A grant secret belongs to whoever provisions, and the sweep skips what it will not address

Issue 225. The mesh seals one credential per consumer beside the provider's
contributions file, and wrote it root-owned. That was right while a module's
own code ran in a container as root; ADR 0198 moved that code under the node's
runtime, as the node's account, and the secret stayed root's. On the control
machine two consumers went unprovisioned for three hours and the only sign
was a line reading 'secret not readable yet', 4330 times.

The same sentence is already written for a module's own secrets a few hundred
lines above — 'a root-owned 0600 file is one that process cannot read'. This
is that rule reaching the other kind of secret the mesh writes for a module.

Issue 226. The sweep met a reference recorded with the store's old address,
read 'I will not address this' as 'the store refuses everything', and
collected none of the 1681 it had found. Two changes: references from build
records are read through Recorded, where the provenance is known — not in
LetGo, which cannot tell one registry host from another and must stay strict
— and a reference the sweep will not address is now ErrNotOurs, skipped,
never a reason to stop. Only the store refusing ends a sweep.

make check: the two failures both fail on main as well — the resolver test
(hq 202/203) and the service-manager test, which reads this machine's own
shell environment.
This commit is contained in:
2026-10-04 12:21:49 +02:00
parent 912e9f4e85
commit 41b20b2782
7 changed files with 263 additions and 19 deletions
+18 -4
View File
@@ -8,6 +8,7 @@ package artifacts
import (
"context"
"errors"
"fmt"
"net/http"
"strings"
@@ -26,7 +27,15 @@ type Store struct {
}
// Gone is the answer when the store does not hold it: the outcome wanted, already true.
var Gone = fmt.Errorf("the store does not hold it")
var Gone = errors.New("the store does not hold it")
// ErrNotOurs is a reference this sweep will not address: not the mesh's own, or naming nothing
// the store holds by digest.
//
// **A fact about the record, not about the store** (novox/hq issue 226). The two deserve opposite
// responses — skip one and go on, abandon the sweep for the other — and collapsing them into "an
// error" is how a cautious loop became one that did nothing while reporting the right number.
var ErrNotOurs = errors.New("not a reference into the mesh's artifact store")
// LetGo asks the store to drop one artifact the mesh recorded making.
//
@@ -38,12 +47,17 @@ var Gone = fmt.Errorf("the store does not hold it")
// wants the artifact absent, and it is. It is distinguished from success only so a caller can say
// which of the two happened.
func (s Store) LetGo(ctx context.Context, reference string) error {
// **Strict, and deliberately** (novox/hq issue 226). Only a reference the mesh keeps in its
// own vocabulary is addressed here. `Recorded` would read `docker.io/library/registry@sha256:…`
// as the mesh's too — it cannot tell one registry host from another — so normalising belongs
// where the provenance is known, which is the sweep reading its own build records, not here
// where the only job is to refuse anything that is not plainly ours.
path, kept := catalogue.InArtifactStore(reference)
if !kept {
// Nothing the mesh put in its own store. Refused rather than attempted: composing a
// delete for a reference of unknown shape is how a sweep reaches something that is not
// the mesh's.
return fmt.Errorf("%s is not a reference into the mesh's artifact store", reference)
// the mesh's. Distinguished from a store that refuses, so a sweep skips this and goes on.
return fmt.Errorf("%w: %s", ErrNotOurs, reference)
}
if s.Address == "" {
return fmt.Errorf("this mesh has no artifact store on its network to ask about %s", reference)
@@ -95,5 +109,5 @@ func split(path string) (repository, kind, digest string, err error) {
if before, after, ok := strings.Cut(path, "/blobs/sha256:"); ok {
return before, "blobs", "sha256:" + after, nil
}
return "", "", "", fmt.Errorf("%q names nothing the store holds by digest", path)
return "", "", "", fmt.Errorf("%w: %q names nothing the store holds by digest", ErrNotOurs, path)
}
+23
View File
@@ -92,3 +92,26 @@ func TestAReferenceThatIsNotTheMeshsOwnIsNeverAsked(t *testing.T) {
t.Fatalf("the store was asked about %v", *asked)
}
}
// A reference this sweep will not address says so as ErrNotOurs, which is a fact about the
// record and not about the store (novox/hq issue 226).
//
// The sweep skips one and abandons itself for the other, so they cannot be the same error. The
// first live run met a reference recorded with the store's old address, read the refusal as "the
// store refuses everything", and collected none of the 1681 it had found.
func TestAReferenceThisSweepWillNotAddressIsToldApartFromAStoreRefusing(t *testing.T) {
store, asked := fakeStore(t, http.StatusAccepted)
for _, reference := range []string{
"docker.io/library/registry@sha256:abc123",
"127.0.0.1:5100/mesh-tools/build@sha256:abc123",
"1.4.2",
} {
err := store.LetGo(context.Background(), reference)
if !errors.Is(err, ErrNotOurs) {
t.Errorf("%s answered %v; a sweep must be able to skip it and go on", reference, err)
}
}
if len(*asked) != 0 {
t.Fatalf("the store was asked about %v", *asked)
}
}