop the name, which would retire
exactly the coverage iteration 2 bought. It now says the job runs this test by
name in pr.yml and main.yml, and that the test name and both allowlists must
move together.
- three new doc comments cited loadConfigFiles, which does not exist; the merge
is inline at the end of Initialize. All three now cite Initialize, as does
item 4 of this message.
- the new constants' justification claimed they keep the merged reader and the
workspace-scoped reader from drifting onto different files, but only the local
path was switched -- Initialize still spells config.yaml literally everywhere
it probes for one. Narrowed the comment to what the constants actually bind,
rather than rewriting config-path literals this PR has no other reason to move.
- the three pinned "nothing listens here" ports (59997/59998/59999) sat inside
the kernel's ephemeral range (32768-60999 here and on the GitHub runners).
That only became load-bearing because the two discriminator fixes worked: the
assertions now mean "we dialed THIS port and were refused", so a process that
bound :0 and landed there would answer the dial and red a correct
implementation. Moved all three to 19997/19998/19999. Each move is proved live
by the assertion that names its constant -- the prime test still finds its port
in the captured output.
Declined: switching migrationSharedTopology (migrate_dolt_mode.go:1238) to the
layered reader. Both lanes graded it optional and explicitly non-regressing --
migrate has always read config.yaml only -- and it is a production behavior
change to `bd migrate dolt-mode` that no further review pass would see, on an
iteration that is otherwise approve. It belongs with the five other unlayered
IsDoltServerMode resolvers iteration 2 recorded (where.go:143,
context_cmd.go:140, doctor/database.go:311, doctor/shared_store.go:45,
doctor/server.go:73) as follow-up.
Iteration 3 validation: gofmt, git diff --check, go vet in the default,
CGO_ENABLED=0/gms_pure_go and GOOS=windows modes, the full internal/config
suite, the five cmd/bd resolver and call-site tests including the 33s prime
binary test at its new port, the migrate-dolt-mode topology test, and the exact
pure-Go CI step command (ok, 0.44s).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Julian Knutsen <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>