feat: a backup is one uncompressed tar of per-table JSON Lines that stays the same bytes day to day (#6769)

The file is minstrel-<UTC stamp>.tar: manifest.json first, then
tables/<table>.jsonl, rows one per line in primary-key order. The run's
time lives only in the manifest and member headers are fixed, so a
deduplicating off-site copy stores unchanged tables once.

WriteFile writes to a hidden .part file, syncs it and renames it, so a
watched folder never shows half a backup. Read refuses anything that
isn't a whole backup: another format, a newer format version, a cut-off
file, or a member holding fewer rows than the manifest says.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-11-02 13:51:02 -05:00
co-authored by Claude Opus 5.5
parent a67897a6cd
commit 8be69b752a
2 changed files with 416 additions and 0 deletions
+226
View File
@@ -0,0 +1,226 @@
package backup
import (
"archive/tar"
"bytes"
"encoding/json"
"errors"
"fmt"
"io"
"os"
"path/filepath"
"slices"
"strings"
"time"
)
// The backup file is laid out so that what didn't change between two days is
// the same bytes both times, which suits a deduplicating off-site copy such as
// Fabled Archives' restic:
//
// - Never compressed. Compression rewrites every byte after the first
// change, so no chunk would match again.
// - One member per table, tables/<table>.jsonl: one row per line, in
// primary-key order, keys sorted, no spacing.
// - What changes every run lives only in manifest.json, the first member.
// Member headers are fixed (mtime 0, owner 0, mode 0644).
// - A new dated name each run, so the target keeps a history and a sender
// only ever adds.
const (
FormatName = "minstrel-backup-tar"
FormatVersion = 1
manifestName = "manifest.json"
tablesDir = "tables/"
tableSuffix = ".jsonl"
namePrefix = "minstrel-"
nameSuffix = ".tar"
// maxMember bounds any one member a read takes in: far past the largest
// library's tracks table, small enough that a hostile file can't fill
// memory.
maxMember = 1 << 30
)
// Manifest is the file's first member.
type Manifest struct {
Format string `json:"format"`
FormatVersion int `json:"format_version"`
CreatedAt time.Time `json:"created_at"`
// Build is the Minstrel version that wrote the file.
Build string `json:"build"`
// Schema is the database's migration version when the file was written.
// A restore refuses a file from a newer schema than its own.
Schema int64 `json:"schema"`
// Rows is each table's row count, so a read can tell a whole member from
// a cut-off one.
Rows map[string]int `json:"rows"`
}
// FileName is a backup's file name: minstrel-20261102T030000Z.tar, in UTC, so
// names sort by time.
func FileName(at time.Time) string {
return namePrefix + at.UTC().Format("20060102T150405Z") + nameSuffix
}
// IsFileName reports whether a name is one FileName gives.
func IsFileName(name string) bool {
stamp, ok := strings.CutPrefix(name, namePrefix)
if !ok {
return false
}
stamp, ok = strings.CutSuffix(stamp, nameSuffix)
if !ok {
return false
}
_, err := time.Parse("20060102T150405Z", stamp)
return err == nil
}
// Write writes c as a backup file to w and returns its manifest.
func Write(w io.Writer, c Contents, at time.Time, build string) (Manifest, error) {
names := make([]string, 0, len(c.Tables))
m := Manifest{
Format: FormatName,
FormatVersion: FormatVersion,
CreatedAt: at.UTC(),
Build: build,
Schema: c.Schema,
Rows: make(map[string]int, len(c.Tables)),
}
for name, lines := range c.Tables {
names = append(names, name)
m.Rows[name] = len(lines)
}
slices.Sort(names)
manifest, err := json.MarshalIndent(m, "", " ")
if err != nil {
return m, err
}
tw := tar.NewWriter(w)
if err := member(tw, manifestName, append(manifest, '\n')); err != nil {
return m, err
}
for _, name := range names {
var body bytes.Buffer
for _, line := range c.Tables[name] {
body.Write(line)
body.WriteByte('\n')
}
if err := member(tw, tablesDir+name+tableSuffix, body.Bytes()); err != nil {
return m, err
}
}
return m, tw.Close()
}
// WriteFile writes c into dir under its dated name. The bytes go to a hidden
// .part file first, are synced, and only then take the backup's name, so
// anything watching the folder never sees half a backup under a backup's
// name, and a run that fails leaves nothing behind.
func WriteFile(dir string, c Contents, at time.Time, build string) (string, Manifest, error) {
final := filepath.Join(dir, FileName(at))
part := filepath.Join(dir, "."+FileName(at)+".part")
f, err := os.OpenFile(part, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o600)
if err != nil {
return "", Manifest{}, fmt.Errorf("backup: create %s: %w", part, err)
}
m, err := Write(f, c, at, build)
if err == nil {
err = f.Sync()
}
if cerr := f.Close(); err == nil {
err = cerr
}
if err == nil {
err = os.Rename(part, final)
}
if err != nil {
_ = os.Remove(part)
return "", m, fmt.Errorf("backup: write %s: %w", final, err)
}
if d, err := os.Open(dir); err == nil {
_ = d.Sync()
_ = d.Close()
}
return final, m, nil
}
func member(tw *tar.Writer, name string, data []byte) error {
hdr := &tar.Header{
Name: name,
Mode: 0o644,
Size: int64(len(data)),
ModTime: time.Unix(0, 0),
Typeflag: tar.TypeReg,
Format: tar.FormatUSTAR,
}
if err := tw.WriteHeader(hdr); err != nil {
return err
}
_, err := tw.Write(data)
return err
}
// Read reads a backup file, refusing anything that isn't a whole one.
func Read(r io.Reader) (Manifest, Contents, error) {
var m Manifest
c := Contents{Tables: map[string][][]byte{}}
tr := tar.NewReader(r)
first := true
for {
hdr, err := tr.Next()
if errors.Is(err, io.EOF) {
break
}
if err != nil {
return m, c, fmt.Errorf("%w: %v", ErrNotABackup, err)
}
if hdr.Size > maxMember {
return m, c, fmt.Errorf("%w: %s is too large", ErrNotABackup, hdr.Name)
}
data, err := io.ReadAll(io.LimitReader(tr, maxMember))
if err != nil {
return m, c, fmt.Errorf("%w: %v", ErrNotABackup, err)
}
if first {
first = false
if hdr.Name != manifestName || json.Unmarshal(data, &m) != nil || m.Format != FormatName {
return m, c, ErrNotABackup
}
if m.FormatVersion > FormatVersion {
return m, c, fmt.Errorf("%w: it was written by a newer Minstrel (format %d)", ErrNotABackup, m.FormatVersion)
}
c.Schema = m.Schema
continue
}
table, ok := strings.CutPrefix(hdr.Name, tablesDir)
if !ok {
continue
}
if table, ok = strings.CutSuffix(table, tableSuffix); !ok || table == "" {
continue
}
var lines [][]byte
for line := range bytes.SplitSeq(data, []byte("\n")) {
if len(line) == 0 {
continue
}
if !json.Valid(line) {
return m, c, fmt.Errorf("%w: a row in %s is not JSON", ErrNotABackup, table)
}
lines = append(lines, line)
}
c.Tables[table] = lines
}
if first {
return m, c, ErrNotABackup
}
for table, want := range m.Rows {
if got := len(c.Tables[table]); got != want {
return m, c, fmt.Errorf("%w: %s holds %d rows, the manifest says %d", ErrNotABackup, table, got, want)
}
}
return m, c, nil
}
+190
View File
@@ -0,0 +1,190 @@
package backup
import (
"archive/tar"
"bytes"
"errors"
"io"
"os"
"path/filepath"
"testing"
"time"
)
func sample() Contents {
return Contents{
Schema: 98,
Tables: map[string][][]byte{
"users": {[]byte(`{"id":"u1","username":"ada"}`), []byte(`{"id":"u2","username":"bo"}`)},
"tracks": {[]byte(`{"file_path":"/music/a.flac","id":"t1"}`)},
"wants": {},
},
}
}
func members(t *testing.T, file []byte) map[string][]byte {
t.Helper()
out := map[string][]byte{}
tr := tar.NewReader(bytes.NewReader(file))
for {
hdr, err := tr.Next()
if errors.Is(err, io.EOF) {
return out
}
if err != nil {
t.Fatal(err)
}
data, _ := io.ReadAll(tr)
out[hdr.Name] = data
}
}
func TestABackupFileReadsBackAsWritten(t *testing.T) {
var buf bytes.Buffer
at := time.Date(2026, 11, 2, 3, 0, 0, 0, time.UTC)
if _, err := Write(&buf, sample(), at, "2026.11.02.0300"); err != nil {
t.Fatal(err)
}
m, c, err := Read(&buf)
if err != nil {
t.Fatal(err)
}
if m.Schema != 98 || c.Schema != 98 || m.Build != "2026.11.02.0300" || !m.CreatedAt.Equal(at) {
t.Errorf("manifest = %+v", m)
}
want := sample()
for name, lines := range want.Tables {
got := c.Tables[name]
if len(got) != len(lines) {
t.Errorf("%s: %d rows, want %d", name, len(got), len(lines))
continue
}
for i := range lines {
if !bytes.Equal(got[i], lines[i]) {
t.Errorf("%s row %d = %s, want %s", name, i, got[i], lines[i])
}
}
}
}
// Two days over unchanged data differ only in the manifest, so a
// deduplicating copy stores the tables once.
func TestOnlyTheManifestChangesBetweenDays(t *testing.T) {
var day1, day2 bytes.Buffer
if _, err := Write(&day1, sample(), time.Date(2026, 11, 1, 3, 0, 0, 0, time.UTC), "a"); err != nil {
t.Fatal(err)
}
if _, err := Write(&day2, sample(), time.Date(2026, 11, 2, 3, 0, 0, 0, time.UTC), "b"); err != nil {
t.Fatal(err)
}
a, b := members(t, day1.Bytes()), members(t, day2.Bytes())
if len(a) != 4 {
t.Fatalf("members = %d, want the manifest and three tables", len(a))
}
for name, data := range a {
same := bytes.Equal(data, b[name])
if name == manifestName && same {
t.Error("the manifest is identical across days; it should carry the run's time")
}
if name != manifestName && !same {
t.Errorf("%s differs between days over the same rows", name)
}
}
// The table members' bytes, headers included, sit unchanged in both files.
tracks := a[tablesDir+"tracks"+tableSuffix]
if !bytes.Contains(day1.Bytes(), tracks) || !bytes.Contains(day2.Bytes(), tracks) {
t.Error("the tracks member is not stored verbatim")
}
}
func TestReadRefusesWhatIsNotAWholeBackup(t *testing.T) {
var good bytes.Buffer
if _, err := Write(&good, sample(), time.Now(), "x"); err != nil {
t.Fatal(err)
}
tarOf := func(name string, data []byte) []byte {
var buf bytes.Buffer
tw := tar.NewWriter(&buf)
_ = member(tw, name, data)
_ = tw.Close()
return buf.Bytes()
}
short := func() []byte {
// The manifest promises two users; the member holds one.
var buf bytes.Buffer
tw := tar.NewWriter(&buf)
_ = member(tw, manifestName, []byte(`{"format":"minstrel-backup-tar","format_version":1,"rows":{"users":2}}`))
_ = member(tw, tablesDir+"users"+tableSuffix, []byte(`{"id":"u1"}`+"\n"))
_ = tw.Close()
return buf.Bytes()
}
cases := map[string][]byte{
"empty": nil,
"not a tar": []byte("hello"),
"cut off": good.Bytes()[:good.Len()/2],
"no manifest": tarOf(tablesDir+"users"+tableSuffix, []byte(`{"id":"u1"}`)),
"another format": tarOf(manifestName, []byte(`{"format":"fablednotary-backup-tar","format_version":1}`)),
"a newer format": tarOf(manifestName, []byte(`{"format":"minstrel-backup-tar","format_version":99}`)),
"a member short": short(),
}
for name, file := range cases {
if _, _, err := Read(bytes.NewReader(file)); !errors.Is(err, ErrNotABackup) {
t.Errorf("%s: err = %v, want ErrNotABackup", name, err)
}
}
}
func TestWriteFileLandsUnderTheDatedNameOnly(t *testing.T) {
dir := t.TempDir()
at := time.Date(2026, 11, 2, 3, 0, 0, 0, time.UTC)
path, _, err := WriteFile(dir, sample(), at, "x")
if err != nil {
t.Fatal(err)
}
if filepath.Base(path) != "minstrel-20261102T030000Z.tar" {
t.Errorf("path = %s", path)
}
entries, _ := os.ReadDir(dir)
if len(entries) != 1 || entries[0].Name() != filepath.Base(path) {
t.Errorf("the folder holds %v, want only the backup", entries)
}
f, err := os.Open(path)
if err != nil {
t.Fatal(err)
}
defer f.Close()
if _, _, err := Read(f); err != nil {
t.Errorf("the written file doesn't read: %v", err)
}
}
func TestAFailedWriteFileLeavesNothingBehind(t *testing.T) {
dir := filepath.Join(t.TempDir(), "missing")
if _, _, err := WriteFile(dir, sample(), time.Now(), "x"); err == nil {
t.Fatal("writing into a missing folder succeeded")
}
if _, err := os.Stat(dir); !os.IsNotExist(err) {
t.Errorf("the failed write created %s", dir)
}
}
func TestFileNamesSortByTimeAndAreRecognised(t *testing.T) {
earlier := FileName(time.Date(2026, 1, 9, 23, 0, 0, 0, time.UTC))
later := FileName(time.Date(2026, 1, 10, 1, 0, 0, 0, time.FixedZone("x", 5*3600)))
if earlier >= later {
t.Errorf("%s does not sort before %s", earlier, later)
}
for name, want := range map[string]bool{
earlier: true,
"minstrel-20261102T030000Z.tar": true,
".minstrel-20261102T030000Z.tar.part": false,
"minstrel-yesterday.tar": false,
"fablednotary-20261102T030000Z.tar": false,
} {
if IsFileName(name) != want {
t.Errorf("IsFileName(%q) = %v, want %v", name, !want, want)
}
}
}