| 1 | package dbreads |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "crypto/ecdh" |
| 6 | "encoding/json" |
| 7 | "errors" |
| 8 | "io" |
| 9 | "net" |
| 10 | "os" |
| 11 | "sync/atomic" |
| 12 | "time" |
| 13 | |
| 14 | "github.com/superfly/ltx" |
| 15 | |
| 16 | "github.com/carlosframework/platform/internal/activator" |
| 17 | "github.com/carlosframework/platform/internal/dbdoor" |
| 18 | ) |
| 19 | |
| 20 | // serve runs one read this agent owns, inside the owner's deadline, and |
| 21 | // writes its result (spec §5.5 step 5). The read slot is given back here; |
| 22 | // the reservation stays until the worker unit is gone (J5). |
| 23 | func (a *Agent) serve(parent context.Context, req Request, p Picked) { |
| 24 | defer a.reads.Done() |
| 25 | defer a.adm.finish(req.ID) |
| 26 | // The deadline is the one picked published (spec §5.7), measured on |
| 27 | // this agent's clock, not a fresh OwnerDeadline from now: an election |
| 28 | // whose answer came back late has already spent part of it, and the |
| 29 | // console stops waiting a minute after the published one (Astra round |
| 30 | // 1, finding 3). In production a.now is time.Now, so this is |
| 31 | // context.WithDeadline(parent, p.Deadline). |
| 32 | started := a.now() |
| 33 | left := p.Deadline.Sub(started) |
| 34 | ctx, cancel := context.WithTimeout(parent, left) |
| 35 | defer cancel() |
| 36 | var cancelled atomic.Bool |
| 37 | var res Result |
| 38 | if left <= 0 { |
| 39 | // Past the deadline before any work: start nothing. |
| 40 | a.logf("dbreads: %s: picked, but its deadline %s had passed when the election answered", req.ID, p.Deadline.Format(time.RFC3339Nano)) |
| 41 | res = Result{State: StateFailed, Code: CodeInterrupted} |
| 42 | if pub, err := ParseReplyKey(req.ReplyKey); err == nil { |
| 43 | res = a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, 0) |
| 44 | } |
| 45 | } else { |
| 46 | go a.watchCancel(ctx, req.ID, cancel, &cancelled) |
| 47 | res = a.read(ctx, req) |
| 48 | } |
| 49 | if cancelled.Load() { |
| 50 | res.State, res.Code = StateCancelled, "" |
| 51 | } |
| 52 | res.Box, res.Started, res.Ended = a.cfg.Box, started, a.now() |
| 53 | body, err := json.Marshal(res) |
| 54 | if err == nil { |
| 55 | wctx, wcancel := context.WithTimeout(context.WithoutCancel(parent), 30*time.Second) |
| 56 | err = putCreateOnly(wctx, a.cfg.Store, ResultKey(req.ID), body) |
| 57 | wcancel() |
| 58 | } |
| 59 | if err != nil { |
| 60 | a.logf("dbreads: %s: write the result: %v", req.ID, err) |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | // watchCancel reads cancel every CancelEvery and ends the read when it |
| 65 | // appears: the console's best-effort word that nobody is waiting. |
| 66 | func (a *Agent) watchCancel(ctx context.Context, id string, cancel context.CancelFunc, seen *atomic.Bool) { |
| 67 | t := time.NewTicker(a.cfg.CancelEvery) |
| 68 | defer t.Stop() |
| 69 | for { |
| 70 | select { |
| 71 | case <-ctx.Done(): |
| 72 | return |
| 73 | case <-t.C: |
| 74 | if _, _, err := a.cfg.Store.Get(ctx, CancelKey(id)); err == nil { |
| 75 | seen.Store(true) |
| 76 | cancel() |
| 77 | return |
| 78 | } |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | // ended seals the summary into a result. |
| 84 | func (a *Agent) ended(pub *ecdh.PublicKey, id, state, code string, s Summary, last int64) Result { |
| 85 | sealed, err := SealSummary(pub, id, s) |
| 86 | if err != nil { |
| 87 | a.logf("dbreads: %s: seal the summary: %v", id, err) |
| 88 | } |
| 89 | return Result{State: state, Code: code, LastSeq: last, Sealed: sealed} |
| 90 | } |
| 91 | |
| 92 | // read is today's "changed while it was being read" retry, box side: one |
| 93 | // fresh plan, with both checks again, then changed-twice. |
| 94 | func (a *Agent) read(ctx context.Context, req Request) Result { |
| 95 | pub, err := ParseReplyKey(req.ReplyKey) |
| 96 | if err != nil { |
| 97 | return Result{State: StateFailed, Code: CodeInterrupted} |
| 98 | } |
| 99 | at, _ := req.AtTime() |
| 100 | for attempt := 0; attempt < 2; attempt++ { |
| 101 | if attempt > 0 { |
| 102 | // The first attempt's worker may still be stopping: keep its |
| 103 | // reservation, and wait for its unit before the next one (J5). |
| 104 | a.adm.retry(req.ID, attempt) |
| 105 | if err := a.adm.settle(ctx, settlePoll); err != nil { |
| 106 | return a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, 0) |
| 107 | } |
| 108 | } |
| 109 | if res, retry := a.attempt(ctx, req, pub, at); !retry { |
| 110 | return res |
| 111 | } |
| 112 | } |
| 113 | // "Changed twice" is itself the replica's word: it waits for check B |
| 114 | // like every other replica-derived ending (Astra round 1, finding 2). |
| 115 | res, _ := a.afterCheckB(ctx, req, pub, CodeChangedTwice, Summary{}) |
| 116 | return res |
| 117 | } |
| 118 | |
| 119 | // afterCheckB is how every ending that carries anything derived from the |
| 120 | // replica leaves a read: a failed restore's error text, the restored size, |
| 121 | // the oldest point kept, "nothing replicated", "no room", "changed twice". |
| 122 | // Check B runs first (spec §8.2), and when it fails the details are |
| 123 | // discarded: the read ends refused with the identity code alone, so a |
| 124 | // requester never learns anything about another instance's replica read |
| 125 | // under a reused prefix (Astra round 1, finding 2). |
| 126 | func (a *Agent) afterCheckB(ctx context.Context, req Request, pub *ecdh.PublicKey, code string, s Summary) (Result, bool) { |
| 127 | _, bcode, err := a.check(ctx, req) |
| 128 | switch { |
| 129 | case err != nil: |
| 130 | a.logf("dbreads: %s: check B before a %s ending: %v", req.ID, code, err) |
| 131 | return a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, 0), false |
| 132 | case bcode != "": |
| 133 | return a.ended(pub, req.ID, StateRefused, bcode, Summary{}, 0), false |
| 134 | } |
| 135 | return a.ended(pub, req.ID, StateFailed, code, s, 0), false |
| 136 | } |
| 137 | |
| 138 | // settlePoll is how often a retry rereads the unit listing while it waits. |
| 139 | var settlePoll = 200 * time.Millisecond |
| 140 | |
| 141 | // attempt is the identity bracket (spec §8.2): check A, plan, stream every |
| 142 | // planned file into a worker that restores and checks it, check B after the |
| 143 | // last byte has left the bucket, and only then run. No output exists before |
| 144 | // check B passes. retry is true when a planned file vanished or changed. |
| 145 | func (a *Agent) attempt(ctx context.Context, req Request, pub *ecdh.PublicKey, at time.Time) (Result, bool) { |
| 146 | fail := func(code string, s Summary) (Result, bool) { |
| 147 | return a.ended(pub, req.ID, StateFailed, code, s, 0), false |
| 148 | } |
| 149 | // failB is fail for an ending derived from the replica: check B first. |
| 150 | failB := func(code string, s Summary) (Result, bool) { return a.afterCheckB(ctx, req, pub, code, s) } |
| 151 | refuse := func(code string) (Result, bool) { |
| 152 | return a.ended(pub, req.ID, StateRefused, code, Summary{}, 0), false |
| 153 | } |
| 154 | attemptStart := time.Now() |
| 155 | pctx, pcancel := context.WithTimeout(ctx, a.cfg.PlanBudget) |
| 156 | defer pcancel() |
| 157 | |
| 158 | rec, code, err := a.check(pctx, req) // check A |
| 159 | if err != nil { |
| 160 | a.logf("dbreads: %s: check A: %v", req.ID, err) |
| 161 | return fail(CodeInterrupted, Summary{}) |
| 162 | } |
| 163 | if code != "" { |
| 164 | return refuse(code) |
| 165 | } |
| 166 | rd, err := a.cfg.OpenReplica(pctx, activator.HibernatePrefix(a.cfg.HibernatePrefix, rec.Host)) |
| 167 | if err != nil { |
| 168 | a.logf("dbreads: %s: open the replica: %v", req.ID, err) |
| 169 | return fail(CodeInterrupted, Summary{}) |
| 170 | } |
| 171 | plan, err := dbdoor.PlanFrom(pctx, rd, at) |
| 172 | var before *dbdoor.BeforeHistoryError |
| 173 | switch { |
| 174 | case errors.Is(err, dbdoor.ErrNothingReplicated): |
| 175 | return failB(CodeNothingReplicated, Summary{}) |
| 176 | case errors.As(err, &before): |
| 177 | return failB(CodeHistoryGone, Summary{Oldest: before.Oldest}) |
| 178 | case errors.Is(err, dbdoor.ErrHistoryGone): |
| 179 | return failB(CodeHistoryGone, Summary{}) |
| 180 | case errors.Is(err, os.ErrNotExist): |
| 181 | return Result{}, true |
| 182 | case err != nil: |
| 183 | a.logf("dbreads: %s: plan: %v", req.ID, err) |
| 184 | return fail(CodeInterrupted, Summary{}) |
| 185 | } |
| 186 | // A pull's cap is on the files it ships, a query's on the copy it |
| 187 | // restores (spec §5.7). Both endings are the replica's word, so both wait |
| 188 | // for check B (Astra round 1, finding 2). |
| 189 | pull := req.Op == OpPull |
| 190 | switch { |
| 191 | case pull && plan.LTXBytes > a.cfg.MaxPull: |
| 192 | return failB(CodeTooLarge, Summary{Restored: plan.LTXBytes, Cap: a.cfg.MaxPull}) |
| 193 | case !pull && plan.RestoredBytes > a.cfg.MaxRestored: |
| 194 | return failB(CodeTooLarge, Summary{Restored: plan.RestoredBytes, Cap: a.cfg.MaxRestored}) |
| 195 | } |
| 196 | need := plan.LTXBytes + plan.RestoredBytes |
| 197 | if pull { |
| 198 | need = plan.LTXBytes // the worker holds the files and restores nothing |
| 199 | } |
| 200 | if err := a.adm.grow(req.ID, need); err != nil { |
| 201 | return failB(CodeNoRoom, Summary{}) |
| 202 | } |
| 203 | conn, err := a.dial(pctx) |
| 204 | if err != nil { |
| 205 | a.logf("dbreads: %s: the worker did not start: %v", req.ID, err) |
| 206 | a.askProbe() |
| 207 | return fail(CodeInterrupted, Summary{}) |
| 208 | } |
| 209 | defer conn.Close() |
| 210 | defer context.AfterFunc(ctx, func() { conn.Close() })() |
| 211 | stopPlanBound := context.AfterFunc(pctx, func() { conn.Close() }) |
| 212 | defer stopPlanBound() |
| 213 | |
| 214 | if err := writeJSONFrame(conn, frameHeader, planHeader{V: ProtocolVersion, Op: req.Op, Files: plan.Planned, |
| 215 | Target: plan.TXID, MaxRestored: a.cfg.MaxRestored}); err != nil { |
| 216 | closeThenProbe(a, conn) |
| 217 | return fail(CodeInterrupted, Summary{}) |
| 218 | } |
| 219 | for _, f := range plan.Planned { |
| 220 | vanished, err := sendFile(pctx, conn, rd, f) |
| 221 | if vanished { |
| 222 | return Result{}, true |
| 223 | } |
| 224 | if err != nil { |
| 225 | // The worker went before ready: its unit may have failed to |
| 226 | // start after the socket accepted (spec §7.1; plan review |
| 227 | // finding 15). Probe once this connection is closed. |
| 228 | closeThenProbe(a, conn) |
| 229 | return fail(CodeInterrupted, Summary{}) |
| 230 | } |
| 231 | } |
| 232 | typ, p, err := readFrame(conn) |
| 233 | switch { |
| 234 | case err != nil: |
| 235 | closeThenProbe(a, conn) |
| 236 | return fail(CodeInterrupted, Summary{}) |
| 237 | case typ == frameEnd: |
| 238 | var e endMsg |
| 239 | json.Unmarshal(p, &e) |
| 240 | if e.Code != CodeTooLarge { |
| 241 | e.Code = CodeCheckFailed |
| 242 | } |
| 243 | return failB(e.Code, Summary{Error: e.Error, Restored: plan.RestoredBytes, Cap: a.cfg.MaxRestored}) |
| 244 | } |
| 245 | var ready readyMsg |
| 246 | if typ != frameReady || json.Unmarshal(p, &ready) != nil || ready.TXID != plan.TXID { |
| 247 | return failB(CodeCheckFailed, Summary{}) |
| 248 | } |
| 249 | if _, code, err := a.check(pctx, req); err != nil || code != "" { // check B |
| 250 | writeFrame(conn, frameStop, nil) |
| 251 | if err != nil { |
| 252 | return fail(CodeInterrupted, Summary{}) |
| 253 | } |
| 254 | return refuse(code) |
| 255 | } |
| 256 | stopPlanBound() // the run is bounded by the worker and the owner's deadline |
| 257 | stream, size, max := StreamRows, ChunkBytes, a.cfg.MaxOutput |
| 258 | if pull { |
| 259 | stream, size, max = StreamLTX, PullChunkBytes, plan.LTXBytes |
| 260 | } |
| 261 | // Phase timings, for operators and for plan 3's throughput measurement |
| 262 | // (plan review finding 26): a query's first phase plans and restores, a |
| 263 | // pull's plans and has the worker verify and hold the files; then the |
| 264 | // output's sealing and confirmed creates. |
| 265 | held := "restored to" |
| 266 | if pull { |
| 267 | held = "held through" |
| 268 | } |
| 269 | a.logf("dbreads: %s: planned %d files (%d bytes) and %s %d in %s", req.ID, plan.Files, plan.LTXBytes, held, plan.TXID, time.Since(attemptStart).Round(time.Millisecond)) |
| 270 | runStart := time.Now() |
| 271 | if err := writeJSONFrame(conn, frameRun, runMsg{SQL: req.SQL}); err != nil { |
| 272 | return fail(CodeInterrupted, Summary{}) |
| 273 | } |
| 274 | cw := newChunkWriter(ctx, a.cfg.Store, req.ID, pub, stream, a.cfg.ChunkEvery, max, size) |
| 275 | endCode, sqlErr, rerr := relay(conn, cw) |
| 276 | last, n, sum, werr := cw.close() |
| 277 | a.logf("dbreads: %s: output %d bytes in %d confirmed chunks in %s", req.ID, n, last, time.Since(runStart).Round(time.Millisecond)) |
| 278 | s := Summary{TXID: plan.TXID, At: plan.At, Bytes: n, SHA256: sum, Error: sqlErr} |
| 279 | if pull { |
| 280 | s.Files = plan.Planned |
| 281 | // A worker that sends more or less than it held is a check |
| 282 | // failure, never an output cap. |
| 283 | if werr == nil && rerr == nil && endCode == "" && (cw.capped() || n != plan.LTXBytes) { |
| 284 | return a.ended(pub, req.ID, StateFailed, CodeCheckFailed, Summary{}, last), false |
| 285 | } |
| 286 | } |
| 287 | switch { |
| 288 | case werr != nil: |
| 289 | a.logf("dbreads: %s: write output: %v", req.ID, werr) |
| 290 | return a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, last), false |
| 291 | case cw.capped(): |
| 292 | return a.ended(pub, req.ID, StateDone, CodeOutputCap, s, last), false |
| 293 | case rerr != nil: |
| 294 | return a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, last), false |
| 295 | case endCode == "": |
| 296 | return a.ended(pub, req.ID, StateDone, "", s, last), false |
| 297 | case endCode == CodeSQLError, endCode == CodeTimeout: |
| 298 | return a.ended(pub, req.ID, StateFailed, endCode, s, last), false |
| 299 | } |
| 300 | return a.ended(pub, req.ID, StateFailed, CodeInterrupted, Summary{}, last), false |
| 301 | } |
| 302 | |
| 303 | // sendFile streams one planned file, whole, from the bucket into the |
| 304 | // worker. A file that is gone, or not the size the plan listed, vanished |
| 305 | // under the plan: the caller re-plans. |
| 306 | func sendFile(ctx context.Context, w io.Writer, rd dbdoor.ReplicaReader, f dbdoor.LTXFile) (bool, error) { |
| 307 | rc, err := rd.OpenLTXFile(ctx, f.Level, ltx.TXID(f.MinTXID), ltx.TXID(f.MaxTXID), 0, 0) |
| 308 | if errors.Is(err, os.ErrNotExist) { |
| 309 | return true, nil |
| 310 | } |
| 311 | if err != nil { |
| 312 | return false, err |
| 313 | } |
| 314 | defer rc.Close() |
| 315 | buf := make([]byte, fileFrameBytes) |
| 316 | var sent int64 |
| 317 | for { |
| 318 | n, rerr := rc.Read(buf) |
| 319 | if n > 0 { |
| 320 | if sent+int64(n) > f.Size { |
| 321 | return true, nil |
| 322 | } |
| 323 | if err := writeFrame(w, frameFile, buf[:n]); err != nil { |
| 324 | return false, err |
| 325 | } |
| 326 | sent += int64(n) |
| 327 | } |
| 328 | if rerr == io.EOF { |
| 329 | break |
| 330 | } |
| 331 | if errors.Is(rerr, os.ErrNotExist) { |
| 332 | return true, nil |
| 333 | } |
| 334 | if rerr != nil { |
| 335 | return false, rerr |
| 336 | } |
| 337 | } |
| 338 | return sent != f.Size, nil |
| 339 | } |
| 340 | |
| 341 | // relay moves the worker's output into the chunk writer until the worker's |
| 342 | // end frame, or until the cap: then the agent stops reading and closes the |
| 343 | // connection, which the worker sees as EOF. |
| 344 | func relay(conn net.Conn, cw *chunkWriter) (string, string, error) { |
| 345 | for { |
| 346 | typ, p, err := readFrame(conn) |
| 347 | if err != nil { |
| 348 | return "", "", err |
| 349 | } |
| 350 | switch typ { |
| 351 | case frameOutput: |
| 352 | if _, err := cw.Write(p); errors.Is(err, errCapped) { |
| 353 | return "", "", nil |
| 354 | } else if err != nil { |
| 355 | return "", "", err |
| 356 | } |
| 357 | case frameEnd: |
| 358 | var e endMsg |
| 359 | if json.Unmarshal(p, &e) != nil { |
| 360 | return "", "", errProtocol |
| 361 | } |
| 362 | return e.Code, e.Error, nil |
| 363 | default: |
| 364 | return "", "", errProtocol |
| 365 | } |
| 366 | } |
| 367 | } |
| 368 | |
amadan