Give each pending assignment its own condition, and keep a raised row until it clears

Second review of #150: one key per machine and module let a newer failure
be cleared in the tick that raised it, pruning could orphan an open
condition, and a build no longer waited for read as never asked although it
may still run.

internal/inventory/fortest.go

@@ -68,3 +68,13 @@ func ForTest(t *testing.T) *Inventory {
 	}
 	return inv
 }
+
+// RemoveNodeForTest removes a machine's record as a person removing it from the store would: no verb of the
+// mesh removes one yet, and what hangs off it must go with it.
+func (i *Inventory) RemoveNodeForTest(ctx context.Context, name string) (bool, error) {
+	tag, err := i.store.Pool().Exec(ctx, `delete from node where name = $1`, name)
+	if err != nil {
+		return false, err
+	}
+	return tag.RowsAffected() == 1, nil
+}

internal/inventory/migrations/0082-an-assignment-waits-for-its-build.sql

@@ -8,8 +8,10 @@
 
 -- Every build the controller asked for: a merge's new module, a plan's tier, a person's `build`, an `assign`
 -- with build. Kept once the ask is made, never before. Its outcome is the `build` row of the same id;
--- an ask without one is still running, or was lost. not_asked is why the ask could not be made or waited
--- for; for a merge that could not ask, the id is the controller's and no build's.
+-- an ask without one is still running, or was lost. not_asked is why the ask could not be handed over; for a
+-- merge that could not ask, the id is the controller's and no build's. outcome_unknown is why an asker that
+-- handed it over stopped waiting: the build may still run, and is read as in flight until its outcome or
+-- its bound.
 create table build_request (
     id          text primary key,
     repository  text not null,
@@ -19,6 +21,7 @@ create table build_request (
     commit_hash text not null default '',
     asked_for   text not null default '',
     not_asked   text,
+    outcome_unknown text,
     asked_at    timestamptz not null default now()
 );
 create index build_request_asked_at on build_request (asked_at);
@@ -28,7 +31,7 @@ create index build_request_asked_at on build_request (asked_at);
 -- assignment was refused then), expired (the build failed, was not registered, or said nothing within its
 -- bound) or withdrawn (`unassign`). raised_at is when an expiry or refusal was raised as a condition,
 -- acknowledged_at when a person took it back with `unassign`, cleared_at when its condition was cleared.
--- Gone with its machine. Ended rows are deleted after 30 days.
+-- Gone with its machine. Ended rows are deleted after 30 days, once any condition raised for them is cleared.
 create table pending_assignment (
     id              bigserial primary key,
     node            uuid not null references node (id) on delete cascade,

internal/inventory/pending.go

@@ -32,9 +32,12 @@ type BuildRequest struct {
 	Commit string
 	// For says who asked: "merge", "plan", "build" (a person), "assign".
 	For string
-	// NotAsked is why the ask could not be made, or not waited for; empty when it was.
+	// NotAsked is why the ask could not be handed over; empty when it was.
 	NotAsked string
-	At       time.Time
+	// OutcomeUnknown is why an asker that handed the build over stopped waiting for it: asked, outcome
+	// unknown. Read as in flight until its outcome or its bound.
+	OutcomeUnknown string
+	At             time.Time
 }
 
 // Name is the module this request is expected to register, read from its directory: the last element of
@@ -80,8 +83,8 @@ func (i *Inventory) RecordBuildRequest(ctx context.Context, a BuildRequest) erro
 	return err
 }
 
-// MarkNotAsked says a build request was asked and its asker could not hand it over, or stopped waiting for
-// it: the words are kept, unless its outcome was heard first.
+// MarkNotAsked says a kept build request was never handed over: the words are kept, unless its outcome was
+// heard first.
 func (i *Inventory) MarkNotAsked(ctx context.Context, id, why string) error {
 	_, err := i.store.Pool().Exec(ctx,
 		`update build_request set not_asked = $2
@@ -89,6 +92,15 @@ func (i *Inventory) MarkNotAsked(ctx context.Context, id, why string) error {
 	return err
 }
 
+// MarkOutcomeUnknown says the asker of a build it handed over stopped waiting for it: asked, outcome unknown.
+// The build may still run; its outcome, when heard, is the last word.
+func (i *Inventory) MarkOutcomeUnknown(ctx context.Context, id, why string) error {
+	_, err := i.store.Pool().Exec(ctx,
+		`update build_request set outcome_unknown = $2
+		  where id = $1 and not exists (select 1 from build where build.id = $1)`, id, why)
+	return err
+}
+
 // RequestsNamed is every build request kept whose directory names the module, or whose build said it built
 // it, newest first, each with its outcome.
 func (i *Inventory) RequestsNamed(ctx context.Context, name string) ([]RequestOutcome, error) {
@@ -126,7 +138,7 @@ func (i *Inventory) RequestedNames(ctx context.Context) ([]string, error) {
 func (i *Inventory) requests(ctx context.Context) ([]RequestOutcome, error) {
 	rows, err := i.store.Pool().Query(ctx,
 		`select a.id, a.repository, a.seat, a.source_path, a.ref, a.commit_hash, a.asked_for,
-		        coalesce(a.not_asked, ''), a.asked_at,
+		        coalesce(a.not_asked, ''), coalesce(a.outcome_unknown, ''), a.asked_at,
 		        b.id is not null, coalesce(b.failed, ''), coalesce(b.module, ''), coalesce(b.at, a.asked_at)
 		   from build_request a left join build b on b.id = a.id
 		  order by a.asked_at desc, a.id desc`)
@@ -138,7 +150,7 @@ func (i *Inventory) requests(ctx context.Context) ([]RequestOutcome, error) {
 	for rows.Next() {
 		var a RequestOutcome
 		if err := rows.Scan(&a.ID, &a.Repository, &a.Seat, &a.Path, &a.Ref, &a.Commit, &a.For, &a.NotAsked,
-			&a.At, &a.Heard, &a.Failed, &a.Module, &a.HeardAt); err != nil {
+			&a.OutcomeUnknown, &a.At, &a.Heard, &a.Failed, &a.Module, &a.HeardAt); err != nil {
 			return nil, err
 		}
 		out = append(out, a)
@@ -342,13 +354,34 @@ func (i *Inventory) SettlePending(ctx context.Context, id int64, from, state, no
 	return tag.RowsAffected() == 1, nil
 }
 
-// ForgetEndedPending deletes ended pending assignments settled more than KeptFor ago, and says how many.
+// ForgetEndedPending deletes ended pending assignments settled more than KeptFor ago, and says how many. One
+// raised as a condition is kept until that condition is cleared: deleting it would leave the condition
+// with nothing to answer it.
 func (i *Inventory) ForgetEndedPending(ctx context.Context, now time.Time) (int64, error) {
 	tag, err := i.store.Pool().Exec(ctx,
-		`delete from pending_assignment where state not in ('waiting', 'applying') and settled_at < $1`,
+		`delete from pending_assignment where state not in ('waiting', 'applying') and settled_at < $1
+		   and (raised_at is null or cleared_at is not null)`,
 		now.Add(-KeptFor))
 	if err != nil {
 		return 0, err
 	}
 	return tag.RowsAffected(), nil
 }
+
+// PendingKnown says which of these pending assignments are still on record.
+func (i *Inventory) PendingKnown(ctx context.Context, ids []int64) (map[int64]bool, error) {
+	rows, err := i.store.Pool().Query(ctx, `select id from pending_assignment where id = any($1)`, ids)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[int64]bool{}
+	for rows.Next() {
+		var id int64
+		if err := rows.Scan(&id); err != nil {
+			return nil, err
+		}
+		out[id] = true
+	}
+	return out, rows.Err()
+}