feat(coordinator): artifact storage foundation (CTX-05, part 1)

Introduce durable, coordinator-owned artifacts — the model the master plan
requires instead of trusting worker-supplied result URIs.

- migration 0003: artifacts table + artifact_kind enum, plus nullable
  input/result_artifact_id columns on jobs and tasks.
- domain.Artifact with NewArtifact/SetContent; storage keys derive from a
  fresh UUID, never from a client filename (no path traversal).
- BlobStore and ArtifactRepository ports.
- blob.FSStore: filesystem blob storage that streams while hashing (SHA-256),
  fsyncs, and atomically renames into place — a failed upload leaves neither a
  committed artifact nor a staging file. Covered by unit tests.
- ArtifactRepo (squirrel) and COORDINATOR_STORAGE_DIR config.

HTTP upload/download handlers and the switch of result submission to
artifact_id come in the next parts.
This commit is contained in:
Efremenko Arhip
2026-07-23 14:06:19 +03:00
parent a5945f2d38
commit dbf578c500
9 changed files with 443 additions and 8 deletions
+69
View File
@@ -0,0 +1,69 @@
package domain
import (
"time"
"github.com/google/uuid"
)
type ArtifactKind string
const (
ArtifactInput ArtifactKind = "input"
ArtifactShard ArtifactKind = "shard"
ArtifactPartialResult ArtifactKind = "partial_result"
ArtifactFinalResult ArtifactKind = "final_result"
ArtifactLog ArtifactKind = "log"
)
// Artifact is a durable file the coordinator owns, described by its metadata.
// The bytes live in blob storage under StorageKey; this struct is what the
// database persists and what every other layer reasons about.
type Artifact struct {
ID uuid.UUID
JobID uuid.UUID
TaskID *uuid.UUID // nil for a job-level input
Kind ArtifactKind
Filename string
StorageKey string
ContentType string
SizeBytes int64
SHA256 string
CreatedAt time.Time
}
// NewArtifact begins an artifact record. Size and checksum are unknown until the
// bytes have been streamed to storage, so they are filled in later by SetContent.
//
// StorageKey is derived from a fresh UUID, never from the client-supplied
// filename — that is what stops a "../../etc/passwd" filename from escaping the
// storage directory.
func NewArtifact(jobID uuid.UUID, taskID *uuid.UUID, kind ArtifactKind,
filename, contentType string, now time.Time) (*Artifact, error) {
if filename == "" || kind == "" {
return nil, ErrInvalidInput
}
if contentType == "" {
contentType = "application/octet-stream"
}
id := uuid.New()
return &Artifact{
ID: id,
JobID: jobID,
TaskID: taskID,
Kind: kind,
Filename: filename,
StorageKey: id.String(),
ContentType: contentType,
CreatedAt: now,
}, nil
}
// SetContent records the size and checksum measured while streaming the bytes
// into storage. Both are computed by the coordinator, never trusted from the
// client — the whole point of owning the artifact.
func (a *Artifact) SetContent(sha256 string, size int64) {
a.SHA256 = sha256
a.SizeBytes = size
}
+9 -8
View File
@@ -8,12 +8,13 @@ import "errors"
//
// Always compare with errors.Is — outer layers may wrap these with %w.
var (
ErrJobNotFound = errors.New("job not found")
ErrTaskNotFound = errors.New("task not found")
ErrWorkerNotFound = errors.New("worker not found")
ErrLeaseConflict = errors.New("task leased to another worker")
ErrStaleAttempt = errors.New("attempt does not match lease")
ErrResultConflict = errors.New("different result already recorded")
ErrInvalidInput = errors.New("invalid input")
ErrTaskNotLeased = errors.New("task is not currently leased")
ErrJobNotFound = errors.New("job not found")
ErrTaskNotFound = errors.New("task not found")
ErrWorkerNotFound = errors.New("worker not found")
ErrArtifactNotFound = errors.New("artifact not found")
ErrLeaseConflict = errors.New("task leased to another worker")
ErrStaleAttempt = errors.New("attempt does not match lease")
ErrResultConflict = errors.New("different result already recorded")
ErrInvalidInput = errors.New("invalid input")
ErrTaskNotLeased = errors.New("task is not currently leased")
)
+3
View File
@@ -29,6 +29,8 @@ type Config struct {
LogLevel string
// Path to a rotated log file. Empty logs to stdout only.
LogFile string
// Directory where artifact bytes are stored.
StorageDir string
// Connection pool upper bound.
DBMaxConns int32
@@ -73,6 +75,7 @@ func LoadConfig() (Config, error) {
Token: getEnv("COORDINATOR_TOKEN", os.Getenv("WORKER_AUTH_TOKEN")),
LogLevel: getEnv("LOG_LEVEL", "info"),
LogFile: os.Getenv("LOG_FILE"),
StorageDir: getEnv("COORDINATOR_STORAGE_DIR", "./data"),
DBMaxConns: 10,
DBConnectTimeout: 30 * time.Second,
RequestTimeout: 15 * time.Second,
+116
View File
@@ -0,0 +1,116 @@
// Package blob stores artifact bytes on the local filesystem. It implements
// usecase.BlobStore; no other layer knows where or how the bytes are kept.
package blob
import (
"context"
"crypto/sha256"
"encoding/hex"
"fmt"
"io"
"os"
"path/filepath"
"strings"
"github.com/emil28092005/SciMesh/coordinator/internal/usecase"
)
// FSStore keeps each artifact as one file under dir, named by its storage key.
type FSStore struct {
dir string
staging string
}
var _ usecase.BlobStore = (*FSStore)(nil)
// NewFSStore prepares the storage and staging directories. Staging lives inside
// dir so a finished file can be renamed into place on the same filesystem —
// rename is only atomic within one filesystem.
func NewFSStore(dir string) (*FSStore, error) {
staging := filepath.Join(dir, ".staging")
if err := os.MkdirAll(staging, 0o755); err != nil {
return nil, fmt.Errorf("create blob dirs: %w", err)
}
return &FSStore{dir: dir, staging: staging}, nil
}
// Put streams r to a staging file while hashing it, then atomically renames it
// into place. A caller that dies mid-upload leaves at most a staging temp file,
// never a half-written artifact that looks complete.
func (s *FSStore) Put(ctx context.Context, key string, r io.Reader) (string, int64, error) {
if err := checkKey(key); err != nil {
return "", 0, err
}
tmp, err := os.CreateTemp(s.staging, key+"-*")
if err != nil {
return "", 0, fmt.Errorf("create staging file: %w", err)
}
tmpName := tmp.Name()
// On any failure past this point, do not leave the temp file behind.
defer func() {
if tmpName != "" {
_ = os.Remove(tmpName)
}
}()
h := sha256.New()
// Tee the stream: one copy to disk, one to the hasher, in a single pass so
// the bytes are never held in memory or read twice.
size, err := io.Copy(io.MultiWriter(tmp, h), &ctxReader{ctx: ctx, r: r})
if err != nil {
_ = tmp.Close()
return "", 0, fmt.Errorf("write artifact: %w", err)
}
// fsync before rename so a crash cannot leave a renamed-but-empty file.
if err := tmp.Sync(); err != nil {
_ = tmp.Close()
return "", 0, fmt.Errorf("sync artifact: %w", err)
}
if err := tmp.Close(); err != nil {
return "", 0, fmt.Errorf("close artifact: %w", err)
}
final := filepath.Join(s.dir, key)
if err := os.Rename(tmpName, final); err != nil {
return "", 0, fmt.Errorf("commit artifact: %w", err)
}
tmpName = "" // committed — the deferred cleanup must not delete it now
return hex.EncodeToString(h.Sum(nil)), size, nil
}
// Open returns the artifact bytes for streaming to a client. The caller closes.
func (s *FSStore) Open(ctx context.Context, key string) (io.ReadCloser, error) {
if err := checkKey(key); err != nil {
return nil, err
}
f, err := os.Open(filepath.Join(s.dir, key))
if err != nil {
return nil, err
}
return f, nil
}
// checkKey rejects anything that could escape the storage directory. Keys are
// coordinator-generated UUIDs, so this is defence in depth, not the only guard.
func checkKey(key string) error {
if key == "" || strings.ContainsAny(key, `/\`) || strings.Contains(key, "..") {
return fmt.Errorf("invalid storage key %q", key)
}
return nil
}
// ctxReader aborts a copy when the request context is cancelled, so a stalled
// or disconnected upload does not tie up a file handle indefinitely.
type ctxReader struct {
ctx context.Context
r io.Reader
}
func (c *ctxReader) Read(p []byte) (int, error) {
if err := c.ctx.Err(); err != nil {
return 0, err
}
return c.r.Read(p)
}
@@ -0,0 +1,115 @@
package blob
import (
"bytes"
"context"
"crypto/sha256"
"encoding/hex"
"io"
"os"
"path/filepath"
"strings"
"testing"
)
func newStore(t *testing.T) *FSStore {
t.Helper()
s, err := NewFSStore(t.TempDir())
if err != nil {
t.Fatalf("NewFSStore: %v", err)
}
return s
}
func TestPutComputesChecksumAndSize(t *testing.T) {
s := newStore(t)
data := bytes.Repeat([]byte("chembl-row\n"), 10000) // ~110 KB, streamed
sum, size, err := s.Put(context.Background(), "key-1", bytes.NewReader(data))
if err != nil {
t.Fatalf("Put: %v", err)
}
want := sha256.Sum256(data)
if sum != hex.EncodeToString(want[:]) {
t.Errorf("sha256 = %s, want %s", sum, hex.EncodeToString(want[:]))
}
if size != int64(len(data)) {
t.Errorf("size = %d, want %d", size, len(data))
}
}
func TestPutThenOpenRoundTrips(t *testing.T) {
s := newStore(t)
data := []byte("partial result csv\n1,2,3\n")
if _, _, err := s.Put(context.Background(), "key-2", bytes.NewReader(data)); err != nil {
t.Fatalf("Put: %v", err)
}
rc, err := s.Open(context.Background(), "key-2")
if err != nil {
t.Fatalf("Open: %v", err)
}
defer rc.Close()
got, _ := io.ReadAll(rc)
if !bytes.Equal(got, data) {
t.Errorf("round-trip mismatch: got %q", got)
}
}
func TestPutLeavesNoStagingFileBehind(t *testing.T) {
s := newStore(t)
if _, _, err := s.Put(context.Background(), "key-3", strings.NewReader("x")); err != nil {
t.Fatalf("Put: %v", err)
}
entries, _ := os.ReadDir(s.staging)
if len(entries) != 0 {
t.Errorf("staging dir not empty after a successful put: %v", entries)
}
}
func TestPutFailureLeavesNoArtifactOrStaging(t *testing.T) {
s := newStore(t)
// A reader that errors partway through simulates a dropped upload.
r := io.MultiReader(strings.NewReader("half"), &erroringReader{})
if _, _, err := s.Put(context.Background(), "key-4", r); err == nil {
t.Fatal("expected an error from a failing reader")
}
if _, err := os.Stat(filepath.Join(s.dir, "key-4")); !os.IsNotExist(err) {
t.Error("a failed put must not leave a committed artifact")
}
if entries, _ := os.ReadDir(s.staging); len(entries) != 0 {
t.Errorf("a failed put must not leave staging files: %v", entries)
}
}
func TestPutRejectsUnsafeKeys(t *testing.T) {
s := newStore(t)
for _, key := range []string{"", "../escape", "a/b", `a\b`, "with..dots"} {
if _, _, err := s.Put(context.Background(), key, strings.NewReader("x")); err == nil {
t.Errorf("key %q should have been rejected", key)
}
}
}
func TestPutHonoursContextCancellation(t *testing.T) {
s := newStore(t)
ctx, cancel := context.WithCancel(context.Background())
cancel() // already cancelled before the copy starts
if _, _, err := s.Put(ctx, "key-5", strings.NewReader("data")); err == nil {
t.Fatal("expected cancellation to abort the put")
}
if _, err := os.Stat(filepath.Join(s.dir, "key-5")); !os.IsNotExist(err) {
t.Error("a cancelled put must not leave an artifact")
}
}
type erroringReader struct{}
func (*erroringReader) Read([]byte) (int, error) { return 0, io.ErrUnexpectedEOF }
@@ -0,0 +1,72 @@
package postgres
import (
"context"
"errors"
"fmt"
sq "github.com/Masterminds/squirrel"
"github.com/google/uuid"
"github.com/jackc/pgx/v5"
"github.com/jackc/pgx/v5/pgxpool"
"github.com/emil28092005/SciMesh/coordinator/internal/domain"
"github.com/emil28092005/SciMesh/coordinator/internal/usecase"
)
// ArtifactRepo implements usecase.ArtifactRepository.
type ArtifactRepo struct {
pool *pgxpool.Pool
}
func NewArtifactRepo(pool *pgxpool.Pool) *ArtifactRepo {
return &ArtifactRepo{pool: pool}
}
var _ usecase.ArtifactRepository = (*ArtifactRepo)(nil)
var artifactColumns = []string{
"id", "job_id", "task_id", "kind", "filename", "storage_key",
"content_type", "size_bytes", "sha256", "created_at",
}
func (r *ArtifactRepo) Insert(ctx context.Context, a *domain.Artifact) error {
sql, args, err := psql.Insert("artifacts").
Columns(artifactColumns...).
Values(a.ID, a.JobID, a.TaskID, string(a.Kind), a.Filename, a.StorageKey,
a.ContentType, a.SizeBytes, a.SHA256, a.CreatedAt).
ToSql()
if err != nil {
return err
}
if _, err := conn(ctx, r.pool).Exec(ctx, sql, args...); err != nil {
return fmt.Errorf("insert artifact: %w", err)
}
return nil
}
func (r *ArtifactRepo) Get(ctx context.Context, id uuid.UUID) (*domain.Artifact, error) {
sql, args, err := psql.Select(artifactColumns...).
From("artifacts").
Where(sq.Eq{"id": id}).
ToSql()
if err != nil {
return nil, err
}
var (
a domain.Artifact
kind string
)
err = conn(ctx, r.pool).QueryRow(ctx, sql, args...).Scan(
&a.ID, &a.JobID, &a.TaskID, &kind, &a.Filename, &a.StorageKey,
&a.ContentType, &a.SizeBytes, &a.SHA256, &a.CreatedAt)
if errors.Is(err, pgx.ErrNoRows) {
return nil, domain.ErrArtifactNotFound
}
if err != nil {
return nil, fmt.Errorf("get artifact: %w", err)
}
a.Kind = domain.ArtifactKind(kind)
return &a, nil
}
+17
View File
@@ -9,6 +9,7 @@ package usecase
import (
"context"
"errors"
"io"
"time"
"github.com/google/uuid"
@@ -68,6 +69,22 @@ type WorkerRepository interface {
Get(ctx context.Context, id uuid.UUID) (*domain.Worker, error)
}
// ArtifactRepository persists artifact metadata. The bytes live in a BlobStore;
// this keeps only the record that points at them.
type ArtifactRepository interface {
Insert(ctx context.Context, a *domain.Artifact) error
Get(ctx context.Context, id uuid.UUID) (*domain.Artifact, error)
}
// BlobStore holds artifact bytes, addressed by an opaque storage key. It streams
// in both directions so a large shard never has to sit in memory, and reports
// the checksum and size it measured while writing — the coordinator's own
// numbers, not the client's claim.
type BlobStore interface {
Put(ctx context.Context, key string, r io.Reader) (sha256 string, size int64, err error)
Open(ctx context.Context, key string) (io.ReadCloser, error)
}
// TxManager runs a function inside one database transaction. The transaction
// travels in the context, so repositories pick it up without this port ever
// mentioning pgx.
@@ -0,0 +1,11 @@
BEGIN;
ALTER TABLE tasks DROP COLUMN IF EXISTS input_artifact_id;
ALTER TABLE tasks DROP COLUMN IF EXISTS result_artifact_id;
ALTER TABLE jobs DROP COLUMN IF EXISTS input_artifact_id;
ALTER TABLE jobs DROP COLUMN IF EXISTS result_artifact_id;
DROP TABLE IF EXISTS artifacts;
DROP TYPE IF EXISTS artifact_kind;
COMMIT;
@@ -0,0 +1,31 @@
BEGIN;
CREATE TYPE artifact_kind AS ENUM ('input','shard','partial_result','final_result','log');
-- A durable file the coordinator owns: input, shard, partial/final result, log.
-- The database is the source of truth; files are found through this metadata,
-- never by scanning directories.
CREATE TABLE artifacts (
id uuid PRIMARY KEY,
job_id uuid NOT NULL REFERENCES jobs(id) ON DELETE CASCADE,
task_id uuid REFERENCES tasks(id) ON DELETE CASCADE, -- null for job-level inputs
kind artifact_kind NOT NULL,
filename text NOT NULL,
storage_key text NOT NULL UNIQUE, -- coordinator-generated, never a client path
content_type text NOT NULL DEFAULT 'application/octet-stream',
size_bytes bigint NOT NULL CHECK (size_bytes >= 0),
sha256 text NOT NULL,
created_at timestamptz NOT NULL DEFAULT now()
);
CREATE INDEX ix_artifacts_job ON artifacts (job_id);
CREATE INDEX ix_artifacts_task ON artifacts (task_id);
-- Jobs and tasks reference their artifacts. Nullable during the transition from
-- URI-based inputs/results to artifact-based ones.
ALTER TABLE jobs ADD COLUMN input_artifact_id uuid REFERENCES artifacts(id);
ALTER TABLE jobs ADD COLUMN result_artifact_id uuid REFERENCES artifacts(id);
ALTER TABLE tasks ADD COLUMN input_artifact_id uuid REFERENCES artifacts(id);
ALTER TABLE tasks ADD COLUMN result_artifact_id uuid REFERENCES artifacts(id);
COMMIT;