Files
mesh-catalog/modules/nats/snapshot/snapshot.go
T
jochen 419e82cded Back up the bus by the server's own snapshot of each stream, not its live files (hq ADR 0235)
The restic holder copied JetStream's store while the server wrote it; such a
copy may not restore. The nats image now carries mesh-nats-snapshot, run by
the declared dump under the module's own bus account (snapshot API only):
every stream one at a time, flow-controlled, into one tar with a manifest of
counts, sequences and checksums. Restore builds a new store beside the live
one with the bus's own server; a person swaps it in. Proven against
throwaway nats 2.11 servers being written to during the snapshot.
2026-10-06 18:20:51 +02:00

447 lines
15 KiB
Go

package main
import (
"archive/tar"
"context"
"crypto/sha256"
"encoding/hex"
"encoding/json"
"errors"
"fmt"
"io"
"os"
"sort"
"strings"
"sync"
"time"
"github.com/klauspost/compress/s2"
"github.com/nats-io/nats.go"
)
// The JetStream API's subjects this program speaks. Each is granted to the bus's own module and
// nothing else is (mesh-controller internal/broker, BusSnapshotGrants); a test holds the two lists
// equal against a real server.
const (
apiStreamNames = "$JS.API.STREAM.NAMES"
apiStreamInfo = "$JS.API.STREAM.INFO."
apiStreamSnapshot = "$JS.API.STREAM.SNAPSHOT."
apiStreamRestore = "$JS.API.STREAM.RESTORE."
)
// ChunkSize is what the server is asked to send at a time: its own default. The server sends up to
// an 8 MiB window ahead of the acknowledgements and waits for them past it, so one stream costs the
// server at most that much in flight, whatever its size.
const ChunkSize = 128 * 1024
// SnapshotOptions bounds a snapshot.
type SnapshotOptions struct {
// StreamTimeout is the longest one stream may take; past it the snapshot fails, naming it.
StreamTimeout time.Duration
// Pause is a rest between two streams, so a night's snapshot is never one long burst.
Pause time.Duration
// TempDir is where one stream's archive waits until it is written out; the system's when empty.
TempDir string
Log func(format string, args ...any)
}
// Manifest is the first entry of an archive: what was taken, when, how long it took, and the checksum
// of every file beside it. It is what `verify` and `restore` hold the archive to.
type Manifest struct {
Format int `json:"format"`
Server string `json:"server_version"`
Started time.Time `json:"started"`
Finished time.Time `json:"finished"`
Seconds float64 `json:"seconds"`
// Messages and StreamBytes are the streams' own totals at their snapshots; ArchiveBytes what the
// archives take, compressed.
Messages uint64 `json:"messages"`
StreamBytes int64 `json:"stream_bytes"`
ArchiveBytes int64 `json:"archive_bytes"`
Streams []StreamEntry `json:"streams"`
// Skipped are streams that hold nothing across a restart (memory storage), said rather than
// silently left out.
Skipped []Skipped `json:"skipped,omitempty"`
}
// StreamEntry is one stream as it was snapshotted.
type StreamEntry struct {
Name string `json:"name"`
Subjects []string `json:"subjects,omitempty"`
Retention string `json:"retention,omitempty"`
Messages uint64 `json:"messages"`
Bytes uint64 `json:"bytes"`
FirstSeq uint64 `json:"first_seq"`
LastSeq uint64 `json:"last_seq"`
Deleted int `json:"deleted,omitempty"`
Consumers int `json:"consumers"`
// Meta is the server's own description of the stream (its configuration and state), in the
// shape the restore API and the nats CLI take back: backup.json.
Meta string `json:"meta"`
MetaSHA256 string `json:"meta_sha256"`
// Archive is the server's snapshot of the stream, as the server sent it: stream.tar.s2.
Archive string `json:"archive"`
ArchiveBytes int64 `json:"archive_bytes"`
ArchiveSHA256 string `json:"archive_sha256"`
Seconds float64 `json:"seconds"`
}
// Skipped is a stream not snapshotted, and why.
type Skipped struct {
Name string `json:"name"`
Why string `json:"why"`
}
// ManifestName is the archive's first entry.
const ManifestName = "manifest.json"
// apiError is the error every JetStream API answer may carry.
type apiError struct {
Code int `json:"code"`
ErrCode int `json:"err_code"`
Description string `json:"description"`
}
func (e *apiError) Error() string { return fmt.Sprintf("%s (%d)", e.Description, e.Code) }
// streamState is what this program reads of a stream's state; the whole of it is kept as the server
// gave it, in the stream's meta.
type streamState struct {
Messages uint64 `json:"messages"`
Bytes uint64 `json:"bytes"`
FirstSeq uint64 `json:"first_seq"`
LastSeq uint64 `json:"last_seq"`
NumDelete int `json:"num_deleted"`
Consumers int `json:"consumer_count"`
}
type streamConfig struct {
Name string `json:"name"`
Subjects []string `json:"subjects"`
Retention string `json:"retention"`
Storage string `json:"storage"`
}
// request asks the JetStream API and decodes its answer, refusing an answer that carries an error.
func request(ctx context.Context, nc *nats.Conn, subject string, body any, into any) error {
var payload []byte
if body != nil {
var err error
if payload, err = json.Marshal(body); err != nil {
return err
}
}
rctx, cancel := context.WithTimeout(ctx, 30*time.Second)
defer cancel()
msg, err := nc.RequestWithContext(rctx, subject, payload)
if err != nil {
if errors.Is(err, context.DeadlineExceeded) || errors.Is(err, nats.ErrTimeout) || errors.Is(err, nats.ErrNoResponders) {
return fmt.Errorf("%s was not answered (%v): the bus refuses a subject this user is not granted by not answering it", subject, err)
}
return fmt.Errorf("%s: %w", subject, err)
}
var wrapped struct {
Error *apiError `json:"error"`
}
if err := json.Unmarshal(msg.Data, &wrapped); err != nil {
return fmt.Errorf("%s answered something that is not JSON: %q", subject, truncate(msg.Data))
}
if wrapped.Error != nil {
return fmt.Errorf("%s: %w", subject, wrapped.Error)
}
if into != nil {
return json.Unmarshal(msg.Data, into)
}
return nil
}
func truncate(b []byte) string {
if len(b) > 200 {
return string(b[:200]) + "…"
}
return string(b)
}
// StreamNames is every stream in the account, in name order, all pages of them.
func StreamNames(ctx context.Context, nc *nats.Conn) ([]string, error) {
var names []string
for {
var page struct {
Total int `json:"total"`
Offset int `json:"offset"`
Limit int `json:"limit"`
Streams []string `json:"streams"`
}
if err := request(ctx, nc, apiStreamNames, map[string]int{"offset": len(names)}, &page); err != nil {
return nil, err
}
names = append(names, page.Streams...)
if len(page.Streams) == 0 || len(names) >= page.Total {
break
}
}
sort.Strings(names)
return names, nil
}
// Snapshot writes every stream, one at a time, as a tar: the manifest first, then each stream's
// meta and archive under streams/<name>/. Any stream failing fails the whole: a partial copy of the
// bus written where a whole one is expected is the silent failure this exists to prevent.
func Snapshot(ctx context.Context, nc *nats.Conn, out io.Writer, o SnapshotOptions) (*Manifest, error) {
if o.StreamTimeout <= 0 {
o.StreamTimeout = 10 * time.Minute
}
if o.Log == nil {
o.Log = func(string, ...any) {}
}
m := &Manifest{Format: 1, Server: nc.ConnectedServerVersion(), Started: time.Now().UTC()}
names, err := StreamNames(ctx, nc)
if err != nil {
return nil, err
}
if len(names) == 0 {
// The mesh's own streams always exist once the controller has run; a bus with none is not
// one to call backed up.
return nil, errors.New("the bus holds no streams; a snapshot of nothing would pass for a backup")
}
type taken struct {
entry StreamEntry
meta []byte
archive *os.File
}
var files []taken
defer func() {
for _, t := range files {
t.archive.Close()
os.Remove(t.archive.Name())
}
}()
for i, name := range names {
if i > 0 && o.Pause > 0 {
select {
case <-time.After(o.Pause):
case <-ctx.Done():
return nil, ctx.Err()
}
}
var info struct {
Config streamConfig `json:"config"`
}
if err := request(ctx, nc, apiStreamInfo+name, nil, &info); err != nil {
return nil, err
}
if info.Config.Storage == "memory" {
m.Skipped = append(m.Skipped, Skipped{Name: name, Why: "memory storage: it holds nothing across a restart, and the server cannot snapshot it"})
o.Log("%s: skipped, memory storage", name)
continue
}
f, err := os.CreateTemp(o.TempDir, "mesh-nats-snapshot-*")
if err != nil {
return nil, err
}
files = append(files, taken{archive: f})
sctx, cancel := context.WithTimeout(ctx, o.StreamTimeout)
entry, meta, err := snapshotStream(sctx, nc, name, f)
cancel()
if err != nil {
return nil, fmt.Errorf("stream %s: %w", name, err)
}
files[len(files)-1].entry, files[len(files)-1].meta = entry, meta
m.Messages += entry.Messages
m.StreamBytes += int64(entry.Bytes)
m.ArchiveBytes += entry.ArchiveBytes
o.Log("%s: %d message(s), %s, last sequence %d, %d consumer(s) → %s in %.2fs",
name, entry.Messages, human(int64(entry.Bytes)), entry.LastSeq, entry.Consumers, human(entry.ArchiveBytes), entry.Seconds)
}
m.Finished = time.Now().UTC()
m.Seconds = m.Finished.Sub(m.Started).Seconds()
for _, t := range files {
m.Streams = append(m.Streams, t.entry)
}
tw := tar.NewWriter(out)
manifest, err := json.MarshalIndent(m, "", " ")
if err != nil {
return nil, err
}
if err := writeEntry(tw, ManifestName, append(manifest, '\n'), m.Finished); err != nil {
return nil, err
}
for _, t := range files {
if err := writeEntry(tw, t.entry.Meta, t.meta, m.Finished); err != nil {
return nil, err
}
if _, err := t.archive.Seek(0, io.SeekStart); err != nil {
return nil, err
}
hdr := &tar.Header{Name: t.entry.Archive, Mode: 0o600, Size: t.entry.ArchiveBytes, ModTime: m.Finished, Format: tar.FormatPAX}
if err := tw.WriteHeader(hdr); err != nil {
return nil, err
}
if _, err := io.Copy(tw, t.archive); err != nil {
return nil, err
}
}
if err := tw.Close(); err != nil {
return nil, err
}
return m, nil
}
func writeEntry(tw *tar.Writer, name string, body []byte, at time.Time) error {
if err := tw.WriteHeader(&tar.Header{Name: name, Mode: 0o600, Size: int64(len(body)), ModTime: at, Format: tar.FormatPAX}); err != nil {
return err
}
_, err := tw.Write(body)
return err
}
// snapshotStream takes one stream's snapshot into f: the server's chunks, each acknowledged as it
// arrives so the server's window keeps moving, until the server's last, empty message. It answers
// the stream's entry (the archive's size and checksum included) and its meta.
//
// The protocol is the server's (nats-server jetstream_api.go, streamSnapshot): the request names a
// subject to deliver to; the server answers with the stream's configuration and its state at the
// snapshot, then sends the archive in chunks whose reply subject is the acknowledgement it waits for
// past its window — two seconds without one and it gives up with "408 No Flow Response". The end is
// an empty message whose status header says 204, or the error that ended it.
func snapshotStream(ctx context.Context, nc *nats.Conn, name string, f *os.File) (StreamEntry, []byte, error) {
started := time.Now()
deliver := nc.NewRespInbox()
hash := sha256.New()
var written int64
var mu sync.Mutex
var failed error
done := make(chan struct{})
var once sync.Once
finish := func(err error) {
mu.Lock()
if failed == nil {
failed = err
}
mu.Unlock()
once.Do(func() { close(done) })
}
sub, err := nc.Subscribe(deliver, func(msg *nats.Msg) {
if len(msg.Data) == 0 {
if status := msg.Header.Get("Status"); status != "" && status != "204" {
finish(fmt.Errorf("the server ended the snapshot: %s %s", status, msg.Header.Get("Description")))
return
}
finish(nil)
return
}
// Acknowledged before it is written: the chunk is already in memory, and a slow disk must not
// read to the server as a reader that went away.
if msg.Reply != "" {
if err := nc.Publish(msg.Reply, nil); err != nil {
finish(fmt.Errorf("acknowledging a chunk: %w", err))
return
}
}
if _, err := f.Write(msg.Data); err != nil {
finish(fmt.Errorf("writing the archive: %w", err))
return
}
hash.Write(msg.Data)
written += int64(len(msg.Data))
})
if err != nil {
return StreamEntry{}, nil, err
}
defer sub.Unsubscribe()
// Every chunk is held until the callback has it; the window bounds how many there can be.
if err := sub.SetPendingLimits(-1, -1); err != nil {
return StreamEntry{}, nil, err
}
if err := nc.FlushWithContext(ctx); err != nil {
return StreamEntry{}, nil, err
}
var resp struct {
Config json.RawMessage `json:"config"`
State json.RawMessage `json:"state"`
}
req := map[string]any{"deliver_subject": deliver, "chunk_size": ChunkSize}
if err := request(ctx, nc, apiStreamSnapshot+name, req, &resp); err != nil {
return StreamEntry{}, nil, err
}
select {
case <-done:
case <-ctx.Done():
return StreamEntry{}, nil, fmt.Errorf("not finished within its bound: %w", ctx.Err())
}
// The subscription is drained before the archive is read back, so no callback is still writing.
_ = sub.Unsubscribe()
mu.Lock()
err = failed
mu.Unlock()
if err != nil {
return StreamEntry{}, nil, err
}
if written == 0 {
return StreamEntry{}, nil, errors.New("the server sent an empty archive")
}
if err := f.Sync(); err != nil {
return StreamEntry{}, nil, err
}
var cfg streamConfig
var state streamState
if err := json.Unmarshal(resp.Config, &cfg); err != nil {
return StreamEntry{}, nil, fmt.Errorf("its configuration: %w", err)
}
if err := json.Unmarshal(resp.State, &state); err != nil {
return StreamEntry{}, nil, fmt.Errorf("its state: %w", err)
}
// The archive reads to its end — s2 frames, then a tar — before it is called a snapshot.
if _, err := f.Seek(0, io.SeekStart); err != nil {
return StreamEntry{}, nil, err
}
if err := readsThrough(f); err != nil {
return StreamEntry{}, nil, fmt.Errorf("the archive the server sent does not read: %w", err)
}
// The meta, in the shape the restore API takes and the nats CLI writes (backup.json): the
// server's own configuration and state, kept as raw JSON so a field this program does not know
// is never dropped on the way back.
meta, err := json.MarshalIndent(map[string]json.RawMessage{"config": resp.Config, "state": resp.State}, "", " ")
if err != nil {
return StreamEntry{}, nil, err
}
metaSum := sha256.Sum256(meta)
return StreamEntry{
Name: name, Subjects: cfg.Subjects, Retention: cfg.Retention,
Messages: state.Messages, Bytes: state.Bytes, FirstSeq: state.FirstSeq, LastSeq: state.LastSeq,
Deleted: state.NumDelete, Consumers: state.Consumers,
Meta: "streams/" + name + "/backup.json", MetaSHA256: hex.EncodeToString(metaSum[:]),
Archive: "streams/" + name + "/stream.tar.s2", ArchiveBytes: written,
ArchiveSHA256: hex.EncodeToString(hash.Sum(nil)), Seconds: time.Since(started).Seconds(),
}, meta, nil
}
// readsThrough is whether a stream's archive decompresses and lists to its end.
func readsThrough(r io.Reader) error {
tr := tar.NewReader(s2.NewReader(r))
for {
_, err := tr.Next()
if err == io.EOF {
return nil
}
if err != nil {
return err
}
if _, err := io.Copy(io.Discard, tr); err != nil {
return err
}
}
}
// validName is a stream name that is one safe path component.
func validName(name string) bool {
return name != "" && name != "." && name != ".." && !strings.ContainsAny(name, "/\\\x00 *>")
}