textmachine/platform/internal/pgstore/sink_test.go

411 lines
16 KiB
Go

package pgstore
import (
"context"
"encoding/json"
"sync"
"testing"
"time"
"github.com/jackc/pgx/v5"
"textmachine/platform/internal/ingest"
"textmachine/platform/internal/money"
)
// sinkFixture is one live run with a chapter to fold units into.
func sinkFixture(t *testing.T) (*Store, *RunSink, string, StartedRun) {
t.Helper()
s, ctx := testDB(t)
now := fundedAccount(t, s, ctx, "u1", "10")
seedBook(t, s, ctx, "bk1", "u1", 500)
exec(t, s, ctx, `insert into chapters (id, book_id, number, units_total) values ('c1','bk1',1,50)`)
run, err := s.StartRun(ctx, StartRunInput{UserID: "u1", BookID: "bk1", CeilingChapters: 10,
Ceiling: money.MicroUSD(300_000), Now: now}, 0, nil)
if err != nil {
t.Fatal(err)
}
sink := s.NewRunSink(run.AttemptID, run.ID, "bk1")
if err := sink.Begin(t.Context(), ingest.Hello{EngineRunID: "eng-1"}); err != nil {
t.Fatal(err)
}
return s, sink, run.ID, run
}
func apply(t *testing.T, s *RunSink, seq int64, typ ingest.Type, data any) error {
t.Helper()
body, err := json.Marshal(data)
if err != nil {
t.Fatal(err)
}
return s.Apply(t.Context(), ingest.Envelope{Seq: seq, Type: typ, Data: body}, ingest.Cursor{Offset: seq * 10})
}
// assertBookFirst proves an operation takes the BOOK's row lock BEFORE the row `probe` asks about,
// which is the package's global order (lockBook) and the thing a concurrency stress test can only
// sample.
//
// It holds the book lock, starts the operation, waits until the operation is actually blocked, and
// then asks whether the other row is still free. Free means the operation is waiting for the book
// with nothing else held — book-first. Taken means it grabbed the other row on the way, which is the
// order that deadlocks against every other transaction in this package.
func assertBookFirst(t *testing.T, s *Store, ctx context.Context, bookID string, probe func(context.Context, pgx.Tx) error, op func() error) {
t.Helper()
holder, err := s.pool.Begin(ctx)
if err != nil {
t.Fatal(err)
}
defer func() { _ = holder.Rollback(ctx) }()
var id string
if err := holder.QueryRow(ctx, `select id from books where id = $1 for update`, bookID).Scan(&id); err != nil {
t.Fatal(err)
}
done := make(chan error, 1)
go func() { done <- op() }()
// Wait for the operation to block on something. Without this the probe below could run before the
// operation had taken any lock at all and pass for the wrong reason.
deadline := time.Now().Add(20 * time.Second)
for {
// pg_stat_activity, not pg_locks: a transaction waiting for a ROW lock waits on the HOLDER's
// transactionid, and those pg_locks rows carry a null database, so a database-scoped query over
// pg_locks never sees the wait this test is built on.
var blocked int
if err := s.pool.QueryRow(ctx, `
select count(*) from pg_stat_activity
where datname = current_database() and wait_event_type = 'Lock'`).Scan(&blocked); err != nil {
t.Fatal(err)
}
if blocked > 0 {
break
}
if time.Now().After(deadline) {
t.Fatal("the operation never blocked on the book lock")
}
time.Sleep(10 * time.Millisecond)
}
tx, err := s.pool.Begin(ctx)
if err != nil {
t.Fatal(err)
}
err = probe(ctx, tx)
_ = tx.Rollback(ctx)
if err != nil {
t.Errorf("the row was already locked while the book was not: %v — this order deadlocks", err)
}
if err := holder.Rollback(ctx); err != nil {
t.Fatal(err)
}
if err := <-done; err != nil {
t.Errorf("the operation itself failed: %v", err)
}
}
// lockedAttempt and lockedReservation are the probes assertBookFirst asks with: NOWAIT turns "someone
// else holds this row" into an error instead of a wait.
func lockedAttempt(attemptID int64) func(context.Context, pgx.Tx) error {
return func(ctx context.Context, tx pgx.Tx) error {
var id int64
return tx.QueryRow(ctx, `select id from run_attempts where id = $1 for update nowait`, attemptID).Scan(&id)
}
}
func lockedReservation(key string) func(context.Context, pgx.Tx) error {
return func(ctx context.Context, tx pgx.Tx) error {
var id string
return tx.QueryRow(ctx,
`select engine_run_id from reservations where engine_run_id = $1 for update nowait`, key).Scan(&id)
}
}
// Every path that touches a run's book and one of the rows below takes them in the package's order,
// asserted directly rather than sampled: a stress test can only fail to observe an inversion.
func TestTheMaterializerTakesTheBookBeforeTheAttempt(t *testing.T) {
s, sink, _, run := sinkFixture(t)
ctx := t.Context()
body, err := json.Marshal(ingest.Progress{Draft: ingest.Counter{Done: 1, Total: 50}})
if err != nil {
t.Fatal(err)
}
assertBookFirst(t, s, ctx, "bk1", lockedAttempt(run.AttemptID), func() error {
return sink.Apply(ctx, ingest.Envelope{Seq: 2, Type: ingest.TypeProgress, Data: body}, ingest.Cursor{Offset: 20})
})
}
func TestARestartTakesTheBookBeforeTheAttempt(t *testing.T) {
s, _, runID, run := sinkFixture(t)
ctx := t.Context()
assertBookFirst(t, s, ctx, "bk1", lockedAttempt(run.AttemptID), func() error {
_, err := s.RestartRun(ctx, RestartInput{RunID: runID, AttemptID: run.AttemptID,
UserID: "u1", BookID: "bk1", Ceiling: money.MicroUSD(1_000), Now: time.Now().UTC()})
return err
})
}
// Deleting a book takes the book before the reservations it clears. No cycle exists for this pair
// today — the delete only touches CLOSED reservations, and nothing else locks those — but an
// invariant with an undeclared exception stops being one, and the next writer of this package reads
// the invariant, not the exception.
func TestDeletingABookTakesItBeforeItsReservations(t *testing.T) {
s, ctx := testDB(t)
now := fundedAccount(t, s, ctx, "u1", "10")
seedBook(t, s, ctx, "bk1", "u1", 500)
if err := s.Hold(ctx, "u1", "bk1", "run_x#1", money.MicroUSD(300_000), now); err != nil {
t.Fatal(err)
}
if err := s.Release(ctx, "run_x#1", now); err != nil {
t.Fatal(err)
}
assertBookFirst(t, s, ctx, "bk1", lockedReservation("run_x#1"), func() error {
return s.DeleteBook(ctx, "bk1")
})
}
// The materializer and the reconciler work on one run at the same time as a matter of course — a
// live stream while a sweep ends an attempt — and they touch the same two rows. Taking them in
// opposite orders is a deadlock Postgres resolves by aborting a side, which costs a sweep or an API
// call and, worse, used to reach the reconciler as "this journal cannot be read" and quarantine the
// projection of a paying run for good.
//
// ⚠ Measured before the order was made global (lockBook): 258 of 300 concurrent pairs aborted.
func TestAMaterializerAndAReconcilerOnOneRunDoNotDeadlock(t *testing.T) {
s, sink, runID, run := sinkFixture(t)
const pairs = 60
var wg sync.WaitGroup
errs := make(chan error, 2*pairs)
body, err := json.Marshal(ingest.Progress{Draft: ingest.Counter{Done: 1, Total: 50}})
if err != nil {
t.Fatal(err)
}
ctx := t.Context()
now := time.Now().UTC()
// The restart path takes the same two rows and used to take them attempt-first as well. Its own
// LOGICAL outcome is not what this test is about — only the first of these closes the attempt and
// the rest legitimately find nothing — so its non-transient errors are tolerated and only a broken
// lock is not.
restartErrs := make(chan error, pairs)
for i := range pairs {
wg.Add(3)
go func() {
defer wg.Done()
seq := int64(i) + 2
errs <- sink.Apply(ctx, ingest.Envelope{Seq: seq, Type: ingest.TypeProgress, Data: body},
ingest.Cursor{Offset: seq * 10})
}()
go func() {
defer wg.Done()
_, err := s.FinishRun(ctx, runID, run.AttemptID, "ready", "exit-code", nil, now)
errs <- err
}()
go func() {
defer wg.Done()
_, err := s.RestartRun(ctx, RestartInput{RunID: runID, AttemptID: run.AttemptID,
UserID: "u1", BookID: "bk1", Ceiling: money.MicroUSD(1_000), Now: now})
restartErrs <- err
}()
}
wg.Wait()
close(errs)
close(restartErrs)
for err := range errs {
if IsTransient(err) {
t.Fatalf("the materializer and the reconciler deadlocked: %v", err)
}
if err != nil {
t.Errorf("concurrent write: %v", err)
}
}
for err := range restartErrs {
if IsTransient(err) {
t.Fatalf("the materializer and a restart deadlocked: %v", err)
}
}
}
// A DRAFT unit is not a finished unit: `units_done` counts a unit only once its edit resolved, which
// is the definition the whole per-phase progress rests on.
func TestADraftUnitMovesTheDraftCounterAndNotTheDoneOne(t *testing.T) {
s, sink, _, _ := sinkFixture(t)
if err := apply(t, sink, 2, ingest.TypeUnitDone,
ingest.UnitDone{Chapter: 1, Unit: 1, Wave: "draft", Shipped: true}); err != nil {
t.Fatal(err)
}
var draft, edit, done int
if err := s.pool.QueryRow(t.Context(),
`select units_draft_done, units_edit_done, units_done from chapters where id='c1'`).
Scan(&draft, &edit, &done); err != nil {
t.Fatal(err)
}
if draft != 1 || edit != 0 || done != 0 {
t.Errorf("after a draft unit: draft=%d edit=%d done=%d", draft, edit, done)
}
}
// A ceiling event that did NOT halt the run must not pause it.
func TestACeilingEventThatDidNotHaltChangesNothing(t *testing.T) {
s, sink, runID, _ := sinkFixture(t)
if err := apply(t, sink, 2, ingest.TypeCeiling, ingest.Ceiling{Halted: false}); err != nil {
t.Fatal(err)
}
var status, reason string
if err := s.pool.QueryRow(t.Context(),
`select status, coalesce(paused_reason,'') from runs where id=$1`, runID).Scan(&status, &reason); err != nil {
t.Fatal(err)
}
if status != "translating" || reason != "" {
t.Errorf("a ceiling event with halted=false gave status %q reason %q", status, reason)
}
}
// The bank stop arriving on the STREAM is the same product fact as exit code 3, and the exit path is
// pinned elsewhere; this is the other half.
func TestABankStopOnTheStreamPutsTheRunInAwaitingBank(t *testing.T) {
s, sink, runID, _ := sinkFixture(t)
if err := apply(t, sink, 2, ingest.TypeBankStop, ingest.BankStop{TermsProposed: 7}); err != nil {
t.Fatal(err)
}
var status string
if err := s.pool.QueryRow(t.Context(), `select status from runs where id=$1`, runID).Scan(&status); err != nil {
t.Fatal(err)
}
if status != "awaiting_bank" {
t.Errorf("after a bank stop the run is %q", status)
}
}
// An event with no payload is refused rather than applied as a zero struct: a `progress` line with
// empty data would otherwise write draft 0/0 over a live projection.
func TestAnEventWithNoPayloadIsRefused(t *testing.T) {
_, sink, _, _ := sinkFixture(t)
err := sink.Apply(t.Context(), ingest.Envelope{Seq: 2, Type: ingest.TypeProgress}, ingest.Cursor{})
if err == nil {
t.Fatal("an event with no data was applied")
}
}
// An ETA of zero is ABSENT, not zero: the screen renders without an estimate rather than showing
// "0 s left".
func TestAZeroEtaIsStoredAsAbsent(t *testing.T) {
s, sink, runID, _ := sinkFixture(t)
if err := apply(t, sink, 2, ingest.TypeProgress, ingest.Progress{
Draft: ingest.Counter{Done: 1, Total: 10}, ETASeconds: 0}); err != nil {
t.Fatal(err)
}
var eta *int
if err := s.pool.QueryRow(t.Context(), `select eta_seconds from runs where id=$1`, runID).Scan(&eta); err != nil {
t.Fatal(err)
}
if eta != nil {
t.Errorf("an ETA of zero was stored as %d", *eta)
}
}
// The resync materializes the engine's PER-WAVE split and never LOWERS a counter the stream has
// already moved further.
//
// ⚠ This replaces a pin whose premise the engine retired: while status carried only one end-to-end
// aggregate (backlog row 99) the resync could put it in the edit counter and nothing else, and the
// old test asserted that the draft counter stayed untouched. The split landed on 09.08 (D39.122), so
// asserting the old shape would now pin a limitation instead of a property.
func TestTheResyncMaterializesThePhaseSplitAndNeverLowersACounter(t *testing.T) {
s, sink, runID, _ := sinkFixture(t)
if err := apply(t, sink, 2, ingest.TypeProgress, ingest.Progress{
Draft: ingest.Counter{Done: 9, Total: 20}, Edit: ingest.Counter{Done: 7, Total: 20}}); err != nil {
t.Fatal(err)
}
// A report taken BEFORE the engine's own figures caught up, folded over a projection the stream
// has already moved: nothing may walk backwards.
stale := ingest.StatusReport{TotalUnits: 20, Done: 3,
Progress: ingest.Progress{Draft: ingest.Counter{Done: 5, Total: 20}, Edit: ingest.Counter{Done: 3, Total: 20}}}
if err := s.ApplyStatus(t.Context(), runID, "bk1", stale, time.Now().UTC()); err != nil {
t.Fatal(err)
}
var draft, edit int
if err := s.pool.QueryRow(t.Context(),
`select draft_done, edit_done from runs where id=$1`, runID).Scan(&draft, &edit); err != nil {
t.Fatal(err)
}
if draft != 9 || edit != 7 {
t.Errorf("a stale resync walked the projection back to draft=%d edit=%d, from draft=9 edit=7", draft, edit)
}
// A fresher one moves BOTH waves, which is the half no aggregate could do.
fresh := ingest.StatusReport{TotalUnits: 20, Done: 11,
Progress: ingest.Progress{Draft: ingest.Counter{Done: 20, Total: 20}, Edit: ingest.Counter{Done: 11, Total: 20}}}
if err := s.ApplyStatus(t.Context(), runID, "bk1", fresh, time.Now().UTC()); err != nil {
t.Fatal(err)
}
var dtotal, etotal int
if err := s.pool.QueryRow(t.Context(),
`select draft_done, draft_total, edit_done, edit_total from runs where id=$1`, runID).
Scan(&draft, &dtotal, &edit, &etotal); err != nil {
t.Fatal(err)
}
if draft != 20 || dtotal != 20 || edit != 11 || etotal != 20 {
t.Errorf("after a fresher resync: draft %d/%d, edit %d/%d", draft, dtotal, edit, etotal)
}
}
// What a resync produces is stale by up to the poll interval, and the ONLY thing that says how stale
// is the stamp it leaves. A run being tailed has none: the stream carries its own freshness.
func TestAResyncRecordsWhenItWasTaken(t *testing.T) {
s, _, runID, _ := sinkFixture(t)
var stamp *time.Time
if err := s.pool.QueryRow(t.Context(), `select last_resync_at from runs where id=$1`, runID).Scan(&stamp); err != nil {
t.Fatal(err)
}
if stamp != nil {
t.Fatalf("a run that was never resynced is stamped %s", stamp)
}
now := time.Now().UTC().Truncate(time.Millisecond)
if err := s.ApplyStatus(t.Context(), runID, "bk1", ingest.StatusReport{TotalUnits: 20}, now); err != nil {
t.Fatal(err)
}
if err := s.pool.QueryRow(t.Context(), `select last_resync_at from runs where id=$1`, runID).Scan(&stamp); err != nil {
t.Fatal(err)
}
if stamp == nil || !stamp.UTC().Equal(now) {
t.Fatalf("the resync stamped %v, want %s", stamp, now)
}
// And EVERY resync moves it. A stamp that only records the first one answers "when did polling
// begin", which is not the question a reader of a stale projection is asking.
later := now.Add(7 * time.Minute)
if err := s.ApplyStatus(t.Context(), runID, "bk1", ingest.StatusReport{TotalUnits: 20}, later); err != nil {
t.Fatal(err)
}
if err := s.pool.QueryRow(t.Context(), `select last_resync_at from runs where id=$1`, runID).Scan(&stamp); err != nil {
t.Fatal(err)
}
if stamp == nil || !stamp.UTC().Equal(later) {
t.Errorf("the second resync left the stamp at %v, want %s", stamp, later)
}
}
// Cross-family review of the acceptance dofix (M2): a stop that lands before its engine was spawned
// leaves an attempt that is ENDED and carries no engine run id — a shape this zone could not produce
// before the stop path existed. An old materializer still holding that attempt would then bind it to
// the handshake of the attempt that REPLACED it and apply the same journal a second time, counting
// every unit and chapter twice. Binding is refused for an attempt that is over.
func TestAMaterializerOfAnEndedAttemptDoesNotAdoptTheNextAttemptsEngine(t *testing.T) {
s, ctx := testDB(t)
now := fundedAccount(t, s, ctx, "u1", "10")
seedBook(t, s, ctx, "bk1", "u1", 500)
run, err := s.StartRun(ctx, StartRunInput{UserID: "u1", BookID: "bk1", CeilingChapters: 10,
Ceiling: money.MicroUSD(300_000), Now: now}, 0, nil)
if err != nil {
t.Fatal(err)
}
// The stop-before-spawn shape: the attempt is closed without ever having been bound.
exec(t, s, ctx, `update run_attempts set ended_at = now() where id = $1`, run.AttemptID)
stale := s.NewRunSink(run.AttemptID, run.ID, "bk1")
if err := stale.Begin(ctx, ingest.Hello{EngineRunID: "eng-of-the-next-attempt"}); err == nil {
t.Fatal("an ended attempt adopted an engine run: its journal would be materialized twice")
}
var bound *string
if err := s.pool.QueryRow(ctx,
`select engine_run_id from run_attempts where id = $1`, run.AttemptID).Scan(&bound); err != nil {
t.Fatal(err)
}
if bound != nil {
t.Fatalf("the ended attempt was bound to %q", *bound)
}
}