[Performance] Streaming persistence rewrites the full message on every delta

Offen
#4,569 2 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 1 Tag

@akushonkamen arbeitet bereits daran.

Seit 13.10.2026.

  • #4575 von @akushonkamen — offen

Bewertung

Schwierigkeit
4/5
Geschätzter Aufwand
3-5 Tage
Anfängerfreundlichkeit
42/100
Issue-Typ
Bug
Klarheit
Klar beschrieben
Aktivitätsstatus
Aktiv
Tech-Stack
go, sqlite

Rechercherichtung

Start with pkg/runtime/persistence_observer.go and pkg/session/store.go to trace how each delta reaches the SQLite update; read pkg/runtime/observer.go to understand synchronous event delivery. Use the supplied fake-provider probe to reproduce the write growth, then add regression coverage for bounded checkpoint writes, final persisted content, completion, and cancellation. Done means persistence reduces repeated full-message writes while preserving the stated recovery behavior.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

area/runtime area/sessions

Severity: medium · Type: performance

Location

Reviewed commit: cd65d7c776dda80d554735ee20572d09cfe2d791.

Issue details

Incremental persistence repeatedly serializes the growing message, yielding quadratic cumulative payload volume for fixed-size deltas. The SQLite adapter executes one full-row JSON update per delta, on the event-delivery path. This buys fine-grained recovery, but long responses and reasoning streams pay for all their previous content again on every chunk.

A fake provider plus counting store exercised the actual public runtime and its stock persistence observer, with no model API requests:

16-byte text chunks Final text bytes Assistant writes (including finalization) Cumulative text bytes passed to store JSON bytes serialized with the same json.Marshal(msg.Message) as SQLite
1,024 16,384 1,025 8,413,184 8,500,393
2,048 32,768 2,049 33,603,584 33,777,833
4,096 65,536 4,097 134,316,032 134,664,361

Doubling the final response roughly quadruples cumulative payload volume. These numbers measure the observer/store contract and JSON serialization volume; they are not a disk-I/O or end-to-end latency benchmark. The SQLite source establishes that each update serializes and sends the full message to an SQL update.

Save the following as /tmp/persistence-probe.go and run go run /tmp/persistence-probe.go from the repository root:

package main

import (
	"context"
	"encoding/json"
	"fmt"
	"io"
	"strings"

	"github.com/docker/docker-agent/pkg/agent"
	"github.com/docker/docker-agent/pkg/chat"
	"github.com/docker/docker-agent/pkg/config/latest"
	"github.com/docker/docker-agent/pkg/model/provider/contracts"
	"github.com/docker/docker-agent/pkg/modelsdev"
	"github.com/docker/docker-agent/pkg/runtime"
	"github.com/docker/docker-agent/pkg/session"
	"github.com/docker/docker-agent/pkg/team"
	"github.com/docker/docker-agent/pkg/tools"
)

type countingStore struct {
	session.Store
	writes, contentBytes, jsonBytes int
}

func (s *countingStore) AddMessage(ctx context.Context, id string, msg *session.Message) (int64, error) {
	s.record(msg)
	return s.Store.AddMessage(ctx, id, msg)
}
func (s *countingStore) UpdateMessage(ctx context.Context, sid string, id int64, msg *session.Message) error {
	s.record(msg)
	return s.Store.UpdateMessage(ctx, sid, id, msg)
}
func (s *countingStore) record(msg *session.Message) {
	if msg.Message.Role != chat.MessageRoleAssistant { return }
	s.writes++
	s.contentBytes += len(msg.Message.Content)
	b, err := json.Marshal(msg.Message)
	if err != nil { panic(err) }
	s.jsonBytes += len(b)
}

type fakeProvider struct { n int }
func (p fakeProvider) ID() modelsdev.ID { return modelsdev.NewID("test", "mock") }
func (p fakeProvider) BaseConfig() contracts.Config {
	return contracts.Config{ModelConfig: latest.ModelConfig{Provider: "test", Model: "mock"}}
}
func (p fakeProvider) CreateChatCompletionStream(context.Context, []chat.Message, []tools.Tool) (chat.MessageStream, error) {
	return &fakeStream{n: p.n}, nil
}
type fakeStream struct { n, i int }
func (s *fakeStream) Recv() (chat.MessageStreamResponse, error) {
	if s.i > s.n { return chat.MessageStreamResponse{}, io.EOF }
	choice := chat.MessageStreamChoice{}
	if s.i == s.n { choice.FinishReason = chat.FinishReasonStop } else { choice.Delta.Content = strings.Repeat("x", 16) }
	s.i++
	return chat.MessageStreamResponse{Choices: []chat.MessageStreamChoice{choice}}, nil
}
func (*fakeStream) Close() {}
type fakeCatalog struct{}
func (fakeCatalog) GetModel(context.Context, modelsdev.ID) (*modelsdev.Model, error) { return nil, nil }
func (fakeCatalog) GetDatabase(context.Context) (*modelsdev.Database, error) { return &modelsdev.Database{}, nil }

func main() {
	for _, n := range []int{1024, 2048, 4096} {
		ctx := context.Background()
		store := &countingStore{Store: session.NewInMemorySessionStore()}
		a := agent.New("root", "Respond.", agent.WithModel(fakeProvider{n: n}))
		rt, err := runtime.New(ctx, team.New(team.WithAgents(a)), runtime.WithSessionStore(store), runtime.WithModelStore(fakeCatalog{}), runtime.WithSessionCompaction(false))
		if err != nil { panic(err) }
		sess := session.New(session.WithUserMessage("hi"), session.WithNonInteractive(true))
		for ev := range rt.RunStream(ctx, sess) {
			if e, ok := ev.(*runtime.ErrorEvent); ok { panic(e.Error) }
		}
		if err := rt.Close(); err != nil { panic(err) }
		fmt.Printf("chunks=%d final_bytes=%d assistant_writes=%d cumulative_content_bytes=%d serialized_json_bytes=%d\n", n, n*16, store.writes, store.contentBytes, store.jsonBytes)
	}
}
Impact

Large text/reasoning replies cause avoidable serialization, allocations and database updates. Because persistence is synchronous, slower storage also delays delivery to the terminal/UI and downstream observers. Concurrent sessions amplify pressure on the store.

Recommended fix

Coalesce persistence updates by time and/or added byte count, with bounded buffering and a guaranteed final flush on message completion and lifecycle teardown. Preserve partial recovery semantics explicitly. Alternatively store append-only deltas with periodic snapshots.

Add a fake-provider/counting-store regression that bounds checkpoint writes as chunk count grows, verifies the final persisted content, and covers completion/cancellation. Benchmark the SQLite-backed path before selecting checkpoint intervals.

Vorherrschende Sprache
Go
Sterne
4.7k
Forks
565
Ø Merge
18 Std. 52 Min.
Gemergte PRs (30 T.)
221

Entwicklungsumgebung

In Codespaces öffnen

Startet den Dev-Container des Projekts im Browser, mit Ihrem eigenen GitHub-Konto.

  • Enthält ein Dockerfile oder eine Docker-Compose-Datei
  • Keine Pull-Request-Vorlage
  • Kein Beitragsleitfaden

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus docker/docker-agent

Alle Issues in docker/docker-agent

Ähnliche Issues

Weitere Issues zu Go

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.