Wait for a held node without pinning a pool connection, and give up after a bounded wait naming it (hq ADR 0100)

This commit is contained in:
2026-09-22 18:31:10 +02:00
parent 2e6b9d30bd
commit 65187d4de0
2 changed files with 95 additions and 9 deletions
+51 -9
View File
@@ -2,8 +2,22 @@ package inventory
import ( import (
"context" "context"
"errors"
"fmt"
"slices" "slices"
"strings"
"sync" "sync"
"time"
)
// ErrNodeBusy is a node another act kept holding for longer than a caller waits.
var ErrNodeBusy = errors.New("another act is composing or sending that node's declaration")
// HoldWaitFor is how long HoldNodes waits for a node another act holds, and HoldPoll how often it
// looks again. Variables so a test need not wait minutes.
var (
HoldWaitFor = 2 * time.Minute
HoldPoll = 250 * time.Millisecond
) )
// HoldNodes serialises composing and sending a declaration per node (novox/hq ADR 0100): while // HoldNodes serialises composing and sending a declaration per node (novox/hq ADR 0100): while
@@ -11,16 +25,39 @@ import (
// adopted could send that declaration after `converge --yes` sent the converged one, and the node // adopted could send that declaration after `converge --yes` sent the converged one, and the node
// would return to adopted with nobody having asked. // would return to adopted with nobody having asked.
// //
// Session-level advisory locks on one connection, taken in name order so two callers holding // Session-level advisory locks on one connection, all or none: a set not wholly free is given back
// overlapping sets cannot each wait on the other. Release gives every one back, and may be called // at once, so two callers holding overlapping sets never each wait on the other. **A waiter pins
// more than once; a caller whose context ends while waiting holds nothing. // no connection.** It looks again every HoldPoll with a connection borrowed for the look, and gives
// up after HoldWaitFor with ErrNodeBusy naming the node — so a stuck holder costs the pool one
// connection, never one per caller queued behind it. Release gives every one back, and may be
// called more than once.
func (i *Inventory) HoldNodes(ctx context.Context, names []string) (func(), error) { func (i *Inventory) HoldNodes(ctx context.Context, names []string) (func(), error) {
sorted := slices.Clone(names) sorted := slices.Clone(names)
slices.Sort(sorted) slices.Sort(sorted)
sorted = slices.Compact(sorted) sorted = slices.Compact(sorted)
deadline := time.Now().Add(HoldWaitFor)
for {
release, busy, err := i.tryHold(ctx, sorted)
if err != nil || busy == "" {
return release, err
}
if time.Now().After(deadline) {
return nil, fmt.Errorf("%w: %s has been held for over %s — try again once it is done",
ErrNodeBusy, busy, HoldWaitFor)
}
select {
case <-ctx.Done():
return nil, ctx.Err()
case <-time.After(HoldPoll):
}
}
}
// tryHold takes every named node's lock or none, and says which node was busy when it took none.
func (i *Inventory) tryHold(ctx context.Context, sorted []string) (func(), string, error) {
conn, err := i.store.Pool().Acquire(ctx) conn, err := i.store.Pool().Acquire(ctx)
if err != nil { if err != nil {
return nil, err return nil, "", err
} }
var once sync.Once var once sync.Once
release := func() { release := func() {
@@ -36,12 +73,17 @@ func (i *Inventory) HoldNodes(ctx context.Context, names []string) (func(), erro
}) })
} }
for _, name := range sorted { for _, name := range sorted {
if _, err := conn.Exec(ctx, var took bool
`select pg_advisory_lock(hashtext('mesh-node-declaration:' || $1)::bigint)`, if err := conn.QueryRow(ctx,
name); err != nil { `select pg_try_advisory_lock(hashtext('mesh-node-declaration:' || $1)::bigint)`,
name).Scan(&took); err != nil {
release() release()
return nil, err return nil, "", err
}
if !took {
release()
return nil, strings.TrimSpace(name), nil
} }
} }
return release, nil return release, "", nil
} }
+44
View File
@@ -1,6 +1,8 @@
package inventory package inventory
import ( import (
"errors"
"strings"
"testing" "testing"
"time" "time"
) )
@@ -42,3 +44,45 @@ func TestHoldingANodeMakesTheNextHolderWait(t *testing.T) {
t.Fatal("releasing the node did not let the next holder in") t.Fatal("releasing the node did not let the next holder in")
} }
} }
// A waiter pins no pool connection while it waits, and gives up after a bounded wait saying which
// node is busy.
func TestAWaiterPinsNoConnectionAndGivesUp(t *testing.T) {
inv := fresh(t)
ctx := t.Context()
release, err := inv.HoldNodes(ctx, []string{"anchor"})
if err != nil {
t.Fatal(err)
}
defer release()
savedWait, savedPoll := HoldWaitFor, HoldPoll
HoldWaitFor, HoldPoll = 1500*time.Millisecond, 50*time.Millisecond
defer func() { HoldWaitFor, HoldPoll = savedWait, savedPoll }()
pool := inv.store.Pool()
base := pool.Stat().AcquiredConns()
const waiters = 3
done := make(chan error, waiters)
for range waiters {
go func() {
_, err := inv.HoldNodes(ctx, []string{"anchor"})
done <- err
}()
}
// While they wait, the pool lends nothing to them for longer than a look.
pinned := 0
for range 10 {
time.Sleep(60 * time.Millisecond)
if n := int(pool.Stat().AcquiredConns() - base); n > pinned {
pinned = n
}
}
if pinned >= waiters {
t.Fatalf("%d connections were held by %d waiters", pinned, waiters)
}
for range waiters {
if err := <-done; !errors.Is(err, ErrNodeBusy) || !strings.Contains(err.Error(), "anchor") {
t.Fatalf("a waiter did not give up naming the busy node: %v", err)
}
}
}