| 1 | package main |
| 2 | |
| 3 | // 🤖 s3BlobStore — blobStore implementation #2 (#550 Phase 1): attachment |
| 4 | // bytes in an S3-compatible object store (Cloudflare R2 is the intended |
| 5 | // target — zero egress; see docs/tier-economics.md). Enabled per instance by |
| 6 | // LCHAT_BLOB_S3_BUCKET in its env; unset keeps today's on-disk layout, so |
| 7 | // self-hosters and the fleet flip independently, instance by instance. |
| 8 | // |
| 9 | // The store carries the same E2EE posture as disk: bytes are client-side |
| 10 | // ciphertext, keys/filenames/types never reach here, and object keys are the |
| 11 | // server-minted blob ids — the object store learns exactly what the local |
| 12 | // disk knew, nothing more. |
| 13 | // |
| 14 | // Shape decisions, in the order they matter: |
| 15 | // |
| 16 | // - WRAPS the disk store as a fallback, never replaces it: blobs uploaded |
| 17 | // before the flip live on disk, and open/remove consult S3 first, then |
| 18 | // disk. New writes always land in S3, so the disk set only shrinks — |
| 19 | // the migration sweep (#550 Phase 2) drains it and the fallback then |
| 20 | // never fires. Marker-free by construction. |
| 21 | // - Staged writes SPOOL TO DISK (the same temp-in-blobs-dir shape as the |
| 22 | // disk writer) and upload on commit. S3 has no rename, but a completed |
| 23 | // PUT is atomic — an object is fetchable only in full — so "invisible |
| 24 | // until commit" holds with no staging keys to garbage-collect. The spool |
| 25 | // also gives the PUT a known ContentLength (R2 requires one). |
| 26 | // - Reads are CHUNKED RANGED GETs behind an io.ReadSeekCloser: |
| 27 | // http.ServeContent seeks (size probe, Range requests — video scrubs) |
| 28 | // and then reads in 32KB slices; fetching per slice would be an S3 |
| 29 | // request per 32KB. The reader fetches aligned 512KB windows instead — |
| 30 | // ~2 requests/MB served, and a scrub costs one window, not a full |
| 31 | // download. (#836's lesson: request count is the S3 bill, watch it.) |
| 32 | // |
| 33 | // Config (instance env): |
| 34 | // |
| 35 | // LCHAT_BLOB_S3_BUCKET the switch — set = S3 store, unset = disk |
| 36 | // LCHAT_BLOB_S3_PREFIX key prefix; default = the DB's basename (the |
| 37 | // host), so a shared fleet bucket needs no |
| 38 | // per-instance wiring |
| 39 | // LCHAT_BLOB_S3_REGION optional ("auto" for R2) |
| 40 | // LCHAT_BLOB_S3_ENDPOINT optional custom endpoint (R2's |
| 41 | // https://<account>.r2.cloudflarestorage.com); |
| 42 | // implies path-style addressing |
| 43 | // |
| 44 | // Credentials ride the standard AWS chain (AWS_ACCESS_KEY_ID / |
| 45 | // AWS_SECRET_ACCESS_KEY in the instance env file — R2 issues S3-compatible |
| 46 | // pairs). |
| 47 | |
| 48 | import ( |
| 49 | "context" |
| 50 | "errors" |
| 51 | "fmt" |
| 52 | "io" |
| 53 | "io/fs" |
| 54 | "os" |
| 55 | "path" |
| 56 | "strings" |
| 57 | "time" |
| 58 | |
| 59 | "github.com/aws/aws-sdk-go-v2/aws" |
| 60 | awsconfig "github.com/aws/aws-sdk-go-v2/config" |
| 61 | "github.com/aws/aws-sdk-go-v2/credentials" |
| 62 | awss3 "github.com/aws/aws-sdk-go-v2/service/s3" |
| 63 | "github.com/aws/smithy-go" |
| 64 | ) |
| 65 | |
| 66 | // s3ReadChunk is the ranged-GET window. 512KB ≈ 2 requests per MB served; |
| 67 | // small enough that a video scrub's first paint isn't waiting on a huge |
| 68 | // window, large enough that request count stays a rounding error on the bill. |
| 69 | const s3ReadChunk = 512 << 10 |
| 70 | |
| 71 | // s3BlobTimeout bounds each individual S3 call (not a whole serve — a member |
| 72 | // streaming a long video holds the reader open far longer than any one GET). |
| 73 | const s3BlobTimeout = 60 * time.Second |
| 74 | |
| 75 | type s3BlobStore struct { |
| 76 | client *awss3.Client |
| 77 | bucket string |
| 78 | prefix string |
| 79 | // fallback serves blobs the remote lacks (pre-flip disk blobs for the |
| 80 | // primary store). Nil for a mirror store — a miss there is just a miss. |
| 81 | fallback blobStore |
| 82 | // spoolDir stages in-flight uploads — the disk store's dir, so the |
| 83 | // temp-name convention ("." never appears in a valid blob id) keeps |
| 84 | // staged bytes unfetchable exactly as before. |
| 85 | spoolDir string |
| 86 | } |
| 87 | |
| 88 | // s3BlobConfig reads the env switch. Nil config = feature off (disk store). |
| 89 | type s3BlobConfig struct { |
| 90 | Bucket string |
| 91 | Prefix string |
| 92 | Region string |
| 93 | Endpoint string |
| 94 | // Static credentials for THIS store — primary and mirror are different |
| 95 | // providers with different key pairs, and the AWS default chain can only |
| 96 | // hold one. Empty = the default chain (the primary's usual shape). |
| 97 | AccessKey string |
| 98 | SecretKey string |
| 99 | } |
| 100 | |
| 101 | // s3BlobEnv reads one store's env quartet (+ key pair) under a prefix: |
| 102 | // LCHAT_BLOB_S3_* for the primary, LCHAT_BLOB_S3_MIRROR_* for the mirror. |
| 103 | func s3BlobEnv(envPrefix string) *s3BlobConfig { |
| 104 | bucket := strings.TrimSpace(os.Getenv(envPrefix + "BUCKET")) |
| 105 | if bucket == "" { |
| 106 | return nil |
| 107 | } |
| 108 | return &s3BlobConfig{ |
| 109 | Bucket: bucket, |
| 110 | Prefix: strings.Trim(strings.TrimSpace(os.Getenv(envPrefix+"PREFIX")), "/"), |
| 111 | Region: strings.TrimSpace(os.Getenv(envPrefix + "REGION")), |
| 112 | Endpoint: strings.TrimSpace(os.Getenv(envPrefix + "ENDPOINT")), |
| 113 | AccessKey: strings.TrimSpace(os.Getenv(envPrefix + "ACCESS_KEY")), |
| 114 | SecretKey: strings.TrimSpace(os.Getenv(envPrefix + "SECRET_KEY")), |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | func s3BlobConfigFromEnv() *s3BlobConfig { return s3BlobEnv("LCHAT_BLOB_S3_") } |
| 119 | func s3BlobMirrorConfigFromEnv() *s3BlobConfig { return s3BlobEnv("LCHAT_BLOB_S3_MIRROR_") } |
| 120 | |
| 121 | // newS3BlobStore builds one remote store. fallback may be nil (a mirror); |
| 122 | // spoolDir stages writes; defaultPrefix is the DB basename — used when the |
| 123 | // env names none. |
| 124 | func newS3BlobStore(cfg *s3BlobConfig, spoolDir string, fallback blobStore, defaultPrefix string) (*s3BlobStore, error) { |
| 125 | var opts []func(*awsconfig.LoadOptions) error |
| 126 | if cfg.Region != "" { |
| 127 | opts = append(opts, awsconfig.WithRegion(cfg.Region)) |
| 128 | } |
| 129 | if cfg.AccessKey != "" { |
| 130 | opts = append(opts, awsconfig.WithCredentialsProvider( |
| 131 | credentials.NewStaticCredentialsProvider(cfg.AccessKey, cfg.SecretKey, ""))) |
| 132 | } |
| 133 | awsCfg, err := awsconfig.LoadDefaultConfig(context.Background(), opts...) |
| 134 | if err != nil { |
| 135 | return nil, fmt.Errorf("blob s3: load aws config: %w", err) |
| 136 | } |
| 137 | client := awss3.NewFromConfig(awsCfg, func(o *awss3.Options) { |
| 138 | if cfg.Endpoint != "" { |
| 139 | o.BaseEndpoint = &cfg.Endpoint |
| 140 | // Custom endpoints (R2, test fakes) address by path, not vhost. |
| 141 | o.UsePathStyle = true |
| 142 | } |
| 143 | // The SDK's default-on flexible checksums wrap PutObject bodies in |
| 144 | // aws-chunked framing, which R2 rejects; when-required sends plain |
| 145 | // bodies (ContentLength is always known — the spool measured it). |
| 146 | o.RequestChecksumCalculation = aws.RequestChecksumCalculationWhenRequired |
| 147 | o.ResponseChecksumValidation = aws.ResponseChecksumValidationWhenRequired |
| 148 | }) |
| 149 | prefix := cfg.Prefix |
| 150 | if prefix == "" { |
| 151 | prefix = defaultPrefix |
| 152 | } |
| 153 | return &s3BlobStore{client: client, bucket: cfg.Bucket, prefix: prefix, |
| 154 | fallback: fallback, spoolDir: spoolDir}, nil |
| 155 | } |
| 156 | |
| 157 | func (s *s3BlobStore) key(id string) string { return path.Join(s.prefix, id) } |
| 158 | |
| 159 | // s3NotFound folds the three shapes a missing object answers with (NoSuchKey |
| 160 | // on GET, NotFound on HEAD, and bare 404s from S3-compatibles) into one test. |
| 161 | func s3NotFound(err error) bool { |
| 162 | var apiErr smithy.APIError |
| 163 | if errors.As(err, &apiErr) { |
| 164 | switch apiErr.ErrorCode() { |
| 165 | case "NoSuchKey", "NotFound", "404": |
| 166 | return true |
| 167 | } |
| 168 | } |
| 169 | return false |
| 170 | } |
| 171 | |
| 172 | func (s *s3BlobStore) create() (blobWriter, error) { |
| 173 | f, err := os.CreateTemp(s.spoolDir, "up.*.part") |
| 174 | if err != nil { |
| 175 | return nil, err |
| 176 | } |
| 177 | return &s3BlobWriter{f: f, store: s}, nil |
| 178 | } |
| 179 | |
| 180 | func (s *s3BlobStore) open(id string) (io.ReadSeekCloser, int64, time.Time, error) { |
| 181 | ctx, cancel := context.WithTimeout(context.Background(), s3BlobTimeout) |
| 182 | defer cancel() |
| 183 | key := s.key(id) |
| 184 | head, err := s.client.HeadObject(ctx, &awss3.HeadObjectInput{Bucket: &s.bucket, Key: &key}) |
| 185 | if err != nil { |
| 186 | if s3NotFound(err) { |
| 187 | if s.fallback != nil { |
| 188 | // Pre-flip blob (or genuinely absent — the fallback answers |
| 189 | // with fs.ErrNotExist either way, which is the contract). |
| 190 | return s.fallback.open(id) |
| 191 | } |
| 192 | return nil, 0, time.Time{}, fs.ErrNotExist |
| 193 | } |
| 194 | return nil, 0, time.Time{}, err |
| 195 | } |
| 196 | size := int64(0) |
| 197 | if head.ContentLength != nil { |
| 198 | size = *head.ContentLength |
| 199 | } |
| 200 | mod := time.Time{} |
| 201 | if head.LastModified != nil { |
| 202 | mod = *head.LastModified |
| 203 | } |
| 204 | return &s3Reader{store: s, key: key, size: size}, size, mod, nil |
| 205 | } |
| 206 | |
| 207 | func (s *s3BlobStore) remove(id string) error { |
| 208 | ctx, cancel := context.WithTimeout(context.Background(), s3BlobTimeout) |
| 209 | defer cancel() |
| 210 | key := s.key(id) |
| 211 | if _, err := s.client.DeleteObject(ctx, &awss3.DeleteObjectInput{Bucket: &s.bucket, Key: &key}); err != nil && !s3NotFound(err) { |
| 212 | return err |
| 213 | } |
| 214 | // A pre-flip blob's bytes live on disk — GC must reap those too. The |
| 215 | // fallback treats a missing file as success, so this is free for S3-era ids. |
| 216 | if s.fallback != nil { |
| 217 | return s.fallback.remove(id) |
| 218 | } |
| 219 | return nil |
| 220 | } |
| 221 | |
| 222 | type s3BlobWriter struct { |
| 223 | f *os.File |
| 224 | store *s3BlobStore |
| 225 | } |
| 226 | |
| 227 | func (w *s3BlobWriter) Write(p []byte) (int, error) { return w.f.Write(p) } |
| 228 | |
| 229 | func (w *s3BlobWriter) commit(id string) error { |
| 230 | defer os.Remove(w.f.Name()) |
| 231 | size, err := w.f.Seek(0, io.SeekEnd) |
| 232 | if err != nil { |
| 233 | w.f.Close() |
| 234 | return err |
| 235 | } |
| 236 | if _, err := w.f.Seek(0, io.SeekStart); err != nil { |
| 237 | w.f.Close() |
| 238 | return err |
| 239 | } |
| 240 | ctx, cancel := context.WithTimeout(context.Background(), s3BlobTimeout) |
| 241 | defer cancel() |
| 242 | key := w.store.key(id) |
| 243 | _, err = w.store.client.PutObject(ctx, &awss3.PutObjectInput{ |
| 244 | Bucket: &w.store.bucket, |
| 245 | Key: &key, |
| 246 | Body: w.f, |
| 247 | ContentLength: &size, |
| 248 | }) |
| 249 | w.f.Close() |
| 250 | return err |
| 251 | } |
| 252 | |
| 253 | func (w *s3BlobWriter) abort() { |
| 254 | w.f.Close() |
| 255 | os.Remove(w.f.Name()) |
| 256 | } |
| 257 | |
| 258 | // s3Reader is a lazy io.ReadSeekCloser over one object: seeks are position |
| 259 | // math (http.ServeContent's size probe costs nothing), reads fetch aligned |
| 260 | // s3ReadChunk windows on demand. |
| 261 | type s3Reader struct { |
| 262 | store *s3BlobStore |
| 263 | key string |
| 264 | size int64 |
| 265 | pos int64 |
| 266 | // The current window: buf holds object bytes [bufStart, bufStart+len(buf)). |
| 267 | buf []byte |
| 268 | bufStart int64 |
| 269 | } |
| 270 | |
| 271 | func (r *s3Reader) Read(p []byte) (int, error) { |
| 272 | if r.pos >= r.size { |
| 273 | return 0, io.EOF |
| 274 | } |
| 275 | // Miss (or first read): fetch the aligned window containing pos. |
| 276 | if r.buf == nil || r.pos < r.bufStart || r.pos >= r.bufStart+int64(len(r.buf)) { |
| 277 | start := (r.pos / s3ReadChunk) * s3ReadChunk |
| 278 | end := start + s3ReadChunk - 1 // Range is inclusive |
| 279 | if end >= r.size { |
| 280 | end = r.size - 1 |
| 281 | } |
| 282 | ctx, cancel := context.WithTimeout(context.Background(), s3BlobTimeout) |
| 283 | rng := fmt.Sprintf("bytes=%d-%d", start, end) |
| 284 | out, err := r.store.client.GetObject(ctx, &awss3.GetObjectInput{ |
| 285 | Bucket: &r.store.bucket, Key: &r.key, Range: &rng, |
| 286 | }) |
| 287 | if err != nil { |
| 288 | cancel() |
| 289 | if s3NotFound(err) { |
| 290 | // Deleted mid-serve (GC race) — surface as the fs error the |
| 291 | // handler already maps to 404. |
| 292 | return 0, fs.ErrNotExist |
| 293 | } |
| 294 | return 0, err |
| 295 | } |
| 296 | buf, err := io.ReadAll(out.Body) |
| 297 | out.Body.Close() |
| 298 | cancel() |
| 299 | if err != nil { |
| 300 | return 0, err |
| 301 | } |
| 302 | r.buf, r.bufStart = buf, start |
| 303 | } |
| 304 | n := copy(p, r.buf[r.pos-r.bufStart:]) |
| 305 | r.pos += int64(n) |
| 306 | return n, nil |
| 307 | } |
| 308 | |
| 309 | func (r *s3Reader) Seek(offset int64, whence int) (int64, error) { |
| 310 | var next int64 |
| 311 | switch whence { |
| 312 | case io.SeekStart: |
| 313 | next = offset |
| 314 | case io.SeekCurrent: |
| 315 | next = r.pos + offset |
| 316 | case io.SeekEnd: |
| 317 | next = r.size + offset |
| 318 | default: |
| 319 | return 0, fmt.Errorf("blob s3: bad whence %d", whence) |
| 320 | } |
| 321 | if next < 0 { |
| 322 | return 0, fmt.Errorf("blob s3: negative seek") |
| 323 | } |
| 324 | r.pos = next |
| 325 | return next, nil |
| 326 | } |
| 327 | |
| 328 | func (r *s3Reader) Close() error { |
| 329 | r.buf = nil |
| 330 | return nil |
| 331 | } |
| 332 | |
amadan