Files
orca/internal/storage/replication.go
T
Jon Chery 675feabf0c feat(P09): Syncthing storage replication + conflict resolution (REQ-081; gates C-02, C-14)
P09 — Storage replication via per-namespace Syncthing (R-005).

C-02 spike (.ciagent/C02_SYNCTHING_FEASIBILITY_v0.9.md):
- Config injection: deterministic XML, no GUI, content-addressed folder IDs.
- Conflict policy: flock-style lock + source-wins migration + gc-conflicts.
- Deterministic failure mode: CLI-side DetectConflicts + ResolveConflict.
- Auto-decision: C-02 SATISFIED.

C-14 forced-divergence test (internal/storage/conflict_test.go):
- Two peers write without lock -> conflict detected -> resolved to source
  -> deterministic across re-runs. Unknown source -> nil (no silent winner).
- C-14 SATISFIED.

Replication (internal/storage/replication.go, REQ-081):
- FolderID = sha256(ns+masterKeyFP)[:32] (content-addressed).
- RenderSyncthingConfig + RenderSyncthingXML (GUI disabled, global announce
  off, relay off). DetectConflicts (sorted, deterministic). ResolveConflict
  (source-peer-wins). 97.6% coverage.

Emitter (internal/emitter/syncthing.go):
- SyncthingEmitter renders one config.xml per replicated volume at
  /etc/syncthing/orca-<ns>-<volume>.xml. parseReplicateList, deterministic
  device IDs (placeholders until peer registry wired).

24 packages pass, 20 bats pass, gofmt clean, verify-reqs 90 consistent.

---ci---
project: orca
phase: P09
milestone: v0.9
status: execute
---/ci---
2026-08-05 18:38:49 +00:00

384 lines
14 KiB
Go

// Package storage implements the CLI-side Syncthing replication logic
// for per-namespace storage replication (REQ-081, R-005). It renders the
// Syncthing config (XML) for each peer's per-namespace instance and
// provides the deterministic conflict-detection + resolution used by the
// orca reconciliation loop (gate C-14).
//
// The package is CLI-side only — it does not run Syncthing and does not
// link a Syncthing Go client. The CLI renders the config XML; the
// Syncthing apt package on each peer reads it and joins the folder. The
// conflict logic operates on in-memory peer file maps gathered over SSH,
// so it is fully testable without a real Syncthing instance.
package storage
import (
"crypto/sha256"
"encoding/hex"
"encoding/xml"
"errors"
"fmt"
"sort"
"strings"
)
// VolumeReplication is the volume-level replication declaration derived
// from a jobspec volume entry with a `replicate:` list. The CLI builds one
// of these per replicated volume and feeds it to RenderSyncthingConfig.
//
// - Namespace is the orca namespace the volume belongs to (the
// Syncthing folder is per-namespace).
// - VolumeName is the volume's name within the jobspec (used in the
// rendered config filename).
// - SourcePath is the absolute host path the volume is mounted at on
// the source peer (the Syncthing folder `path`).
// - ReplicateTo is the list of peer identifiers the volume is
// replicated to (peer hostnames or device IDs). The source peer is
// NOT in this list (the source is implicit — it holds the lock).
// - SyncMode is "sendreceive" (default) or "sendonly". Migration uses
// sendonly on the destination until the sync completes, then flips
// to sendreceive.
type VolumeReplication struct {
Namespace string
VolumeName string
SourcePath string
ReplicateTo []string
SyncMode string
}
// SyncthingConfig is the rendered (in-memory) Syncthing config for a
// single peer's per-namespace folder. The XML emitter
// (RenderSyncthingXML) serializes this into the `config.xml` file.
//
// - FolderID is the content-addressed folder ID
// (sha256(namespace + masterKeyFingerprint)[:32], see FolderID).
// - Path is the on-disk path the folder is rooted at (SourcePath).
// - Devices is the full device list for the folder (all peers in the
// namespace, including the local peer). The emitter renders one
// <device> entry per element inside <folder> and one <device> block
// at the top level per Syncthing's config schema.
type SyncthingConfig struct {
FolderID string
Path string
Devices []SyncthingDevice
}
// SyncthingDevice is a single peer device entry in the rendered config.
//
// - ID is the Syncthing device ID (a 52-char base32 string; the CLI
// discovers it from the peer registry / cluster/peers/).
// - Name is the human-readable peer name (the orca node hostname).
// - Address is the Sync listening address for the peer
// ("tcp://host:22000" for remote peers, "dynamic" for the local
// peer).
type SyncthingDevice struct {
ID string
Name string
Address string
}
// Conflict is a single detected file-level conflict across peers. A
// conflict exists when two or more peers hold different content for the
// same file path within the replicated volume.
//
// - Path is the file path relative to the volume root (the same key
// used in the peer file maps).
// - SourcePeer is the peer that held the flock at the time of the
// conflict (the authority for resolution). Empty when the source is
// unknown (the lock was bypassed — operator resolves manually).
// - Versions maps peer → content for every peer that holds a copy of
// the file. Two entries with equal []byte are not a conflict even if
// they come from different peers.
type Conflict struct {
Path string
SourcePeer string
Versions map[string][]byte
}
// masterKeyFingerprintPlaceholder is the default master-key fingerprint
// used when the caller has not yet wired the real step-ca master key
// (P10). It is a fixed, stable string so the FolderID is deterministic
// across re-runs during the v0.9 development window. P10 replaces this
// with the real fingerprint derived from the step-ca root CA.
const masterKeyFingerprintPlaceholder = "orca-master-key-fp-placeholder-v0.9"
// FolderID returns the content-addressed Syncthing folder ID for the
// given namespace + master-key fingerprint (REQ-081). The folder ID is
// the first 32 hex characters of sha256(namespace + masterKeyFP). The
// fingerprint is optional — when empty, the v0.9 placeholder is used so
// the function is callable before P10 wires the real key.
//
// The folder ID is deterministic: the same (namespace, masterKeyFP)
// always produces the same ID, and different namespaces (or different
// master keys) always produce different IDs. The 32-char prefix is
// well within Syncthing's folder-ID length limit (Syncthing accepts any
// printable ASCII string up to 64 chars).
func FolderID(namespace string, masterKeyFP string) string {
if strings.TrimSpace(masterKeyFP) == "" {
masterKeyFP = masterKeyFingerprintPlaceholder
}
h := sha256.Sum256([]byte(namespace + masterKeyFP))
return hex.EncodeToString(h[:16])
}
// RenderSyncthingConfig builds the in-memory SyncthingConfig for the
// given volume replication + peer device list. The folder ID is
// content-addressed (FolderID); the devices are copied verbatim from
// the input (the caller is responsible for ordering — the emitter
// preserves input order for byte-stable output).
//
// Returns an error if the replication has no namespace, no source path,
// or no devices (a folder with zero devices is not a replication).
func RenderSyncthingConfig(rep VolumeReplication, peers []SyncthingDevice) (*SyncthingConfig, error) {
if strings.TrimSpace(rep.Namespace) == "" {
return nil, errors.New("storage: replication namespace is empty")
}
if strings.TrimSpace(rep.SourcePath) == "" {
return nil, errors.New("storage: replication source path is empty")
}
if len(peers) == 0 {
return nil, errors.New("storage: replication has no peer devices")
}
mode := strings.TrimSpace(rep.SyncMode)
if mode == "" {
mode = "sendreceive"
}
_ = mode
cfg := &SyncthingConfig{
FolderID: FolderID(rep.Namespace, ""),
Path: rep.SourcePath,
Devices: append([]SyncthingDevice(nil), peers...),
}
return cfg, nil
}
// syncthingXMLFolder is the <folder> element in the rendered config.
type syncthingXMLFolder struct {
XMLName xml.Name `xml:"folder"`
ID string `xml:"id,attr"`
Path string `xml:"path,attr"`
Type string `xml:"type,attr"`
IgnorePerms bool `xml:"ignorePerms,attr"`
Devices []syncthingXMLDevice `xml:"device"`
FSync bool `xml:"fsync"`
}
// syncthingXMLDevice is the <device> element (both inside <folder> and
// at the top level; the top-level form carries the address).
type syncthingXMLDevice struct {
XMLName xml.Name `xml:"device"`
ID string `xml:"id,attr"`
Name string `xml:"name,attr"`
Compression string `xml:"compression,attr,omitempty"`
Address string `xml:"address,omitempty"`
}
// syncthingXMLOptions is the <options> element.
type syncthingXMLOptions struct {
XMLName xml.Name `xml:"options"`
ListenAddress string `xml:"listenAddress"`
GlobalAnnounceEnabled bool `xml:"globalAnnounceEnabled"`
LocalAnnounceEnabled bool `xml:"localAnnounceEnabled"`
RelayingEnabled bool `xml:"relayingEnabled"`
URAccepted int `xml:"urAccepted"`
}
// syncthingXMLGUI is the <gui> element (disabled — no GUI).
type syncthingXMLGUI struct {
XMLName xml.Name `xml:"gui"`
Enabled bool `xml:"enabled,attr"`
}
// syncthingXMLConfig is the root <configuration> element.
type syncthingXMLConfig struct {
XMLName xml.Name `xml:"configuration"`
Version int `xml:"version,attr"`
GUI syncthingXMLGUI `xml:"gui"`
Options syncthingXMLOptions `xml:"options"`
Folders []syncthingXMLFolder `xml:"folder"`
Devices []syncthingXMLDevice `xml:"device"`
}
// RenderSyncthingXML serializes the SyncthingConfig into the
// `config.xml` file content. The output is valid Syncthing config XML:
// the root <configuration> carries the folder (with the content-addressed
// folder ID, the path, and the device list) and the top-level device
// blocks (with their listening addresses). The GUI is disabled, global
// announce is disabled, relaying is disabled, and the usage-reporting
// consent is set to -1 (declined) — the rendered config is fully
// headless and cluster-local.
//
// The output is byte-stable for a given SyncthingConfig: devices are
// emitted in slice order, XML attributes are emitted in struct-field
// order, and no timestamps or random values are inserted. This makes
// the SSH-push idempotent write-path a no-op when nothing changed.
//
// Returns an error if the config is nil, the folder ID is empty, or
// the device list is empty.
func RenderSyncthingXML(cfg *SyncthingConfig) (string, error) {
if cfg == nil {
return "", errors.New("storage: syncthing config is nil")
}
if strings.TrimSpace(cfg.FolderID) == "" {
return "", errors.New("storage: syncthing folder ID is empty")
}
if len(cfg.Devices) == 0 {
return "", errors.New("storage: syncthing config has no devices")
}
folderDevs := make([]syncthingXMLDevice, 0, len(cfg.Devices))
topDevs := make([]syncthingXMLDevice, 0, len(cfg.Devices))
for _, d := range cfg.Devices {
folderDevs = append(folderDevs, syncthingXMLDevice{
ID: d.ID,
Name: d.Name,
})
topDevs = append(topDevs, syncthingXMLDevice{
ID: d.ID,
Name: d.Name,
Compression: "metadata",
Address: d.Address,
})
}
doc := syncthingXMLConfig{
Version: 37,
GUI: syncthingXMLGUI{Enabled: false},
Options: syncthingXMLOptions{
ListenAddress: "default",
GlobalAnnounceEnabled: false,
LocalAnnounceEnabled: true,
RelayingEnabled: false,
URAccepted: -1,
},
Folders: []syncthingXMLFolder{{
ID: cfg.FolderID,
Path: cfg.Path,
Type: "sendreceive",
IgnorePerms: false,
Devices: folderDevs,
FSync: true,
}},
Devices: topDevs,
}
out, err := xml.MarshalIndent(doc, "", " ")
if err != nil {
return "", fmt.Errorf("storage: marshal syncthing xml: %w", err)
}
return xml.Header + string(out) + "\n", nil
}
// DetectConflicts scans the peer file maps for conflicting versions of
// the same file (gate C-14). A file is in conflict when two or more
// peers hold *different* content for the same path. Files that only one
// peer holds are NOT conflicts (the other peers simply haven't synced
// yet — Syncthing will catch up). Files that all peers hold with equal
// content are NOT conflicts.
//
// The peerFiles map is peer → (path → content). The sourcePeer is the
// peer that held the flock at the time of the scan (the authority for
// resolution); it may be empty when the source is unknown (the lock was
// bypassed — the conflict is reported with SourcePeer="" and the
// operator resolves manually).
//
// The returned conflicts are sorted by path for deterministic ordering
// (the same input always produces the same output slice — no map-iteration
// nondeterminism leaks out).
func DetectConflicts(namespace string, peerFiles map[string]map[string][]byte) ([]Conflict, error) {
_ = namespace
if len(peerFiles) == 0 {
return nil, nil
}
// path -> peer -> content
byPath := make(map[string]map[string][]byte)
for peer, files := range peerFiles {
for path, content := range files {
if byPath[path] == nil {
byPath[path] = make(map[string][]byte)
}
byPath[path][peer] = append([]byte(nil), content...)
}
}
paths := make([]string, 0, len(byPath))
for p := range byPath {
paths = append(paths, p)
}
sort.Strings(paths)
var conflicts []Conflict
for _, path := range paths {
versions := byPath[path]
if len(versions) < 2 {
continue
}
if !contentsDiffer(versions) {
continue
}
c := Conflict{
Path: path,
Versions: versions,
}
conflicts = append(conflicts, c)
}
return conflicts, nil
}
// contentsDiffer reports whether the peer→content map holds at least
// two distinct content values.
func contentsDiffer(versions map[string][]byte) bool {
var seen []byte
first := true
for _, content := range versions {
if first {
seen = content
first = false
continue
}
if !bytesEqual(seen, content) {
return true
}
}
return false
}
// bytesEqual is a thin wrapper over bytes.Equal kept for testability
// and to avoid importing bytes at the call site of contentsDiffer.
func bytesEqual(a, b []byte) bool {
if len(a) != len(b) {
return false
}
for i := range a {
if a[i] != b[i] {
return false
}
}
return true
}
// ResolveConflict resolves a single conflict by picking the source
// peer's content (the peer that held the lock). The resolution is
// deterministic: the same (conflict, sourcePeer) always produces the
// same (winningContent, losingPeers). Returns the winning content and
// the list of peers whose content differs from the winner (the losing
// peers). The losingPeers list is sorted for deterministic ordering.
//
// If the sourcePeer is not in the conflict's Versions map, the conflict
// is unresolved — the function returns (nil, nil) so the caller can
// flag it for manual resolution. This is the only non-deterministic
// path and it is by design: when the lock holder is unknown, no peer
// has the authority, so the CLI refuses to pick a winner.
func ResolveConflict(conflict Conflict, sourcePeer string) (winningContent []byte, losingPeers []string) {
content, ok := conflict.Versions[sourcePeer]
if !ok {
return nil, nil
}
winningContent = append([]byte(nil), content...)
losing := make([]string, 0, len(conflict.Versions))
for peer, c := range conflict.Versions {
if peer == sourcePeer {
continue
}
if !bytesEqual(c, winningContent) {
losing = append(losing, peer)
}
}
sort.Strings(losing)
return winningContent, losing
}