diff --git a/cmd/mesh-controller/gate.go b/cmd/mesh-controller/gate.go index 304186a2..8a1e1a1c 100644 --- a/cmd/mesh-controller/gate.go +++ b/cmd/mesh-controller/gate.go @@ -1443,6 +1443,12 @@ func probeGates(ctx context.Context, d *doctor) ([]conditions.Observation, error } // A release held after a failed one, while builds still wait for a gate (ADR 0236). out = append(out, backlogObservation()...) + // And a rest that went silent (novox/hq ADR 0353 §4). + silent, err := silentObservation(ctx, d.open.inventory) + if err != nil { + return nil, err + } + out = append(out, silent...) return sortedFound(dedupeObservations(out)), nil } @@ -1532,6 +1538,12 @@ func gateLine(g *inventory.PlanGate) string { case inventory.GateFailed: what = "FAILED" } + switch { + case g.Verdict == "" && g.Pending: + what = "kept, not yet sent" // novox/hq ADR 0353 §8 + case g.Verdict == "" && g.Silent: + what = "silent on its send, left" // ADR 0353 §4 + } if g.Verdict == "" && g.Waiting() { what = "waits for a busy container on " + g.Busy.Machine } diff --git a/cmd/mesh-controller/held_back.go b/cmd/mesh-controller/held_back.go index 0d90cc19..a0b4c6ba 100644 --- a/cmd/mesh-controller/held_back.go +++ b/cmd/mesh-controller/held_back.go @@ -151,7 +151,7 @@ func heldMachines(ctx context.Context, open *stores, names []string) (map[string for _, mv := range f.moves(node, modules, same, known, false) { if planStillToSend(plans, mv.Module, node) == "" { why = append(why, fmt.Sprintf("%s would move from %s to %s, which has passed no gate yet — a release "+ - "plan sends it, one machine at a time, judged", mv.Module, buildName(mv.From), buildName(mv.To))) + "plan sends it, one machine first and then the rest at once, judged", mv.Module, buildName(mv.From), buildName(mv.To))) } } if len(why) > 0 { diff --git a/cmd/mesh-controller/plain_words.go b/cmd/mesh-controller/plain_words.go index a5f705b1..7d60b16a 100644 --- a/cmd/mesh-controller/plain_words.go +++ b/cmd/mesh-controller/plain_words.go @@ -456,6 +456,9 @@ var plainWordings = map[string]func(conditions.Observation) words{ "release-held": worded(func(o conditions.Observation) words { return releaseHeldWords(nil, o.Also) }), + "release-silent": worded(func(o conditions.Observation) words { + return releaseSilentWords(o.Also) + }), "facts-stale": worded(func(o conditions.Observation) words { return words{Headline: "Merge checks use outdated facts", Explanation: "The facts a merge check judges a change against are out of date; they are gathered again " + @@ -841,6 +844,19 @@ func releaseHeldWords(modules, machines []string) words { Resolved: "Resolved: the held updates are released"} } +// releaseSilentWords are the plain words of machines that said nothing on an update they were sent (ADR 0353 §4). +func releaseSilentWords(machines []string) words { + where := "Some machines" + if len(machines) > 0 { + where = conditions.Capital(namesWords(machines, 4)) + } + return words{Headline: "Machines did not answer an update", + Needs: "check that they are awake and on the mesh; the next walk sends them what waits.", + Explanation: where + " were sent module updates and said nothing about them within fifteen minutes, so " + + "they were left as they were. What they run is not known to be the update.", + Resolved: "Resolved: the next walk reached them"} +} + // reloginNeeds is what an account waiting for its groups needs (ADR 0252). func reloginNeeds(node string) string { return fmt.Sprintf("log out of %s completely and log in again, or restart it.", node) diff --git a/cmd/mesh-controller/release.go b/cmd/mesh-controller/release.go index ca9c0bd4..37d68311 100644 --- a/cmd/mesh-controller/release.go +++ b/cmd/mesh-controller/release.go @@ -39,8 +39,8 @@ import ( // from the same manifest is no move. // // **The release plan** is what walks the waiting moves through: whenever moves wait for a gate that no -// started plan is walking, the mesh opens one — every such machine, one at a time, the control node last, -// each sent and judged before the next. A release plan that fails stops the next one opening on its own: +// started plan is walking, the mesh opens one — every such machine, the first sent and judged, then the rest +// at once, each judged by its own gate (novox/hq ADR 0353, in place of one at a time). A release plan that fails stops the next one opening on its own: // a person releases it again (`upgrade release-backlog --why`), after the condition says why. // sendScope is what a send may carry that no gate has seen. @@ -81,6 +81,15 @@ type moveFacts struct { srcs map[string]map[string]string passed map[string]map[string]bool plans []inventory.Plan + // self is the release walking these moves, whose own gates are no other walk's (novox/hq ADR 0353). + self string +} + +type walkingKey struct{} + +// withWalking says a release plan is the one sending: its own gates never hold its own sends back. +func withWalking(ctx context.Context, id string) context.Context { + return context.WithValue(ctx, walkingKey{}, id) } func readMoveFacts(ctx context.Context, inv *inventory.Inventory) (moveFacts, error) { @@ -99,6 +108,7 @@ func readMoveFacts(ctx context.Context, inv *inventory.Inventory) (moveFacts, er return f, err } f.plans, err = inv.OpenPlans(ctx) + f.self, _ = ctx.Value(walkingKey{}).(string) return f, err } @@ -193,17 +203,24 @@ func (f moveFacts) moves(node string, modules []string, sent map[string]string, // feet. A release keeps no module records: its open gate's carried moves are its walk. func (f moveFacts) walkedBy(module, node, to string) string { for _, p := range f.plans { - if !p.Open() { + if !p.Open() || (p.Release != nil && p.ID == f.self) { continue } if p.Release != nil { - g := p.Release.Gate - if g == nil || g.Verdict != "" { - continue + // The first machine's gate, and every gate of the rest still judging (novox/hq ADR 0353). + gates := []*inventory.PlanGate{p.Release.Gate} + for _, g := range p.Release.Gates { + gates = append(gates, g) } - for _, c := range g.Carried { - if c.Module == module && !(slices.Contains(g.Machines, node) && sameCommit(c.To, to)) { - return p.ID + for _, g := range gates { + // A send kept and not yet made walks nothing yet (ADR 0353 §8). + if g == nil || g.Verdict != "" || g.Silent || g.Pending { + continue + } + for _, c := range g.Carried { + if c.Module == module && !(slices.Contains(g.Machines, node) && sameCommit(c.To, to)) { + return p.ID + } } } continue @@ -303,8 +320,8 @@ func ungatedIn(ctx context.Context, open *stores, names []string, addedHolder st } } if len(refused) > 0 { - return nil, fmt.Errorf("%w: %s — a walk sends them, one machine at a time, each judged (novox/hq ADR "+ - "0236); `upgrade backlog` lists them", errUngated, strings.Join(refused, "; ")) + return nil, fmt.Errorf("%w: %s — a walk sends them, one machine first and then the rest at once, each judged (novox/hq ADR "+ + "0236, ADR 0353); `upgrade backlog` lists them", errUngated, strings.Join(refused, "; ")) } return kept, nil } @@ -320,19 +337,35 @@ func ungatedIn(ctx context.Context, open *stores, names []string, addedHolder st // And the grants step first, when the send changes what the bus's user list must say (novox/hq issue 490): // answered for the gate to keep, nil when there was none. func gatedSend(ctx context.Context, open *stores, node string, owns []inventory.CarriedMove) ([]inventory.CarriedMove, []string, *inventory.GrantsStep, error) { - inv := open.inventory - f, err := readMoveFacts(ctx, inv) + moves, err := gatedMoves(ctx, open, node, owns) + if err != nil || len(moves) == 0 { + return nil, nil, nil, err + } + sayRecreations(ctx, open, moves) + sent, grants, err := sendJudged(ctx, open, node) if err != nil { return nil, nil, nil, err } + return moves, sent, grants, nil +} + +// gatedMoves is what a gated send to one machine would carry, with the build each moves to, made of nothing: +// gatedSend's reading, which a walk also keeps before it sends (novox/hq ADR 0353 §8). Refused the same way +// while a plan that has started walks one of those builds elsewhere. +func gatedMoves(ctx context.Context, open *stores, node string, owns []inventory.CarriedMove) ([]inventory.CarriedMove, error) { + inv := open.inventory + f, err := readMoveFacts(ctx, inv) + if err != nil { + return nil, err + } moves, err := machineMoves(ctx, open, f, node, true) if err != nil { - return nil, nil, nil, err + return nil, err } // A move the send keeps is not made, and is no move for the gate to judge (novox/hq issue 201): a core module // kept over an older record judged as moving back to it would read the gate as superseded by its own send. if moves, err = withoutKept(ctx, open, node, moves); err != nil { - return nil, nil, nil, err + return nil, err } own := func(module string) bool { return slices.ContainsFunc(owns, func(o inventory.CarriedMove) bool { return o.Module == module }) @@ -342,7 +375,7 @@ func gatedSend(ctx context.Context, open *stores, node string, owns []inventory. continue } if id := f.walkedBy(mv.Module, node, mv.To); id != "" { - return nil, nil, nil, fmt.Errorf("%w: %s's build %s waits on %s, which %s is walking", errWalkedElsewhere, + return nil, fmt.Errorf("%w: %s's build %s waits on %s, which %s is walking", errWalkedElsewhere, mv.Module, short(mv.To), node, id) } } @@ -350,7 +383,7 @@ func gatedSend(ctx context.Context, open *stores, node string, owns []inventory. // plan, or a release, on this machine is not crossed with a newer build; this send waits for its gate. for _, o := range owns { if id := f.walkedBy(o.Module, node, o.To); id != "" { - return nil, nil, nil, fmt.Errorf("%w: %s's build %s waits on %s, which %s is walking", errWalkedElsewhere, + return nil, fmt.Errorf("%w: %s's build %s waits on %s, which %s is walking", errWalkedElsewhere, o.Module, short(o.To), node, id) } } @@ -363,30 +396,26 @@ func gatedSend(ctx context.Context, open *stores, node string, owns []inventory. moves[i].Build = o.Build } } - if len(moves) == 0 && len(owns) == 0 { - return nil, nil, nil, nil + if len(moves) == 0 { + return nil, nil } for i := range moves { if moves[i].Build == "" { if moves[i].Build, err = inv.BuildOf(ctx, moves[i].Module, moves[i].To); err != nil { - return nil, nil, nil, err + return nil, err } } } sort.Slice(moves, func(i, j int) bool { return moves[i].Module < moves[j].Module }) - sayRecreations(ctx, open, moves) - sent, grants, err := sendJudged(ctx, open, node) - if err != nil { - return nil, nil, nil, err - } - return moves, sent, grants, nil + return moves, nil } // passCarried keeps a pass as the verdict of every build the gate judged beside its own module. func passCarried(ctx context.Context, open *stores, p *inventory.Plan, g *inventory.PlanGate, except string) { seen := map[string]bool{} for _, c := range g.Carried { - if c.Module == except || c.Build == "" || seen[c.Build] { + // A build put back with another machine's failure is not this gate's to pass (novox/hq ADR 0353 §3). + if c.Module == except || c.Build == "" || seen[c.Build] || slices.Contains(g.Returned, c.Module) { continue } seen[c.Build] = true @@ -695,7 +724,7 @@ func releaseBacklog(ctx context.Context, open *stores, by string) (*inventory.Pl p := inventory.Plan{ID: fmt.Sprintf("release-%d", now.UnixNano()), Repository: releaseRepository, Created: now, State: inventory.PlanRolling, Tiers: [][]string{names}, Modules: map[string]*inventory.PlanModule{}, Release: &inventory.PlanRelease{Order: order, By: by}, - Note: fmt.Sprintf("%d build(s) wait for a gate on %s; one machine at a time, each judged", len(names), + Note: fmt.Sprintf("%d build(s) wait for a gate on %s; one machine first, then the rest at once, each judged", len(names), strings.Join(order, ", "))} if err := inv.SavePlan(ctx, &p); err != nil { return nil, err @@ -704,14 +733,23 @@ func releaseBacklog(ctx context.Context, open *stores, by string) (*inventory.Pl return &p, nil } -// advanceRelease takes one step of a release plan: the next machine sent everything waiting there under a -// gate, or the machine being judged judged once more; the plan done when every machine is. +// advanceRelease takes one step of a release plan: the first machine sent everything waiting there under a +// gate and judged; once it passed, every other machine sent at once, each under its own gate (novox/hq ADR +// 0353 §1), and judged; the plan done when every machine is. +// +// **A send is kept before it is made** (ADR 0353 §8): the step that chooses a machine keeps its gate pending with +// what it will carry, and the next step makes the send. A controller replaced between the send and keeping it +// finds the gate pending, reads the machine's last send, and neither sends again nor leaves the machine unjudged. func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, error) { r := p.Release if r.Settings != nil { return advanceSettingsWalk(ctx, open, p) } + ctx = withWalking(ctx, p.ID) now := time.Now().UTC() + if r.Rest { + return advanceRest(ctx, open, p, now) + } if r.Gate == nil { heard, err := releaseHeard(ctx, open) if err != nil { @@ -724,7 +762,7 @@ func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, r.Next++ continue } - moves, sent, grants, err := gatedSend(ctx, open, node, nil) + moves, err := gatedMoves(ctx, open, node, nil) if errors.Is(err, errWalkedElsewhere) { note := fmt.Sprintf("waiting before %s: %v", node, err) changed := p.Note != note @@ -732,23 +770,15 @@ func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, return changed, nil } if err != nil { - return false, fmt.Errorf("sending %s what waits there: %w", node, err) + return false, fmt.Errorf("reading what waits on %s: %w", node, err) } if len(moves) == 0 { r.Done = append(r.Done, node) r.Next++ continue } - r.Gate = &inventory.PlanGate{Machines: sent, Since: &now, Carried: moves, Sent: sentNow(ctx, open.inventory, sent), - Grants: grants} - p.Note = fmt.Sprintf("sent %s %d build(s) that waited for a gate; judging them there", node, len(moves)) - if said := grantsSaid(grants); said != "" { - p.Note += "; " + said - } - if said := recreationsSaid(moves); said != "" { - p.Note += "; " + said - } - fmt.Printf("%s: %s\n", p.ID, p.Note) + r.Gate = &inventory.PlanGate{Machines: []string{node}, Carried: moves, Pending: true} + p.Note = fmt.Sprintf("sending %s %d build(s) that wait for a gate, first", node, len(moves)) return true, nil } p.State, p.Tier = inventory.PlanDone, len(p.Tiers) @@ -759,6 +789,19 @@ func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, return true, nil } g := r.Gate + if g.Pending { + node := firstOf(g.Machines) + out, err := sendKept(ctx, open, p, node, g, now) + if err != nil || out == keptWaiting { + return out == keptWaiting, err + } + if out == keptNothing { + r.Gate = nil + r.Done = append(r.Done, node) + r.Next++ + } + return true, nil + } verdict, err := judgeMoves(ctx, open, g, nil, now) if err != nil { return false, err @@ -777,6 +820,11 @@ func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, r.Done = append(r.Done, firstOf(g.Machines)) r.Next++ r.Gate = nil + // **Then the rest at once** (novox/hq ADR 0353 §1): every other machine heard from, each judged by its own + // gate, in place of one machine at a time (ADR 0236 §22). + r.Rest = true + p.Note = fmt.Sprintf("passed on %s; the rest are sent at once, each judged", firstOf(g.Machines)) + fmt.Printf("%s: %s\n", p.ID, p.Note) return true, nil case inventory.GateSuperseded: // Another send moved a judged module on the judged machine (novox/hq issue 352): no verdict on what @@ -797,6 +845,453 @@ func advanceRelease(ctx context.Context, open *stores, p *inventory.Plan) (bool, return true, nil } +// What sendKept did with a kept send. +const ( + keptSent = iota // sent now + keptAlready // the machine was already sent it: judged from that send, nothing sent + keptNothing // nothing waits there any more: no gate + keptWaiting // another walk judges a build it would carry: it waits +) + +// sendKept makes a send a gate kept before it was made (novox/hq ADR 0353 §8) — unless the machine's last send +// already carries what the gate is for, made before a controller was replaced and before it was kept: then +// nothing is sent, and the gate judges from that send (§6). +func sendKept(ctx context.Context, open *stores, p *inventory.Plan, node string, g *inventory.PlanGate, + now time.Time) (int, error) { + inv := open.inventory + moves, sent, grants, err := gatedSend(ctx, open, node, nil) + if errors.Is(err, errWalkedElsewhere) { + p.Note = fmt.Sprintf("waiting before %s: %v", node, err) + return keptWaiting, nil + } + if err != nil { + return keptSent, fmt.Errorf("sending %s what waits there: %w", node, err) + } + if len(moves) > 0 { + g.Machines, g.Carried, g.Since, g.Grants, g.Pending = sent, moves, &now, grants, false + g.Sent = sentNow(ctx, inv, sent) + p.Note = fmt.Sprintf("sent %s %d build(s) that waited for a gate; judging them there", node, len(moves)) + if said := grantsSaid(grants); said != "" { + p.Note += "; " + said + } + if said := recreationsSaid(moves); said != "" { + p.Note += "; " + said + } + fmt.Printf("%s: %s\n", p.ID, p.Note) + return keptSent, nil + } + if !alreadySent(ctx, inv, node, g.Carried) { + return keptNothing, nil + } + since := now + if s, found, err := inv.SentTo(ctx, node); err == nil && found { + if !s.At.IsZero() { + since = s.At + } + g.Sent = map[string]inventory.SentDeclaration{node: s} + } + g.Since, g.Pending = &since, false + p.Note = fmt.Sprintf("%s was already sent the %d build(s) kept for it, at %s: judged from that send, nothing sent again", + node, len(g.Carried), since.Local().Format("15:04:05")) + fmt.Printf("%s: %s\n", p.ID, p.Note) + return keptAlready, nil +} + +// alreadySent says a machine's last send carries every build of the moves: a send made and not kept. +func alreadySent(ctx context.Context, inv *inventory.Inventory, node string, moves []inventory.CarriedMove) bool { + if len(moves) == 0 { + return false + } + sent, known, err := inv.SentBuilds(ctx, node) + if err != nil || !known { + return false + } + for _, mv := range moves { + if !sameCommit(sent[mv.Module], mv.To) { + return false + } + } + return true +} + +// advanceRest takes one step of a release plan's rest (novox/hq ADR 0353): the sends kept for a step made, every +// machine of the step judged by its own gate, one that fails put back by itself while the others go on to their +// verdicts, and the next step — or the end — once every gate of the step has its verdict or is silent. +func advanceRest(ctx context.Context, open *stores, p *inventory.Plan, now time.Time) (bool, error) { + r := p.Release + // The sends kept for this step, in the plan's order: the control node, last in it, last (§1). + var pending []string + for _, n := range r.Order { + if g := r.Gates[n]; g != nil && g.Pending { + pending = append(pending, n) + } + } + if len(pending) > 0 { + // **A kept send carries only what was kept** (ADR 0353 §2): a build that passed no gate and came to wait on + // a machine since its step was chosen could reach a second machine in it; the step is chosen again. + if grew, err := keptStepGrew(ctx, open, r, pending); err != nil { + return false, err + } else if grew != "" { + for _, n := range pending { + delete(r.Gates, n) + } + opened, waiting, err := openRestStep(ctx, open, p) + if err != nil { + return true, err + } + p.Note = "chosen again: " + grew + if !opened && waiting != "" { + p.Note += "; " + waiting + } + return true, nil + } + var sent, already, waiting []string + for _, node := range pending { + out, err := sendKept(ctx, open, p, node, r.Gates[node], now) + if err != nil { + return true, err + } + switch out { + case keptSent: + sent = append(sent, node) + case keptAlready: + already = append(already, node) + case keptNothing: + delete(r.Gates, node) + r.Done = append(r.Done, node) + case keptWaiting: + waiting = append(waiting, node) + } + } + var said []string + if len(sent) > 0 { + said = append(said, "sent "+strings.Join(sent, ", ")+" at once, each judged there") + } + if len(already) > 0 { + said = append(said, strings.Join(already, ", ")+" already sent it: judged from that send") + } + if len(waiting) > 0 { + said = append(said, "waiting before "+strings.Join(waiting, ", ")+": another walk judges a build it carries") + } + if len(said) > 0 { + p.Note = strings.Join(said, "; ") + } + return len(sent)+len(already) > 0 || len(waiting) < len(pending), nil + } + reports, err := open.inventory.LastReports(ctx) + if err != nil { + return false, err + } + byNode := map[string]inventory.Reported{} + for _, rep := range reports { + byNode[rep.Node] = rep + } + changed, noteWas := false, p.Note + var notes, judging []string + batched, back := batchingRollbacks(ctx) + for _, node := range r.Order { + g := r.Gates[node] + if g == nil || g.Verdict != "" || g.Silent { + continue + } + // What another machine's failure already put back is not this gate's to judge (ADR 0353 §3). + if len(g.Carried) > 0 && len(g.Returned) >= len(g.Carried) { + decide(g, inventory.GateFailed, "what it was sent was put back with another machine's failure", now) + r.Failed = append(r.Failed, node) + changed = true + continue + } + // **A machine that says nothing on its send is left silent, not failed** (ADR 0353 §4): it is not judged + // until it reports, and past ApplySilentAfter it is listed and left. + if !reportedOnSend(byNode[node], node, g) { + if now.Sub(*g.Since) > inventory.ApplySilentAfter { + g.Silent = true + r.Silent = append(r.Silent, node) + changed = true + notes = append(notes, fmt.Sprintf("%s said nothing on its send within %s: left silent", node, + inventory.ApplySilentAfter)) + continue + } + judging = append(judging, node+": not yet reported on its send") + continue + } + verdict, err := judgeMoves(ctx, open, g, nil, now) + if err != nil { + return changed, err + } + switch verdict { + case "": + if putBackBroken(ctx, open, p, g, "", nil) { + changed = true + notes = append(notes, p.Note) + } + judging = append(judging, node+": "+gateLine(g)) + case inventory.GatePassed: + passCarried(ctx, open, p, g, "") + r.Done = append(r.Done, node) + changed = true + case inventory.GateSuperseded: + // Another send moved a judged module there (novox/hq issue 352): nothing judged, nothing put back on + // it; what still waits is released again. + r.Done = append(r.Done, node) + changed = true + notes = append(notes, fmt.Sprintf("superseded on %s: %s", node, g.Why)) + default: + // **Put back by itself** (ADR 0353 §3): what was found wanting goes back at once, and every other + // machine of the step goes on to its own verdict. + stateWas := p.State + p.Note = "" + failCarried(batched, open, p, g, "") + p.State = stateWas + r.Failed = append(r.Failed, node) + changed = true + notes = append(notes, fmt.Sprintf("failed its gate on %s: %s — %s", node, g.Why, p.Note)) + returnedElsewhere(ctx, open, r, node, g) + } + } + p.Note = "" + sendRollbacks(ctx, open, p, back) + if p.Note != "" { + notes = append(notes, strings.TrimPrefix(p.Note, "; ")) + } + if len(judging) > 0 { + note := strings.Join(append(notes, "judging "+strings.Join(judging, "; ")), "; ") + p.Note = note + return changed || noteWas != note, nil + } + // Every gate of the step has its verdict or is silent: the next step, unless a machine failed (§3). + if len(r.Failed) == 0 { + opened, waiting, err := openRestStep(ctx, open, p) + if err != nil { + return changed, err + } + if opened { + return true, nil + } + if waiting != "" { + note := strings.Join(append(notes, waiting), "; ") + p.Note = note + return changed || noteWas != note, nil + } + } + p.Tier = len(p.Tiers) + end := fmt.Sprintf("released on %s", orNone(strings.Join(r.Done, ", "))) + if len(r.Silent) > 0 { + end += "; silent on their send, left: " + strings.Join(r.Silent, ", ") + } + if len(r.Skipped) > 0 { + end += "; not heard from, left as they were: " + strings.Join(r.Skipped, ", ") + } + p.State = inventory.PlanDone + if len(r.Failed) > 0 { + p.State = inventory.PlanFailed + end = "failed on " + strings.Join(r.Failed, ", ") + "; " + end + } else if len(r.Silent) > 0 && len(r.Gates) == len(r.Silent) { + // **A rest that said nothing is no success** (ADR 0353 §4): failure is loud; the condition says it. + p.State = inventory.PlanFailed + end = "no machine of the rest reported on its send; " + end + } + p.Note = strings.Join(append(notes, end), "; ") + fmt.Printf("%s: %s\n", p.ID, p.Note) + return true, nil +} + +// keptStepGrew says why a step's kept sends would now carry a build no gate passed that they were not kept with; +// empty when none would. +func keptStepGrew(ctx context.Context, open *stores, r *inventory.PlanRelease, pending []string) (string, error) { + f, err := readMoveFacts(ctx, open.inventory) + if err != nil { + return "", err + } + shelf, err := open.inventory.Catalogue(ctx) + if err != nil { + return "", err + } + for _, node := range pending { + moves, err := gatedMoves(ctx, open, node, nil) + if errors.Is(err, errWalkedElsewhere) { + continue // sendKept says it waits + } + if err != nil { + return "", err + } + kept := r.Gates[node].Carried + for _, mv := range moves { + // The node-engine and the resolver keep one machine at a time, passed or not (ADR 0353 §9). + if (f.gated(mv.Module, mv.To) && !oneAtATime(mv.Module, shelf[mv.Module])) || slices.ContainsFunc(kept, func(c inventory.CarriedMove) bool { + return c.Module == mv.Module && sameCommit(c.To, mv.To) + }) { + continue + } + return fmt.Sprintf("%s's build %s, which this step did not keep for it, came to wait on %s", mv.Module, short(mv.To), node), nil + } + } + return "", nil +} + +// reportedOnSend says a machine's last report is about the send its gate judges (novox/hq issue 352): the +// declaration kept on the gate, or one sequenced after it; a gate that kept none reads the send made last. +func reportedOnSend(r inventory.Reported, node string, g *inventory.PlanGate) bool { + if r.At == nil { + return false + } + if s, kept := g.Sent[node]; kept { + return s.ReportsOn(r) + } + return r.Current +} + +// returnedElsewhere marks, on every other gate of the rest still judging, the builds a failed machine's +// rollback put back wherever they were sent (ADR 0236 §5): no longer theirs to judge (ADR 0353 §3). +func returnedElsewhere(ctx context.Context, open *stores, r *inventory.PlanRelease, failed string, g *inventory.PlanGate) { + for _, c := range g.Carried { + if c.Build == "" { + continue + } + if marked, err := open.inventory.GateFailed(ctx, c.Build); err != nil || !marked { + continue + } + for node, other := range r.Gates { + if node == failed || other.Verdict != "" || slices.Contains(other.Returned, c.Module) { + continue + } + for _, oc := range other.Carried { + if oc.Module == c.Module && sameCommit(oc.To, c.To) { + other.Returned = append(other.Returned, c.Module) + } + } + } + } +} + +// openRestStep chooses the next step of a release plan's rest and keeps its sends (novox/hq ADR 0353 §1, §2): +// every machine left that is heard from and has something waiting, except that a build no gate has passed goes +// to one machine of the step only. Answers whether a step was kept, or why the walk waits. +func openRestStep(ctx context.Context, open *stores, p *inventory.Plan) (bool, string, error) { + r := p.Release + heard, err := releaseHeard(ctx, open) + if err != nil { + return false, "", err + } + f, err := readMoveFacts(ctx, open.inventory) + if err != nil { + return false, "", err + } + waiting := map[string][]inventory.CarriedMove{} + var nodes, walked []string + for _, node := range r.Order[min(r.Next, len(r.Order)):] { + if r.Gates[node] != nil || slices.Contains(r.Done, node) || slices.Contains(r.Skipped, node) || + slices.Contains(r.Failed, node) || slices.Contains(r.Silent, node) { + continue + } + if !heard[node] { + r.Skipped = append(r.Skipped, node) + continue + } + moves, err := gatedMoves(ctx, open, node, nil) + if errors.Is(err, errWalkedElsewhere) { + walked = append(walked, fmt.Sprintf("%s (%v)", node, err)) + continue + } + if err != nil { + return false, "", fmt.Errorf("reading what waits on %s: %w", node, err) + } + if len(moves) == 0 { + r.Done = append(r.Done, node) + continue + } + waiting[node] = moves + nodes = append(nodes, node) + } + shelf, err := open.inventory.Catalogue(ctx) + if err != nil { + return false, "", err + } + step := restStep(nodes, waiting, f.gated, func(module string) bool { return oneAtATime(module, shelf[module]) }) + if len(step) == 0 { + if len(walked) > 0 { + return false, "waiting before " + strings.Join(walked, "; "), nil + } + return false, "", nil + } + if r.Gates == nil { + r.Gates = map[string]*inventory.PlanGate{} + } + for _, node := range step { + r.Gates[node] = &inventory.PlanGate{Machines: []string{node}, Carried: waiting[node], Pending: true} + } + p.Note = fmt.Sprintf("sending %s at once, each judged there", strings.Join(step, ", ")) + if left := len(nodes) - len(step); left > 0 { + p.Note += fmt.Sprintf("; %d machine(s) wait for a build that passed no gate yet to pass on one", left) + } + return true, "", nil +} + +// oneAtATime is a module whose builds go to the rest one machine at a time (novox/hq ADR 0353 §9): the +// node-engine, which applies every send, and a module holding the mesh's resolver, which names the bus. +func oneAtATime(module string, m catalogue.Manifest) bool { + if module == hostModule { + return true + } + return slices.ContainsFunc(m.Claims, func(c catalogue.Claim) bool { return c.Name == resolverSeat }) +} + +// restStep is the machines of one step of the rest (novox/hq ADR 0353 §2), in the order given: each machine, unless +// a build it would carry has passed no gate and an earlier machine of the step already carries it — that build's +// first machine. Pure. +// +// **The node-engine and the mesh's resolver still go one machine at a time** (ADR 0353 §9): a step carries a +// build of either to one machine only, judged before the next, since a fault that cuts a machine off the bus +// shows only as its silence and no rollback reaches it. +func restStep(nodes []string, waiting map[string][]inventory.CarriedMove, passed func(module, commit string) bool, + alone func(module string) bool) []string { + first := map[string]string{} + var out []string + for _, n := range nodes { + var unpassed []string + for _, mv := range waiting[n] { + if !passed(mv.Module, mv.To) || alone(mv.Module) { + unpassed = append(unpassed, mv.Module+"@"+mv.To) + } + } + if slices.ContainsFunc(unpassed, func(k string) bool { _, taken := first[k]; return taken }) { + continue + } + for _, k := range unpassed { + first[k] = n + } + out = append(out, n) + } + return out +} + +// silentObservation is what the gate's probe says of the newest release whose rest went silent (novox/hq ADR 0353 +// §4): machines that never reported on their send, left as they were — loud, never a success said by nothing. +func silentObservation(ctx context.Context, inv *inventory.Inventory) ([]conditions.Observation, error) { + recent, err := inv.RecentPlans(ctx, 20) + if err != nil { + return nil, err + } + for _, p := range recent { + if p.Release == nil || p.Release.Settings != nil || p.Open() { + continue + } + if len(p.Release.Silent) == 0 { + return nil, nil + } + machines := append([]string(nil), p.Release.Silent...) + sort.Strings(machines) + o := conditions.Observation{Scope: conditions.ScopeMesh, ID: "release", Token: "silent", Kind: "release-silent", + Severity: conditions.Warning, Resolver: conditions.ResolverOperator, Also: machines, + Summary: fmt.Sprintf("%s: %s said nothing on the send of the builds that waited for a gate within %s, and were "+ + "left as they were", p.ID, strings.Join(machines, ", "), inventory.ApplySilentAfter)} + w := releaseSilentWords(machines) + o.Headline, o.Explanation, o.Resolved, o.Needs = w.Headline, w.Explanation, w.Resolved, w.Needs + return []conditions.Observation{o}, nil + } + return nil, nil +} + // backlogObservation is what the gate's probe says of a release held after a failure. func backlogObservation() []conditions.Observation { if backlogNow.held == "" || len(backlogNow.waiting) == 0 { diff --git a/cmd/mesh-controller/release_plan.go b/cmd/mesh-controller/release_plan.go index 77060d8b..68fdca5d 100644 --- a/cmd/mesh-controller/release_plan.go +++ b/cmd/mesh-controller/release_plan.go @@ -595,7 +595,7 @@ func advancePlans(ctx context.Context, open *stores) { return } defer release() - // What waits for a gate, released one machine at a time (ADR 0236). + // What waits for a gate, released one machine first, then the rest at once (ADR 0236, ADR 0353). if _, err := releaseBacklog(ctx, open, ""); err != nil { fmt.Printf("plans: what waits for a gate could not be looked at: %v\n", err) } @@ -1056,10 +1056,36 @@ func firstSend(ctx context.Context, open *stores, p *inventory.Plan, node string runningOf map[string][]string) error { inv := open.inventory before, beforeKnown, _ := inv.SentBuilds(ctx, node) + // **A machine already sent what this send is for is not sent again** (novox/hq ADR 0353 §6): sent by this plan + // before a controller was replaced and before it was kept, or by another walk. Judged from that send, and from + // the readings of a gate already judging it there. + // Only a send the machine reported it applied is one it has; one it never reported on is sent, as before + // (issue 280: what it ran before is then that build, and a failure leaves it as it was). + if beforeKnown && slices.IndexFunc(modules, func(m string) bool { return !sameCommit(before[m], p.Modules[m].Commit) }) < 0 && + appliedLastSend(ctx, inv, node) { + if others, err := gatedMoves(ctx, open, node, nil); err == nil && len(others) == 0 { + adoptSend(ctx, open, p, node, modules, before, runningOf) + return nil + } + } + // **What the machine ran before is kept before the send is made** (novox/hq ADR 0353 §6): a controller replaced + // between the send and keeping it reads it here, and a failure still puts the machine back. + kept := false + for _, m := range modules { + s := p.Modules[m] + if beforeKnown && (s.Previous == "" || !sameCommit(before[m], s.Commit)) && s.Previous != before[m] { + s.Previous, kept = before[m], true + } + } + if kept { + if err := inv.SavePlan(ctx, p); err != nil { + return fmt.Errorf("what %s ran before is not kept, so nothing is sent: %w", node, err) + } + } owns := make([]inventory.CarriedMove, 0, len(modules)) for _, m := range modules { s := p.Modules[m] - owns = append(owns, inventory.CarriedMove{Module: m, Node: node, From: before[m], To: s.Commit, Build: s.Build}) + owns = append(owns, inventory.CarriedMove{Module: m, Node: node, From: s.Previous, To: s.Commit, Build: s.Build}) } // A gated send (ADR 0236): everything waiting on the machine goes with the tier, and the gate judges // all of it there. @@ -1074,10 +1100,7 @@ func firstSend(ctx context.Context, open *stores, p *inventory.Plan, node string sentWhat := sentNow(ctx, inv, sent) for _, m := range modules { s := p.Modules[m] - // What the first machine ran before: what a failed gate puts back (ADR 0236). - if beforeKnown { - s.Previous = before[m] - } + // What the first machine ran before (ADR 0236), kept above before the send. s.First, s.FirstAt = sent, &now s.Gate = &inventory.PlanGate{Component: coreComponent(m), Machines: firstRunning(sent, runningOf[m]), From: s.Previous, To: s.Commit, Since: &now, Sent: sentWhat} @@ -1106,6 +1129,113 @@ func firstSend(ctx context.Context, open *stores, p *inventory.Plan, node string return nil } +// adoptSend records a first send the machine already had (novox/hq ADR 0353 §6, §7): nothing is sent; the gate counts +// from the machine's last send, and takes over the readings of an open gate of another walk judging the same build +// there — whose carried move also says what the machine ran before, for a rollback. +func adoptSend(ctx context.Context, open *stores, p *inventory.Plan, node string, modules []string, + before map[string]string, runningOf map[string][]string) { + inv := open.inventory + since := time.Now().UTC() + last, found, err := inv.SentTo(ctx, node) + if err == nil && found && !last.At.IsZero() { + since = last.At + } + lead := modules[0] + other, id, from := judgingThere(ctx, inv, p.ID, lead, node, p.Modules[lead].Commit, p.Modules[lead].Build) + for _, m := range modules { + s := p.Modules[m] + // What it ran before: kept by this walk before its send was made (§6); else what the other walk's send + // moved it from; else the build itself, so a failure leaves it as it was (issue 280). + switch { + case s.Previous != "": + case m == lead && from != "": + s.Previous = from + default: + s.Previous = before[m] + } + at := since + s.First, s.FirstAt = []string{node}, &at + s.Gate = &inventory.PlanGate{Component: coreComponent(m), Machines: firstRunning([]string{node}, runningOf[m]), + From: s.Previous, To: s.Commit, Since: &at} + if found { + s.Gate.Sent = map[string]inventory.SentDeclaration{node: last} + } + s.GatedBy = "" + if m == lead { + s.Gate.TakeReadings(other, modules, id) + } else { + s.GatedBy = lead + } + } + p.State = inventory.PlanRolling + p.Note = fmt.Sprintf("tier %d built; %s already runs %s's build: nothing sent again, judged from that send", + p.Tier, node, strings.Join(modules, ", ")) + if id != "" { + p.Note += ", going on from the readings of " + id + } + fmt.Printf("%s: %s\n", p.ID, p.Note) +} + +// appliedLastSend says a machine reported it applied the declaration it was sent last. +func appliedLastSend(ctx context.Context, inv *inventory.Inventory, node string) bool { + reports, err := inv.LastReports(ctx) + if err != nil { + return false + } + for _, r := range reports { + if r.Node == node { + return r.Current && r.At != nil && r.Outcome == inventory.OutcomeApplied + } + } + return false +} + +// judgingThere is an open gate of another walk judging a module's build on a machine (novox/hq ADR 0353 §7), the +// walk's id, and what the machine ran before; nil when none does. +func judgingThere(ctx context.Context, inv *inventory.Inventory, self, module, node, commit, build string) (*inventory.PlanGate, string, string) { + plans, err := inv.OpenPlans(ctx) + if err != nil { + return nil, "", "" + } + // The same build, by its id where both name one, else by its commit. + same := func(b, c string) bool { + if b != "" && build != "" { + return b == build + } + return sameCommit(c, commit) + } + usable := func(g *inventory.PlanGate) bool { + return g != nil && !g.Pending && g.Since != nil && (g.Verdict == "" || g.Verdict == inventory.GatePassed) && + slices.Contains(g.Machines, node) + } + for _, q := range plans { + if q.ID == self { + continue + } + if r := q.Release; r != nil { + gates := []*inventory.PlanGate{r.Gate} + for _, g := range r.Gates { + gates = append(gates, g) + } + for _, g := range gates { + if !usable(g) { + continue + } + for _, c := range g.Carried { + if c.Module == module && c.Node == node && same(c.Build, c.To) { + return g, q.ID, c.From + } + } + } + continue + } + if s := q.Modules[module]; s != nil && usable(s.Gate) && same(s.Build, s.Commit) { + return s.Gate, q.ID, s.Previous + } + } + return nil, "", "" +} + // passedWith keeps, on a module sent with others, the verdict of the gate that judged the send: the gate // on the first of them passed, and its pass was kept for every build it carried (passCarried). func passedWith(module string, s *inventory.PlanModule, lead string, g *inventory.PlanGate) { @@ -1594,9 +1724,19 @@ func plansCommand(ctx context.Context, args []string) error { // A release plan's walk (ADR 0236): machines done, the one judged, those to come. fmt.Printf(" machines in order: %s; done: %s; skipped: %s\n", strings.Join(r.Order, ", "), orNone(strings.Join(r.Done, ", ")), orNone(strings.Join(r.Skipped, ", "))) - if r.Gate != nil { - fmt.Printf(" %s\n", gateLine(r.Gate)) - for _, c := range r.Gate.Carried { + if len(r.Failed)+len(r.Silent) > 0 { + fmt.Printf(" failed: %s; silent: %s\n", orNone(strings.Join(r.Failed, ", ")), orNone(strings.Join(r.Silent, ", "))) + } + gates := []*inventory.PlanGate{r.Gate} + for _, n := range r.Order { + gates = append(gates, r.Gates[n]) + } + for _, g := range gates { + if g == nil { + continue + } + fmt.Printf(" %s\n", gateLine(g)) + for _, c := range g.Carried { fmt.Printf(" %-22s %s → %s\n", c.Module, short(c.From), short(c.To)) if c.Recreates != "" { fmt.Printf(" %-22s %s\n", "", c.Recreates) diff --git a/cmd/mesh-controller/rest_at_once_test.go b/cmd/mesh-controller/rest_at_once_test.go new file mode 100644 index 00000000..8b6f7155 --- /dev/null +++ b/cmd/mesh-controller/rest_at_once_test.go @@ -0,0 +1,372 @@ +package main + +import ( + "context" + "slices" + "strings" + "testing" + "time" + + "github.com/novox/mesh-controller/internal/catalogue" + "github.com/novox/mesh-controller/internal/inventory" +) + +// novox/hq ADR 0353: after its first machine passes, a build goes to every other machine at once, each judged by +// its own gate and a failing one put back by itself; and a gate's readings outlive the controller that took them — +// a machine already sent what a send is for is not sent again, and its judging goes on from what was read. +// +// Seen 2026-10-14: release-1791930194184991192 sent ace, g14, shanks and the control node one after another in +// eleven minutes; the new controller its last send started sent the control node again the controller build it +// had just applied, and judged it from its first reading. + +// aBacklogOnThree is the backlog fixture with a third machine, desk, running app at c1 like the others. +func aBacklogOnThree(t *testing.T) *backlogMesh { + t.Helper() + b := aBacklog(t) + ctx := t.Context() + inv := b.open.inventory + if _, err := inv.AddNode(ctx, "desk"); err != nil { + t.Fatal(err) + } + if _, err := inv.Assign(ctx, "desk", "app"); err != nil { + t.Fatal(err) + } + if err := inv.RecordSent(ctx, nodeID(t, b.open, "desk"), "d-desk", map[string]string{"app": "c1"}); err != nil { + t.Fatal(err) + } + wasHeard := releaseHeard + releaseHeard = func(context.Context, *stores) (map[string]bool, error) { + return map[string]bool{"anchor": true, "laptop": true, "desk": true}, nil + } + t.Cleanup(func() { releaseHeard = wasHeard }) + return b +} + +// events records every send and every judging, in order. +func events(t *testing.T) *[]string { + t.Helper() + var said []string + wasSend, wasGather := sendRollout, gatherGateFacts + sendRollout = func(ctx context.Context, open *stores, names []string) ([]string, error) { + said = append(said, "send "+strings.Join(names, ",")) + return wasSend(ctx, open, names) + } + gatherGateFacts = func(ctx context.Context, open *stores, component string) (gateFacts, error) { + said = append(said, "judge") + return wasGather(ctx, open, component) + } + t.Cleanup(func() { sendRollout, gatherGateFacts = wasSend, wasGather }) + return &said +} + +// Once the first machine passed, the rest are both sent before either is judged (ADR 0353 §1); before, the second +// was judged to its verdict before the third was sent. +func TestAReleaseSendsTheRestAtOnceAfterTheFirstPasses(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklogOnThree(t) + ctx := t.Context() + said := events(t) + gateEvery = 0 + for i := 0; i < 12; i++ { + advancePlans(ctx, b.open) + } + p := b.release(t) + if p.State != inventory.PlanDone || len(p.Release.Done) != 3 { + t.Fatalf("the release is %s: %s %+v", p.State, p.Note, p.Release) + } + desk, laptop := slices.Index(*said, "send desk"), slices.Index(*said, "send laptop") + if desk < 0 || laptop < 0 || slices.Index(*said, "send anchor") > min(desk, laptop) { + t.Fatalf("sends and judgings: %v", *said) + } + if slices.Contains((*said)[min(desk, laptop):max(desk, laptop)], "judge") { + t.Fatalf("the rest were sent one at a time, each judged before the next: %v", *said) + } + if v, found, err := b.open.inventory.GateOf(ctx, "build-late-c2"); err != nil || !found || v.Verdict != inventory.GatePassed { + t.Fatalf("late's pass on laptop was not kept: %+v %v %v", v, found, err) + } +} + +// A machine of the rest whose gate fails is put back at once, and the other machine of the step goes on to its own +// verdict (ADR 0353 §3); before, the walk stopped at the failure and the machine after it was never sent. +func TestARestMachineThatFailsIsPutBackAloneAndTheOthersAreJudgedOn(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklogOnThree(t) + ctx := t.Context() + inv := b.open.inventory + gateEvery = 0 + b.broken["desk"] = true // desk fails what it is sent in the rest's step + advancePlans(ctx, b.open) + for i := 0; i < 12 && b.release(t).Open(); i++ { + advancePlans(ctx, b.open) + } + p := b.release(t) + if p.State != inventory.PlanFailed || !slices.Equal(p.Release.Failed, []string{"desk"}) { + t.Fatalf("the release is %s: %s %+v", p.State, p.Note, p.Release) + } + if !slices.ContainsFunc(b.sent, func(s []string) bool { return slices.Equal(s, []string{"laptop"}) }) { + t.Fatalf("laptop was never sent: %v", b.sent) + } + if !slices.Contains(p.Release.Done, "laptop") { + t.Fatalf("laptop has no verdict of its own: %+v", p.Release) + } + // late waited on laptop alone and was healthy there: its pass is kept, whatever desk did. + if v, found, err := inv.GateOf(ctx, "build-late-c2"); err != nil || !found || v.Verdict != inventory.GatePassed { + t.Fatalf("late's pass on laptop was not kept: %+v %v %v", v, found, err) + } + // app failed on desk: marked, and put back wherever it was sent (ADR 0236). + if failed, _ := inv.GateFailed(ctx, "build-app-c2"); !failed { + t.Fatal("app's build is not marked failed") + } + if sent, _, _ := inv.SentBuilds(ctx, "desk"); sent["app"] != "c1" { + t.Fatalf("desk was not put back: %v", sent) + } +} + +// A walk whose first machine already runs its build under another walk's gate sends nothing and goes on from that +// gate's readings (ADR 0353 §7); before, it sent the machine again and judged from its first reading. +func TestAWalkTakesOverTheReadingsOfAGateJudgingTheSameBuildThere(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklog(t) + ctx := t.Context() + inv := b.open.inventory + advancePlans(ctx, b.open) // the release sends anchor app c2, and reads it once + release := b.release(t) + judging := release.Release.Gate + if judging == nil || judging.Pending || judging.Passes == 0 { + t.Fatalf("the release is not judging anchor: %+v", judging) + } + built := time.Now().UTC() + plan := inventory.Plan{ID: "plan-app", Repository: "novox/mesh-catalog", Branch: "main", Commit: "c2", + Created: built, State: inventory.PlanBuilding, Tiers: [][]string{{"app"}}, + Modules: map[string]*inventory.PlanModule{"app": {State: "built", BuiltAt: &built, Commit: "c2", + Build: "build-app-c2"}}} + if err := inv.SavePlan(ctx, &plan); err != nil { + t.Fatal(err) + } + advancePlans(ctx, b.open) + if len(b.sent) != 1 { + t.Fatalf("anchor was sent again what it runs: %v", b.sent) + } + p, err := inv.PlanByID(ctx, "plan-app") + if err != nil { + t.Fatal(err) + } + g := p.Modules["app"].Gate + if g == nil || g.Since == nil || !g.Since.Equal(*judging.Since) || g.Passes < judging.Passes || g.TookOver != release.ID { + t.Fatalf("the walk's gate starts from zero: %+v (the release's: %+v)", g, judging) + } + if p.Modules["app"].Previous != "c1" { + t.Fatalf("what anchor ran before is %q, not c1", p.Modules["app"].Previous) + } +} + +// A walk's first send that was made and not kept — the controller replaced between the two — is not made again +// by the controller that takes over: the gate counts from that send (ADR 0353 §5, §6). +func TestAFirstSendMadeAndNotKeptIsNotSentAgain(t *testing.T) { + judgeLeavesAsTheCore(t) + g := aGateMesh(t) + ctx := t.Context() + // The old controller sends anchor app c2 and is replaced before it keeps the send. + made := sendRollout + sendRollout = func(ctx context.Context, open *stores, names []string) ([]string, error) { + if _, err := made(ctx, open, names); err != nil { + return nil, err + } + return nil, context.Canceled + } + advancePlans(ctx, g.open) + sendRollout = made + if len(g.sent) != 1 || g.plan(t).Modules["app"].FirstAt != nil { + t.Fatalf("the fixture's send was not made, or was kept: %v", g.sent) + } + last, _, err := g.open.inventory.SentTo(ctx, "anchor") + if err != nil { + t.Fatal(err) + } + advancePlans(ctx, g.open) + if len(g.sent) != 1 { + t.Fatalf("the send was made again: %v", g.sent) + } + s := g.plan(t).Modules["app"] + if s.FirstAt == nil || s.Gate == nil || s.Gate.Since == nil || !s.Gate.Since.Equal(last.At) { + t.Fatalf("the gate does not count from the send made: %+v %+v", s, s.Gate) + } + // What anchor ran before was kept before the send, so a failure still puts it back (§5). + if s.Previous != "c1" || s.Gate.From != "c1" { + t.Fatalf("what anchor ran before is %q (gate from %q), not c1", s.Previous, s.Gate.From) + } +} + +// A release's send kept before it was made, and made, is judged by the controller that takes over without being +// made again (ADR 0353 §8). +func TestAReleaseSendKeptAndMadeIsNotMadeAgain(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklog(t) + ctx := t.Context() + inv := b.open.inventory + p, err := releaseBacklog(ctx, b.open, "") + if err != nil || p == nil { + t.Fatalf("no release opened: %v", err) + } + if _, err := advanceRelease(ctx, b.open, p); err != nil || p.Release.Gate == nil || !p.Release.Gate.Pending { + t.Fatalf("the send was not kept before it was made: %+v %v", p.Release, err) + } + if err := inv.SavePlan(ctx, p); err != nil { + t.Fatal(err) + } + // The send is made, and the controller is replaced before it keeps it. + if _, err := sendRollout(ctx, b.open, []string{"anchor"}); err != nil { + t.Fatal(err) + } + advancePlans(ctx, b.open) + if len(b.sent) != 1 { + t.Fatalf("anchor was sent again: %v", b.sent) + } + g := b.release(t).Release.Gate + if g == nil || g.Pending || g.Since == nil || len(g.Carried) == 0 { + t.Fatalf("anchor is not judged on the send made: %+v", g) + } +} + +// A build no gate has passed goes to one machine of a step, its first; a machine that would be its second waits +// for the next step, and one whose every build passed goes now (ADR 0353 §2). +func TestABuildThatPassedNoGateGoesToOneMachineOfAStep(t *testing.T) { + waiting := map[string][]inventory.CarriedMove{ + "desk": {{Module: "app", To: "c2"}, {Module: "new", To: "n1"}}, + "laptop": {{Module: "app", To: "c2"}, {Module: "new", To: "n1"}}, + "shanks": {{Module: "app", To: "c2"}}, + "novox": {{Module: "app", To: "c2"}, {Module: "ctl", To: "k1"}}, + } + passed := func(module, commit string) bool { return module == "app" && commit == "c2" } + never := func(string) bool { return false } + step := restStep([]string{"desk", "laptop", "shanks", "novox"}, waiting, passed, never) + if !slices.Equal(step, []string{"desk", "shanks", "novox"}) { + t.Fatalf("the step is %v: laptop would be new's second machine before it passed", step) + } + if step := restStep([]string{"laptop"}, waiting, func(string, string) bool { return true }, never); !slices.Equal(step, []string{"laptop"}) { + t.Fatalf("a machine whose every build passed waits: %v", step) + } +} + +// The node-engine and the mesh's resolver go to the rest one machine at a time, even passed (ADR 0353 §9): a step +// carries either to one machine only, while what else passed goes everywhere at once. +func TestTheNodeEngineAndTheResolverGoToTheRestOneAtATime(t *testing.T) { + waiting := map[string][]inventory.CarriedMove{ + "desk": {{Module: hostModule, To: "h2"}}, + "laptop": {{Module: hostModule, To: "h2"}, {Module: "app", To: "c2"}}, + "shanks": {{Module: "app", To: "c2"}}, + "novox": {{Module: "dns", To: "d2"}}, + "ace": {{Module: "dns", To: "d2"}}, + } + resolver := catalogue.Manifest{Module: "dns", Claims: []catalogue.Claim{{Name: resolverSeat}}} + alone := func(m string) bool { + if m == "dns" { + return oneAtATime(m, resolver) + } + return oneAtATime(m, catalogue.Manifest{Module: m}) + } + step := restStep([]string{"desk", "laptop", "shanks", "novox", "ace"}, waiting, + func(string, string) bool { return true }, alone) + if !slices.Equal(step, []string{"desk", "shanks", "novox"}) { + t.Fatalf("the step is %v: the node-engine and the resolver each reach one machine of it", step) + } + if oneAtATime("app", catalogue.Manifest{Module: "app"}) { + t.Fatal("a leaf module is held to one machine at a time") + } +} + +// A first machine that fails still stops the walk (ADR 0236 §5, kept by ADR 0353 §3): the rest are never sent. +func TestAFirstMachineThatFailsStillStopsTheWalk(t *testing.T) { + b := aBacklogOnThree(t) + ctx := t.Context() + gateEvery = 0 + b.broken["anchor"] = true + advancePlans(ctx, b.open) + for i := 0; i < 6 && b.release(t).Open(); i++ { + advancePlans(ctx, b.open) + } + if p := b.release(t); p.State != inventory.PlanFailed || p.Release.Rest { + t.Fatalf("the release is %s: %s %+v", p.State, p.Note, p.Release) + } + for _, s := range b.sent { + if !slices.Equal(s, []string{"anchor"}) { + t.Fatalf("sent %v: the rest after a failed first machine", b.sent) + } + } +} + +// The control node is sent last in a step of the rest (ADR 0353 §1): the machine holding the controller is +// ordered after every other. +func TestTheControlNodeIsSentLastInAStep(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklogOnThree(t) + ctx := t.Context() + register(t, b.open, catalogue.Manifest{Module: catalogue.ControllerSeatName, Version: "1"}) + if _, err := b.open.inventory.Assign(ctx, "anchor", catalogue.ControllerSeatName); err != nil { + t.Fatal(err) + } + said := events(t) + gateEvery = 0 + for i := 0; i < 12 && (len(b.sent) == 0 || b.release(t).Open()); i++ { + advancePlans(ctx, b.open) + } + if p := b.release(t); !slices.Equal(p.Release.Order, []string{"desk", "laptop", "anchor"}) { + t.Fatalf("the order is %v", p.Release.Order) + } + laptop, anchor := slices.Index(*said, "send laptop"), slices.Index(*said, "send anchor") + if laptop < 0 || anchor < laptop || slices.Contains((*said)[laptop:anchor], "judge") { + t.Fatalf("the control node is not sent last in the rest's step: %v", *said) + } +} + +// A machine of the rest that says nothing on its send is left silent after fifteen minutes: not failed, not put +// back; a rest that went silent whole ends failed, said as a condition (ADR 0353 §4). +func TestARestMachineSilentPastItsBoundIsLeftAndSaid(t *testing.T) { + judgeLeavesAsTheCore(t) + b := aBacklog(t) + ctx := t.Context() + inv := b.open.inventory + reporting := sendRollout + sendRollout = func(ctx context.Context, open *stores, names []string) ([]string, error) { + if slices.Equal(names, []string{"laptop"}) { // laptop sleeps: sent, and says nothing + b.sent = append(b.sent, names) + current, _ := open.inventory.CurrentBuilds(ctx) + return names, open.inventory.RecordSent(ctx, nodeID(t, open, "laptop"), "d-laptop-asleep", + map[string]string{"app": current["app"].Commit, "late": current["late"].Commit, "same": current["same"].Commit}) + } + return reporting(ctx, open, names) + } + t.Cleanup(func() { sendRollout = reporting }) + gateEvery = 0 + for i := 0; i < 8; i++ { + advancePlans(ctx, b.open) + } + p := b.release(t) + g := p.Release.Gates["laptop"] + if !p.Open() || g == nil || g.Verdict != "" || g.Silent { + t.Fatalf("laptop is judged before fifteen minutes: %s %s %+v", p.State, p.Note, g) + } + since := time.Now().Add(-inventory.ApplySilentAfter - time.Minute) + g.Since = &since + if err := inv.SavePlan(ctx, &p); err != nil { + t.Fatal(err) + } + advancePlans(ctx, b.open) + p = b.release(t) + if p.State != inventory.PlanFailed || !slices.Equal(p.Release.Silent, []string{"laptop"}) || len(p.Release.Failed) > 0 { + t.Fatalf("the release is %s: %s %+v", p.State, p.Note, p.Release) + } + if failed, _ := inv.GateFailed(ctx, "build-late-c2"); failed { + t.Fatal("a silent machine's build was marked failed") + } + if sent, _, _ := inv.SentBuilds(ctx, "laptop"); sent["late"] != "c2" { + t.Fatalf("laptop was put back: %v", sent) + } + obs, err := silentObservation(ctx, inv) + if err != nil { + t.Fatal(err) + } + if len(obs) != 1 || obs[0].Kind != "release-silent" || !slices.Equal(obs[0].Also, []string{"laptop"}) { + t.Fatalf("the silence is not said: %+v", obs) + } +} diff --git a/cmd/mesh-controller/upgrades.go b/cmd/mesh-controller/upgrades.go index 7f8bc8e2..7570fb79 100644 --- a/cmd/mesh-controller/upgrades.go +++ b/cmd/mesh-controller/upgrades.go @@ -79,7 +79,7 @@ func (f following) Upgraded(ctx context.Context, u link.Upgraded) error { // `rebuild`, a `replay` registered — was sent here one machine after another without any judging; it // now waits for a gate like any other build, and the release plan walks it through the machines, one // at a time, each judged. - fmt.Printf("%s moved to %s outside a plan; a release plan sends it to %s, one machine at a time, each judged "+ + fmt.Printf("%s moved to %s outside a plan; a release plan sends it to %s, one machine first and then the rest at once, each judged "+ "(`upgrade backlog` lists what waits)\n", u.Module, shortCommit(u.Commit), readableList(on)) return nil } diff --git a/cmd/mesh-controller/walk_said.go b/cmd/mesh-controller/walk_said.go index ddb724e4..3e5d9bad 100644 --- a/cmd/mesh-controller/walk_said.go +++ b/cmd/mesh-controller/walk_said.go @@ -177,6 +177,12 @@ func walkReached(p inventory.Plan, running func(string) []string) map[string][]s reached = appendNew(reached, r.Gate.Machines...) carried(r.Gate) } + for _, g := range r.Gates { + if g != nil && !g.Pending { + reached = appendNew(reached, g.Machines...) + carried(g) + } + } for _, tier := range p.Tiers { for _, m := range tier { for _, n := range running(m) { diff --git a/internal/inventory/plans.go b/internal/inventory/plans.go index fd86c2b9..c9f712cf 100644 --- a/internal/inventory/plans.go +++ b/internal/inventory/plans.go @@ -5,6 +5,7 @@ import ( "encoding/json" "errors" "fmt" + "slices" "strings" "time" @@ -391,6 +392,65 @@ type PlanGate struct { // reports its module's container left running because it said busy, the judging waits and has not // started; once it reports the container replaced, Since is moved to then. Nil when nothing waited. Busy *GateBusy `json:"busy,omitempty"` + // Pending says the send this gate judges is kept and not yet made (novox/hq ADR 0353 §8): a controller that + // takes over reads the machine's last send before making it, and never sends it twice. + Pending bool `json:"pending,omitempty"` + // Silent says the machine reported nothing on its send within ApplySilentAfter (novox/hq ADR 0353 §4): left, + // neither failed nor put back. + Silent bool `json:"silent,omitempty"` + // TookOver names the gate whose readings this one took over (novox/hq ADR 0353 §7): another walk's judging of + // the same build on the same machine, which this gate goes on from instead of starting from zero. + TookOver string `json:"took_over,omitempty"` +} + +// TakeReadings makes this gate go on from another's readings of the same build on the same machine (novox/hq ADR +// 0353 §7): when the judging began, when each machine applied, the passes counted and when, and what each found. +// A module's own count is taken only for the modules this gate judges. +func (g *PlanGate) TakeReadings(from *PlanGate, modules []string, id string) { + if from == nil { + return + } + if from.Since != nil { + since := *from.Since + g.Since = &since + } + for n, s := range from.Sent { + if slices.Contains(g.Machines, n) { + if g.Sent == nil { + g.Sent = map[string]SentDeclaration{} + } + g.Sent[n] = s + } + } + for n, at := range from.Applied { + if slices.Contains(g.Machines, n) { + if g.Applied == nil { + g.Applied = map[string]time.Time{} + } + g.Applied[n] = at + } + } + g.Passes, g.Faulted, g.HealthySeen, g.Last = from.Passes, from.Faulted, from.HealthySeen, from.Last + if from.LastPass != nil { + at := *from.LastPass + g.LastPass = &at + } + for _, m := range modules { + if n, counted := from.Healthy[m]; counted { + if g.Healthy == nil { + g.Healthy = map[string]int{} + } + g.Healthy[m] = n + } + if at, counted := from.HealthyAt[m]; counted { + if g.HealthyAt == nil { + g.HealthyAt = map[string]time.Time{} + } + g.HealthyAt[m] = at + } + } + g.Readings = append([]GateReading(nil), from.Readings...) + g.TookOver = id } // GateBusy is a first-node gate's wait for a busy container (novox/hq ADR 0335 §4). @@ -477,9 +537,17 @@ type CarriedMove struct { type PlanRelease struct { Order []string `json:"order"` Next int `json:"next"` - // Gate is the machine being judged; nil between machines. + // Gate is the first machine being judged; nil between machines. Gate *PlanGate `json:"gate,omitempty"` - Done []string `json:"done,omitempty"` + // Rest says the first machine passed (novox/hq ADR 0353 §1): every other machine is sent at once, each judged by + // its own gate in Gates, by machine — a build no gate passed going to one machine of a step only (§2). + Rest bool `json:"rest,omitempty"` + Gates map[string]*PlanGate `json:"gates,omitempty"` + // Failed are the machines of the rest whose gate failed, and Silent those that said nothing on their send + // within ApplySilentAfter (ADR 0353 §3 and §4). + Failed []string `json:"failed,omitempty"` + Silent []string `json:"silent,omitempty"` + Done []string `json:"done,omitempty"` // Skipped are the machines not heard from when their turn came, left as they were. Skipped []string `json:"skipped,omitempty"` // By is the person who released it, empty when the mesh did. diff --git a/testdata/store-tests b/testdata/store-tests index bbfc1467..48692606 100644 --- a/testdata/store-tests +++ b/testdata/store-tests @@ -61,3 +61,11 @@ ./cmd/mesh-controller TestAProviderElsewhereServesItsOwnComposedName ./cmd/mesh-controller TestAReusedCheckBuildIsRegisteredByItsDigestAndGatedAsABuiltOne ./internal/inventory TestAReportOfABusyContainerIsKeptAndReadBack +./cmd/mesh-controller TestAReleaseSendsTheRestAtOnceAfterTheFirstPasses +./cmd/mesh-controller TestARestMachineThatFailsIsPutBackAloneAndTheOthersAreJudgedOn +./cmd/mesh-controller TestAWalkTakesOverTheReadingsOfAGateJudgingTheSameBuildThere +./cmd/mesh-controller TestAFirstSendMadeAndNotKeptIsNotSentAgain +./cmd/mesh-controller TestAReleaseSendKeptAndMadeIsNotMadeAgain +./cmd/mesh-controller TestAFirstMachineThatFailsStillStopsTheWalk +./cmd/mesh-controller TestTheControlNodeIsSentLastInAStep +./cmd/mesh-controller TestARestMachineSilentPastItsBoundIsLeftAndSaid