package main import ( "fmt" "math/rand/v2" "os" "path/filepath" "runtime" "strings" "testing" "time" "github.com/stretchr/testify/require" ) // syntheticMessages is the length of the synthetic log. const syntheticMessages = 100000 // TestReplaySyntheticLog runs a synthetic log through the memory core with the // fake summarizer. After each message it waits for the builds, as a turn does. // It checks that each view stays within its sawtooth, that no view line holds a // whole long message or a part of an unbuilt one, and that a restart in the // middle and at the end loads the same views byte for byte. func TestReplaySyntheticLog(t *testing.T) { start := time.Now() root := t.TempDir() clock := newTestClock() random := rand.New(rand.NewPCG(1, 2)) memory, err := openMemory(t.Context(), root, clock.Now, newFakeSummarizer()) require.NoError(t, err) defer func() { require.NoError(t, memory.Close()) }() batches := 0 low := chatHigh long := make([]bool, syntheticMessages) for id := range syntheticMessages { if id == syntheticMessages/2 { memory = requireRestart(t, memory, root, clock) } clock.now = clock.now.Add(time.Minute) kind, text := syntheticMessage(random, id) long[id] = len(kind)+len(": ")+len(text) > nodeBytes lines := viewLines(memory) _, err = memory.Append(kind, text) require.NoError(t, err) size, merged := requireViews(t, memory, lines, long, id%1000 == 0) if merged { batches++ low = min(low, size) } require.NoError(t, memory.Wait(t.Context())) } memory = requireRestart(t, memory, root, clock) require.Equal(t, syntheticMessages, memory.log.Len()) require.Greater(t, batches, 0) require.NoError(t, memory.Wait(t.Context())) t.Logf("%d messages, %d nodes, %d batches in %s", syntheticMessages, memory.tree.Count(), batches, time.Since(start).Round(time.Second)) t.Logf("chat view: %d lines, %d bytes, lowest after a batch %d bytes", len(memory.chat.lines), memory.chat.Size(), low) t.Logf("compaction view: %d lines, %d bytes", len(memory.cview.lines), memory.cview.Size()) runtime.GC() var stats runtime.MemStats runtime.ReadMemStats(&stats) t.Logf("heap in use: %d KB", stats.HeapInuse/1024) } // syntheticMessage returns the kind and the text of message id. Most messages // are tool calls and their output. Users write short lines. One output in a // hundred is over 30 KB. func syntheticMessage(random *rand.Rand, id int) (string, string) { roll := random.IntN(100) switch { case roll < 5: return "user", syntheticText(id, 10+random.IntN(200)) case roll < 10: return "reply", syntheticText(id, 100+random.IntN(1500)) case roll < 12: return "note", syntheticText(id, 50+random.IntN(600)) case roll < 13: return "work", syntheticText(id, 300+random.IntN(3000)) case roll < 56: return "tool", syntheticText(id, 60+random.IntN(400)) case roll < 57: return "echo", syntheticText(id, 30000+random.IntN(30000)) default: return "echo", syntheticText(id, 20+random.IntN(4000)) } } // syntheticText returns a text of about size bytes that starts with its message // ID and holds some non-ASCII characters. func syntheticText(id, size int) string { head := fmt.Sprintf("m%d ", id) words := "ação data line café " return head + strings.Repeat(words, max(1, (size-len(head))/len(words))) } // requireViews checks the views after a new message, and with full compares // them with the tree. Each view is at most its high limit, and a batch merges // the chat view to its low limit. The leaf of the new message can be built // after the append, so the sizes count it as the pending mark, as the append // did. It returns the size of the chat view and if a batch merged it, from the // number of lines before the message. func requireViews(t *testing.T, memory *Memory, lines int, long []bool, full bool) (int, bool) { t.Helper() memory.mutex.Lock() defer memory.mutex.Unlock() size := appendSize(memory.chat) require.LessOrEqual(t, size, chatHigh) require.LessOrEqual(t, appendSize(memory.cview), compactionHigh) merged := len(memory.chat.lines) < lines if merged { require.LessOrEqual(t, size, chatLow) } requireOnlySummaries(t, memory, long, full) return size, merged } // requireOnlySummaries checks each line of the chat view as the model reads it. // A built line holds its node, a long message has a summary of the fake // summarizer and not its text, and an unbuilt line holds only the pending mark. // With full, it also checks that the view keeps the node of each built line. // The caller holds the lock of memory. func requireOnlySummaries(t *testing.T, memory *Memory, long []bool, full bool) { t.Helper() rendered := strings.Split(memory.chat.Render(), "\n") require.Len(t, rendered, len(memory.chat.lines)+2) for n, line := range memory.chat.lines { head := fmt.Sprintf("%d+%d|", line.index<