Files
mesh-sdk/go/compose/frombus/answer.go
T
jschoubben eff81f1220
mesh/delivery merged: source.main moved to novox/mesh-sdk@b74e23f973ab
mesh/delivery-group group fix/held-module-and-assigner-contract merged: every member is merged
mesh/check checking eff81f1220e2 on b74e23f973ab | novox/mesh-sdk#142 in group fix/held-module-and-assigner-contract
Never refuse a node for a held module's stand-in by its own runtime's rules (issue 774)
On 2026-10-30 at 21:04 mesh-self-check (ace) and mesh-conditions, mesh-healers (novox) were placed before
their bus accounts were issued. Each was held for "needs a secret called broker and none was made"; its bare
stub then lost its own secrets and was refused by ownRuntimes ("runs as its own account and declares no own
secret broker"), a refusal no culprit takes, so 74 modules on ace and 79 on novox failed. A held runtime's stub
met the same kind earlier today ("its build produced 0 bundle(s)").

Quick fix: let the bare stub keep its broker own secret, for this one rule.
Thorough fix (chosen, the fourth of the kind today): a held module's stand-in (Stubbed) has no runtime
composed at all, so no rule about its runtime, bundle or broker can refuse it: ownRuntimes passes it over and
runtimeProcess composes none for a held runtime (what runs of either is carried, as before). A module of its
own account left without its runtime by a held runtime is refused with a typed NoRuntimeToRunError, which makes
the held runtime the culprit, so those modules are held after it and the node composes the rest.
2026-10-30 21:18:06 +01:00

558 lines
22 KiB
Go

package frombus
import (
"context"
"errors"
"fmt"
"maps"
"regexp"
"slices"
"sort"
"strings"
"git.novox.be/novox/mesh-sdk/go/catalogue"
"git.novox.be/novox/mesh-sdk/go/compose"
"git.novox.be/novox/mesh-sdk/go/delivery"
)
// Answer is a node's answer to its proposal's placement candidates (novox/hq ADR 0567 §3, ADR 0559 §4): s.Node composed in shadow with
// s.Placements laid over the trunk's placement (compose.CompositionsShadow, its own sealed values and boxes through r as
// for its own composition), compared with what it runs (delivery.CompareComposed). It applies nothing and writes
// nothing; the proposal and the revision are the candidate's, said back in the answer.
//
// The node's own composition is compared only where the node runs one. fps gives the build each module is composed at: what the node runs, and for a module the candidate places that the
// node does not run, the build the trunk holds; a placed module with none is added, with no build of it known. What the
// composition stands in for is judged: a value of a module the proposal adds (on this node or another: its credential, own
// secrets and boxes are made only once it is placed, ADR 0549 §5) is not yet issued, which is no refusal; any other
// stand-in is a value missing for a module the node runs, a refusal naming it ("node" when it is the node's own).
func Answer(ctx context.Context, r Reader, mesh Mesh, s Scope, fps map[string]string, running []delivery.Composed, proposal string, revision uint64) delivery.CandidateAnswer {
a := delivery.CandidateAnswer{Proposal: proposal, Revision: revision}
refuse := func(err error) delivery.CandidateAnswer {
a.Refused = err.Error()
a.Verdict = delivery.CandidateRefused
return a
}
trunk, err := r.Placements(ctx)
if err != nil {
return refuse(fmt.Errorf("frombus: placement: %w", err))
}
if s.Running == nil {
s.Running = running
}
// Every module it cannot compose is named, not the first only (issue 1126): a build it cannot read or take in, and
// below, a module the composition refuses, each composed as a stub so the rest is composed and compared.
failed := map[string]string{}
in, err := input(ctx, r, mesh, s, func(module string, err error) { failed[module] = err.Error() }, nil)
if err != nil {
return refuse(err)
}
asConsumer, asProvider, err := compose.CredentialsNeeded(in, s.Node)
if err != nil {
return refuse(err)
}
// What the node runs, by module. A consumer's credential is asked only for a module it runs: the vault refuses the
// whole ask for a consuming module not assigned to the node (mesh-vault's own), and a module the candidate adds is
// assigned only after the merge (ADR 0549 §5), so its credential is left missing, a stand-in answered as not yet
// issued under Added (review of L2b, round 1, finding 1).
runs := map[string]bool{}
for _, c := range running {
if c.Module != "" {
runs[c.Module] = true
}
}
var consumed []compose.CredentialKey
for _, k := range asConsumer {
if runs[k.ConsumerModule] {
consumed = append(consumed, k)
}
}
asConsumer = consumed
if len(asConsumer)+len(asProvider) > 0 {
if in.Credentials, in.Issued, err = r.Credentials(ctx, s.Node, asConsumer, asProvider); err != nil {
return refuse(fmt.Errorf("frombus: the credentials of %s: %w", s.Node, err))
}
}
composed, skipped, stood, err := composeEach(in, s.Node, fps, false, s.Running, failed, ranOf(ctx, r, mesh, s.Running, fps))
if err != nil {
// Not one module's: the node's own (a value served elsewhere, the mesh's network), said once as the node's.
if len(failed) == 0 {
return refuse(err)
}
failed[delivery.HaveNodeWord] = err.Error()
composed, skipped, stood = nil, nil, nil
}
// The node's own composition is judged only where the node runs one: one that runs none applies it only once a want
// names it (the node agent's own composer hands it on that want alone), so a candidate does not add it.
runsOwn := false
for _, c := range running {
runsOwn = runsOwn || c.Module == ""
}
if !runsOwn {
for key, c := range composed {
if c.Module == "" {
delete(composed, key)
}
}
}
if _, whole := failed[delivery.HaveNodeWord]; whole {
// Nothing composed to compare: every refusal is said, and nothing else.
for _, m := range sortedKeys(failed) {
a.Refusals = append(a.Refusals, delivery.ModuleRefused{Module: m, Why: failed[m]})
}
heldAsToday(&a, s.HeldToday)
a.Verdict = delivery.VerdictOf(a)
return a
}
// A module the composition folds inside the node's own is judged there: the agent also says it runs by module, from
// the declaration it kept before the switch (agent.Everything), which is no removal (issue 1126, seed #531).
if len(s.InOwn) > 0 {
var judged []delivery.Composed
for _, c := range running {
if c.Module == "" || !slices.Contains(s.InOwn, c.Module) {
judged = append(judged, c)
}
}
running = judged
}
cmp := delivery.CompareComposedWith(composed, running, s.SameValue)
// The answer rides in the node's have key: its shown bodies are bounded, the rest named only.
budget := MaxShown
for i := range cmp.Differs {
kept := cmp.Differs[i].Bodies[:0]
for _, b := range cmp.Differs[i].Bodies {
if n := len(b.Composed) + len(b.Running); n <= budget {
budget -= n
kept = append(kept, b)
} else {
kept = append(kept, delivery.ResourceBodies{ID: b.ID, Why: b.Why})
a.Note = "some differing resources are named without their bodies: the answer holds at most " +
fmt.Sprint(MaxShown) + " bytes of them"
}
}
cmp.Differs[i].Bodies = kept
}
// A refused module was composed as a stub: what it composed is no difference; its refusal is said.
notFailed := func(m string) bool { _, ok := failed[m]; return !ok }
a.Same = slices.DeleteFunc(cmp.Same, func(m string) bool { return !notFailed(m) })
a.Removed = slices.DeleteFunc(cmp.Removed, func(m string) bool { return !notFailed(m) })
a.Differs = slices.DeleteFunc(cmp.Differs, func(d delivery.ModuleDiffers) bool { return !notFailed(d.Module) })
cmp.Added = slices.DeleteFunc(cmp.Added, func(d delivery.ModuleAdded) bool { return !notFailed(d.Module) })
skipped = slices.DeleteFunc(skipped, func(m string) bool { return !notFailed(m) })
// What the proposal adds, by node and module: placed by a candidate and not by the trunk on that node.
added := map[string]bool{}
for node, p := range s.Placements {
if p == nil {
continue
}
was := map[string]bool{}
for _, m := range trunk[node].Modules {
was[m.Module] = true
}
for _, m := range p.Modules {
if !was[m.Module] {
added[node+"/"+m.Module] = true
}
}
}
unissued := map[string][]string{}
refusals := map[string][]string{}
for _, x := range stood {
node, module := standsInFor(x, s.Node)
switch {
case node == s.Node && runs[module]:
// A value missing for a module the node runs is a refusal, whatever the trunk places (the seed's trunk is
// empty, and every module counts as the group's).
refusals[module] = append(refusals[module], x)
case module != "" && added[node+"/"+module]:
if node == s.Node {
unissued[module] = append(unissued[module], x)
} // another node's addition: the module here that carries it differs, which the assigner judges
case module == "" || node != s.Node:
// The node's own, or a value this node holds for a module elsewhere (a grant it provides): said as the node's.
refusals[delivery.HaveNodeWord] = append(refusals[delivery.HaveNodeWord], x)
default:
refusals[module] = append(refusals[module], x)
}
}
addedHere := map[string]delivery.ModuleAdded{}
for _, m := range cmp.Added {
addedHere[m.Module] = m
}
for _, m := range skipped { // placed, and no build given to compose it at
addedHere[m] = delivery.ModuleAdded{Module: m, Unissued: []string{"no build of " + m + " is known to compose it at"}}
}
for m, xs := range unissued {
got := addedHere[m]
got.Module = m
got.Unissued = append(got.Unissued, xs...)
addedHere[m] = got
}
for _, m := range sortedKeys(addedHere) {
a.Added = append(a.Added, addedHere[m])
}
for _, m := range sortedKeys(refusals) {
if _, isAdded := addedHere[m]; isAdded {
continue // a module the candidate adds here: what it lacks is said under Added
}
if _, isFailed := failed[m]; isFailed {
continue // said below, as why it could not be composed
}
a.Refusals = append(a.Refusals, delivery.ModuleRefused{Module: m, Why: "stands in for " + strings.Join(refusals[m], ", ")})
}
for _, m := range sortedKeys(failed) {
a.Refusals = append(a.Refusals, delivery.ModuleRefused{Module: m, Why: failed[m]})
}
sort.Slice(a.Refusals, func(i, j int) bool { return a.Refusals[i].Module < a.Refusals[j].Module })
heldAsToday(&a, s.HeldToday)
a.Verdict = delivery.VerdictOf(a)
return a
}
// heldAsToday moves into Held each refusal of a module the node keeps today as it runs (Scope.HeldToday, with how:
// held, failed, running a build of the old form): the candidate keeps it as the node does, which is no change (issue
// 1126: a seed reproduces today, and today keeps them so). Each says how it stands today (ModuleRefused.Today). No
// module is held behind another (issue 1960): a module that reads from a held one composes with what it last offered.
func heldAsToday(a *delivery.CandidateAnswer, today map[string]string) {
if len(today) == 0 {
return
}
kept := a.Refusals[:0]
for _, r := range a.Refusals {
if w, keep := today[r.Module]; keep {
r.Today = w
a.Held = append(a.Held, r)
} else {
kept = append(kept, r)
}
}
a.Refusals = kept
if len(a.Refusals) == 0 {
a.Refusals = nil
}
}
// maxCulprits bounds the modules one refusal is tried on, each a whole composition.
const maxCulprits = 8
// MaxShown bounds the bytes of shown resource bodies in one answer (delivery.ModuleDiffers.Bodies).
const MaxShown = 256 << 10
var (
grantStandIn = regexp.MustCompile(`^the grant of \S+ for (\S+) on (\S+)$`)
needsStandIn = regexp.MustCompile(`^(\S+) on (\S+) needs `)
)
// standsInFor is the node and module a stand-in of compose's words is for: an own secret `module/name`, a port
// `port module:port`, a grant `the grant of <provision> for <module> on <node>`, a need `<module> on <node> needs ...`;
// the module empty when the words name none (the certificate, the authority).
func standsInFor(x, node string) (string, string) {
if m := grantStandIn.FindStringSubmatch(x); m != nil {
return m[2], m[1]
}
if m := needsStandIn.FindStringSubmatch(x); m != nil {
return m[2], m[1]
}
if rest, ok := strings.CutPrefix(x, "port "); ok {
if module, _, ok := strings.Cut(rest, ":"); ok {
return node, module
}
}
if module, _, ok := strings.Cut(x, "/"); ok && !strings.Contains(x, " ") {
return node, module
}
return node, ""
}
// composeEach is the one rule, for the node that applies and for its answer to a candidate alike, by which a module's
// own refusal refuses that module and never the node (issue 1126: seed-8 fixed it in the answer alone, and the
// applying path still refused all of novox for de-spiegel's unset route names). The node is composed; a refusal is laid
// on the module whose stub (compose.Stub) changes it (culprit), that module is held (compose.Node.Held) and said in
// failed with the refusal's words, and the rest is composed again, until it composes or the refusal is no single
// module's, which is answered as the node's (err). Applied composes as the node applies (compose.Reading) and holds the
// module placed here each stand-in is for; one for no module placed here (the node's own, a value this node holds for a
// module elsewhere) refuses the node (compose.ErrStoodIn). Not applied, it composes in compose.Foreseeing and answers
// every stand-in for the caller to judge (Answer).
//
// **A held module is what runs of it** (review of the apply, round 1): composed from what running holds of it
// (compose.Carried of its own composition there; the node's own carries it wherever it runs there), so what the others
// read of it (the machine ports its containers publish, the routes reaching it) is what runs, and none of its own
// composition is in out: the node keeps what it last applied.
//
// **A hold never spreads** (issue 1126, issue 1960: de-spiegel's hold for an unset value held mailu, which held
// route-proxy, nftables and postgres on novox). No module is held because a module it reads from is held: a consumer,
// a collector, the provider of what the held one contributes to and the filter where it listens each compose with what
// the held module last offered. That offer (what it contributes, contributes many, listens on) is its manifest at the
// build it runs where one is known (ran: a composition of its own running at another build than it is composed at),
// else the manifest it is composed from. Only the part of the offer that cannot compose is left out, by provision
// (culprit): de-spiegel's routes, its names unset, leave route-proxy without them while mailu keeps its smtp account,
// and the hold says so.
func composeEach(in compose.Input, node string, fps map[string]string, applied bool, running []delivery.Composed, failed map[string]string,
ran func(module string) (catalogue.Manifest, bool)) (out map[string]delivery.Composed, skipped, stood []string, err error) {
mode := compose.Foreseeing
if applied {
mode = compose.Reading
}
runs := map[string]*delivery.Composed{}
for i := range running {
if m := running[i].Module; m != "" {
runs[m] = &running[i]
}
}
src := map[string]catalogue.Manifest{} // each held module's manifest before it was held, offering what it last offered
left := map[string]map[string]bool{} // by held module, what of its offer is left out (leaveOut)
hold := func(m, why, leave string) {
if _, done := failed[m]; !done {
failed[m] = why // a held module tried again keeps the words it was first held for
}
if _, ok := src[m]; !ok {
man, ok := manifestOn(in, node, m)
if !ok {
return
}
if ran != nil {
if was, ok := ran(m); ok {
man.Contributes, man.ContributesMany, man.Listens = was.Contributes, was.ContributesMany, was.Listens
}
}
src[m] = man
}
if leave != "" {
if left[m] == nil {
left[m] = map[string]bool{}
}
left[m][leave] = true
if leave != leaveAll {
failed[m] += "; composed without what it contributes to " + leave + ", which does not compose"
}
}
in = heldAs(in, node, m, src[m], left[m], runs[m])
}
for _, m := range sortedKeys(failed) {
hold(m, failed[m], "")
}
for {
out, skipped, stood, err = compose.CompositionsIn(in, node, fps, mode)
if err != nil {
module, leave := culprit(in, node, fps, mode, failed, err)
if module == "" {
return nil, nil, nil, firstReason(err, failed)
}
hold(module, err.Error(), leave)
continue
}
again := false
if applied && len(stood) > 0 {
placed := map[string]bool{}
for _, m := range assignedOn(in, node) {
placed[m] = true
}
of := map[string][]string{}
var nodes []string
for _, x := range stood {
if n, m := standsInFor(x, node); n == node && placed[m] {
of[m] = append(of[m], x)
} else {
nodes = append(nodes, x)
}
}
for _, m := range sortedKeys(of) {
if _, done := failed[m]; !done {
hold(m, "stands in for "+strings.Join(of[m], ", "), "")
again = true
}
}
if !again && len(nodes) > 0 {
return nil, nil, nil, fmt.Errorf("%w on %s: %s", compose.ErrStoodIn, node, strings.Join(nodes, ", "))
}
}
if !again {
break
}
}
skipped = slices.DeleteFunc(skipped, func(m string) bool { _, held := failed[m]; return held })
return out, skipped, stood, nil
}
// firstReason is err, a refusal no culprit is found for, led by the first reason of each held module it names: a held
// module's stand-in refused later never replaces why it was held (issue 774, 2026-10-30: mesh-self-check on ace held for
// "has no bus credential", then refused as "runs as its own account and declares no own secret broker", its bare stub's
// words, which were all the proposal's refusal said).
func firstReason(err error, failed map[string]string) error {
var named []string
for _, m := range sortedKeys(failed) {
if strings.Contains(err.Error(), m) {
named = append(named, m+" is held: "+failed[m])
}
}
if len(named) == 0 {
return err
}
return fmt.Errorf("%s; its stand-in is then refused: %w", strings.Join(named, "; "), err)
}
// blames says the composition with m's stub (again) makes m the culprit of err: it composes, or fails otherwise. The
// runtime's stub is never a runtime (no build is taken in, so it carries no bundle) and always fails otherwise: it is
// the culprit only where its stub composes (issue 774, 2026-10-30 19:30: node-tools held for any error naming it).
func blames(m string, again, err error) bool {
if m == catalogue.RuntimeModule {
// Or where its stub fails only for the modules of their own account it would run (issue 774): they are held
// after it, each for its own words, and the node composes the rest.
var none *catalogue.NoRuntimeToRunError
return again == nil || errors.As(again, &none) && none.Held
}
return again == nil || again.Error() != err.Error()
}
// leaveAll leaves out the whole of a held module's offer and its own secrets (a bare stub); any other word leaves out
// what it contributes to that provision alone.
const leaveAll = "*"
// assignedOn is the modules placed on node, in placement order.
func assignedOn(in compose.Input, node string) []string {
for _, n := range in.Nodes {
if n.Name == node {
return n.Assigned
}
}
return nil
}
// manifestOn is the manifest node composes module from: its Kept, else the catalogue's.
func manifestOn(in compose.Input, node, module string) (catalogue.Manifest, bool) {
for _, n := range in.Nodes {
if n.Name == node {
if man, ok := n.Kept[module]; ok {
return man, true
}
}
}
man, ok := in.Catalogue[module]
return man, ok
}
// culprit is the module placed on node whose composition the error is of, and what of its offer to leave out: the
// first, in placement order, whose stub (compose.Stub) makes the composition in mode succeed or fail otherwise (leave
// ""); else the first whose stub without what it contributes to one provision does (leave that provision: de-spiegel's
// routes, its names unset, and not its smtp account, issue 1960); else the first whose bare stub does (leaveAll: its
// stub contributing nothing, its own secret never issued, issue 1126 seed-8). Empty when none does (the error is not
// one module's). Modules already refused are passed over for their stub.
func culprit(in compose.Input, node string, fps map[string]string, mode compose.Mode, failed map[string]string, err error) (string, string) {
assigned := assignedOn(in, node)
// Only the modules the error names: the words of a module's refusal name it. Each try is a whole composition, so a
// refusal naming none is no module's, answered at once rather than after a composition per module (issue 1126: ace's
// answer took 7m46s).
var order []string
for _, m := range assigned {
if strings.Contains(err.Error(), m) && len(order) < maxCulprits {
order = append(order, m)
}
}
changes := func(m, leave string) bool {
_, _, _, again := compose.CompositionsIn(stubbed(in, node, m, leave), node, fps, mode)
return blames(m, again, err)
}
for _, m := range order {
if _, done := failed[m]; done {
continue
}
if changes(m, "") {
return m, ""
}
}
for _, m := range order {
man, ok := manifestOn(in, node, m)
if !ok {
continue
}
// One already held is tried again too: held for its build, its stub may still refuse (its own secret never
// issued, issue 1126). A provision named in the error is tried first, each at one more composition.
var to []string
for p := range man.Contributes {
to = append(to, p)
}
for p := range man.ContributesMany {
if _, both := man.Contributes[p]; !both {
to = append(to, p)
}
}
sort.SliceStable(to, func(i, j int) bool {
return strings.Contains(err.Error(), " "+to[i]) && !strings.Contains(err.Error(), " "+to[j])
})
for _, p := range to {
if changes(m, p) {
return m, p
}
}
if isBare(man) {
continue
}
if changes(m, leaveAll) {
return m, leaveAll
}
}
return "", ""
}
// isBare says a stub contributes nothing and declares no own secret (stubbed, leaveAll).
func isBare(m catalogue.Manifest) bool {
return len(m.Contributes) == 0 && len(m.ContributesMany) == 0 && len(m.OwnSecrets) == 0
}
// stubbed is in with module held on node as its stub alone (heldAs, nothing of it running), from the manifest node
// composes it from, leave left out of its offer: a culprit's try.
func stubbed(in compose.Input, node, module, leave string) compose.Input {
man, ok := manifestOn(in, node, module)
if !ok {
return in
}
var left map[string]bool
if leave != "" {
left = map[string]bool{leave: true}
}
return heldAs(in, node, module, man, left, nil)
}
// heldAs is in with module held on node (compose.Node.Held): composed from man as what runs of it (compose.Carried of
// runs; its stub, compose.Stub, when runs is nil or holds nothing of it), as node's Kept; with what left names out of
// its offer: a provision's contributions, or with leaveAll everything it contributes and its own secrets, none of a
// held module's own composition being applied, so what it has not been issued (restic's repository on a node it never
// ran on) is no refusal of the rest.
func heldAs(in compose.Input, node, module string, man catalogue.Manifest, left map[string]bool, runs *delivery.Composed) compose.Input {
out := in
out.Nodes = append([]compose.Node(nil), in.Nodes...)
for i := range out.Nodes {
if out.Nodes[i].Name != node {
continue
}
if left[leaveAll] {
man.Contributes, man.ContributesMany, man.OwnSecrets = nil, nil, nil
} else if len(left) > 0 {
contributes := map[string]map[string]any{}
for p, v := range man.Contributes {
if !left[p] {
contributes[p] = v
}
}
many := maps.Clone(man.ContributesMany)
for p := range left {
delete(many, p)
}
man.Contributes, man.ContributesMany = contributes, many
}
kept := make(map[string]catalogue.Manifest, len(out.Nodes[i].Kept)+1)
for k, v := range out.Nodes[i].Kept {
kept[k] = v
}
kept[module] = compose.Carried(man, module, runs)
out.Nodes[i].Kept = kept
held := make(map[string]bool, len(out.Nodes[i].Held)+1)
for k, v := range out.Nodes[i].Held {
held[k] = v
}
held[module] = true
out.Nodes[i].Held = held
}
return out
}