| 1 |
<?php |
| 2 |
|
| 3 |
if (!defined('ABSPATH')) { |
| 4 |
exit; |
| 5 |
} |
| 6 |
|
| 7 |
/** |
| 8 |
* Staged getRedirectsForView pipeline. |
| 9 |
* |
| 10 |
* Replaces the legacy single-shot SQL that JOINed wp_posts/wp_terms/wp_options |
| 11 |
* onto every active redirect and forced an ORDER BY published_status filesort |
| 12 |
* across the full result before LIMIT applied. That shape times out at 45s+ |
| 13 |
* on cold-cache shared hosts (Bruno/Showmetech, multiple 4.1.13 reports). |
| 14 |
* |
| 15 |
* The pipeline writes a precomputed view of every redirect into a shared |
| 16 |
* persistent table (`{wp_abj404_view_done}`). Reads serve directly from |
| 17 |
* that table with WHERE/ORDER/LIMIT applied. Per-tab status filtering and |
| 18 |
* filterText LIKE both apply at read time, so one shared `_done` serves |
| 19 |
* every admin's tab. |
| 20 |
* |
| 21 |
* Concurrency model: one builder at a time per site, gated by a session |
| 22 |
* lock (`GET_LOCK`). Atomic `RENAME TABLE` swap publishes a freshly built |
| 23 |
* buffer to readers. Stale-while-revalidate on every request: if the |
| 24 |
* served snapshot is older than VIEW_DONE_FRESHNESS_TTL_SECONDS, kick off |
| 25 |
* a rebuild for the next request and serve the stale data now. |
| 26 |
*/ |
| 27 |
trait ABJ_404_Solution_DataAccess_ViewQueriesStagedTrait { |
| 28 |
|
| 29 |
// Tunable constants live on ABJ_404_Solution_ViewBuildConfig (see |
| 30 |
// includes/ViewBuildConfig.php) instead of as `const` declarations on |
| 31 |
// this trait, because PHP traits cannot have constants until 8.2 and |
| 32 |
// the plugin declares Requires PHP: 7.4. References below are written |
| 33 |
// as the FQN class constant rather than `self::` so they resolve the |
| 34 |
// same way regardless of which class consumes the trait. |
| 35 |
|
| 36 |
/** @var bool Process-local guard so a single request never rebuilds twice. */ |
| 37 |
private static $viewBuildAlreadyRanThisRequest = false; |
| 38 |
|
| 39 |
// Stage-runner / shutdown-diagnostics state ($viewBuildShutdownLoggerRegistered, |
| 40 |
// $viewBuildStageOpenForShutdown, $viewBuildShutdownStageNumber, |
| 41 |
// $viewBuildShutdownStageKey, $lastBatchProgressDetail) is declared on the |
| 42 |
// sibling ABJ_404_Solution_DataAccess_ViewBuildStageRunnerTrait so the |
| 43 |
// per-stage timing, shutdown logger and inflight marker share one |
| 44 |
// composing-class field set with the runner methods that read/write them. |
| 45 |
|
| 46 |
// $stagedQueryTimeoutSeconds is declared on the sibling |
| 47 |
// ABJ_404_Solution_DataAccess_ViewBuildHelpersTrait so the same field |
| 48 |
// backs both stagedQueryOptions() (helpers trait) and the per-stage |
| 49 |
// writes performed by the orchestrator below. |
| 50 |
|
| 51 |
/** |
| 52 |
* Request-lifetime cache of viewDoneIsServeable(). The AJAX gate, the |
| 53 |
* progress reader, and the pending-build response share the same answer |
| 54 |
* within a single request; without this cache each call reissues a SHOW |
| 55 |
* TABLES probe through the centralized DAO, which on a slow host pays the |
| 56 |
* full diagnostic latency on every probe and pushes the gate response |
| 57 |
* over criterion 6's <2s budget. |
| 58 |
* |
| 59 |
* Reset to null on every fetch entry / write that mutates view_done so a |
| 60 |
* fresh request never sees a stale answer. |
| 61 |
* |
| 62 |
* @var bool|null |
| 63 |
*/ |
| 64 |
private $viewDoneIsServeableCache = null; |
| 65 |
|
| 66 |
// $viewBuildProgressOptionNames (the option-name registry) is declared |
| 67 |
// on the sibling ABJ_404_Solution_DataAccess_ViewBuildHelpersTrait so |
| 68 |
// the progress get/set/clear helpers and this orchestrator share one |
| 69 |
// registry. self::$viewBuildProgressOptionNames resolves to the same |
| 70 |
// composing-class property regardless of which trait references it. |
| 71 |
|
| 72 |
/** @return void */ |
| 73 |
public static function resetViewBuildOncePerRequestGuard(): void { |
| 74 |
self::$viewBuildAlreadyRanThisRequest = false; |
| 75 |
self::$viewBuildShutdownLoggerRegistered = false; |
| 76 |
} |
| 77 |
|
| 78 |
/** @return string */ |
| 79 |
private function viewBuildTableName(): string { |
| 80 |
return $this->doTableNameReplacements('{wp_abj404_view_build}'); |
| 81 |
} |
| 82 |
|
| 83 |
/** @return string */ |
| 84 |
private function viewDoneTableName(): string { |
| 85 |
return $this->doTableNameReplacements('{wp_abj404_view_done}'); |
| 86 |
} |
| 87 |
|
| 88 |
/** @return string */ |
| 89 |
private function viewDeletemeTableName(): string { |
| 90 |
return $this->doTableNameReplacements('{wp_abj404_view_deleteme}'); |
| 91 |
} |
| 92 |
|
| 93 |
/** @return string */ |
| 94 |
private function viewDoneFreshnessOptionName(): string { |
| 95 |
return $this->getLowercasePrefix() . 'abj404_view_done_built_at'; |
| 96 |
} |
| 97 |
|
| 98 |
// Foreground admin/AJAX flows hold this lease briefly so staged-build |
| 99 |
// diagnostics reach the browser instead of being hidden inside cron. Cron |
| 100 |
// checks foregroundViewBuildLeaseActive() and reschedules itself instead |
| 101 |
// of taking the build lock while the lease is held. |
| 102 |
/** @return void */ |
| 103 |
public function claimForegroundViewBuildLease(): void { |
| 104 |
if (!function_exists('update_option')) { return; } |
| 105 |
update_option($this->getLowercasePrefix() . 'abj404_view_build_foreground_until', |
| 106 |
time() + ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_FOREGROUND_LEASE_SECONDS, false); |
| 107 |
} |
| 108 |
/** @return bool */ |
| 109 |
private function foregroundViewBuildLeaseActive(): bool { |
| 110 |
if (!function_exists('get_option')) { return false; } |
| 111 |
$until = get_option($this->getLowercasePrefix() . 'abj404_view_build_foreground_until', 0); |
| 112 |
return is_scalar($until) && intval($until) > time(); |
| 113 |
} |
| 114 |
|
| 115 |
/** |
| 116 |
* Public entry: returns the page of rows the admin Redirects/Captured |
| 117 |
* tab should render. Read-only with respect to view_done; never runs |
| 118 |
* the staged build inline. If view_done is missing or invalidated, a |
| 119 |
* background rebuild is scheduled and ABJ_404_Solution_ViewBuildPendingException |
| 120 |
* is thrown so the caller can translate it into a pending response. |
| 121 |
* |
| 122 |
* The fetch AJAX handler (ViewUpdater::getPaginationLinks) gates on |
| 123 |
* viewDoneIsServeable() before calling this method, so under normal |
| 124 |
* traffic this never throws. Non-AJAX callers (REST API, snapshot |
| 125 |
* warmup pipeline, tests) can hit the pending path; they handle it |
| 126 |
* by retrying once cron / the JS poller advances the build. |
| 127 |
* |
| 128 |
* @param string $sub |
| 129 |
* @param array<string, mixed> $tableOptions |
| 130 |
* @return array<int, array<string, mixed>> |
| 131 |
* @throws ABJ_404_Solution_ViewBuildPendingException |
| 132 |
*/ |
| 133 |
public function runRedirectsForViewStaged(string $sub, array $tableOptions): array { |
| 134 |
// Honor _abj404_query_timeout from the warmup pipeline so staged |
| 135 |
// queries inherit the same per-stage budget legacy code did. Reset |
| 136 |
// on entry so a previous request's value cannot leak across calls. |
| 137 |
$this->stagedQueryTimeoutSeconds = isset($tableOptions['_abj404_query_timeout']) |
| 138 |
&& is_numeric($tableOptions['_abj404_query_timeout']) |
| 139 |
? max(0, intval($tableOptions['_abj404_query_timeout'])) : 0; |
| 140 |
if (!empty($tableOptions['_abj404_force_view_rebuild'])) { |
| 141 |
// Diagnostic ?_abj404_force_view_rebuild=1 path: discard the |
| 142 |
// runner's in-flight state and start fresh. Non-blocking |
| 143 |
// acquire: if a sibling cron / AJAX advance holds the lock, |
| 144 |
// we skip the cleanup and fall through to serve-stale; the |
| 145 |
// already-running build will publish on its own. This is the |
| 146 |
// Phase 3a / Phase 4 successor to the direct invalidateViewDone() |
| 147 |
// pre-call -- the runner-owned primitive preserves the published |
| 148 |
// view_done snapshot for parallel readers until the new S11 |
| 149 |
// RENAME swap, which is the intended force-rebuild contract. |
| 150 |
$this->forceRestartViewBuild(0); |
| 151 |
} |
| 152 |
$builtAt = $this->viewDoneBuiltAt(); |
| 153 |
$isFresh = $builtAt > 0 |
| 154 |
&& (time() - $builtAt) < ABJ_404_Solution_ViewBuildConfig::VIEW_DONE_FRESHNESS_TTL_SECONDS |
| 155 |
&& $this->viewDoneIsServeable(); |
| 156 |
|
| 157 |
if ($isFresh) { |
| 158 |
return $this->readFromViewDone($sub, $tableOptions); |
| 159 |
} |
| 160 |
|
| 161 |
// Stale or invalidated: if view_done is serveable (table exists with |
| 162 |
// rows on disk) we return the snapshot now and kick off a background |
| 163 |
// rebuild for the next request. Invalidated counts as "stale-but- |
| 164 |
// present"; the freshness signal was cleared (so the rebuild gets |
| 165 |
// scheduled) but the data on disk is the most recent successful |
| 166 |
// snapshot and is correct to serve. Hard-stale notice fires when |
| 167 |
// the data on disk exceeds VIEW_DONE_HARD_STALE_NOTICE_AGE_SECONDS. |
| 168 |
if ($this->viewDoneIsServeable()) { |
| 169 |
$this->scheduleViewDoneRebuild(); |
| 170 |
$this->maybeRaiseViewDoneHardStaleNotice(); |
| 171 |
return $this->readFromViewDone($sub, $tableOptions); |
| 172 |
} |
| 173 |
|
| 174 |
// view_done is missing on disk or empty (no usable data). Schedule a |
| 175 |
// background rebuild (cron + ajaxAdvanceViewBuild advance the build) |
| 176 |
// and signal pending back up. Inline build inside a fetch request is |
| 177 |
// intentionally removed: on slow hosts it fatals at max_execution_time |
| 178 |
// and the HTTP 500 / "critical error" payload defeats client-side |
| 179 |
// recovery. |
| 180 |
$this->scheduleViewDoneRebuild(); |
| 181 |
$progress = $this->describeBuildProgressForNotice(); |
| 182 |
throw new ABJ_404_Solution_ViewBuildPendingException( |
| 183 |
'Staged view build pending; background rebuild scheduled. Progress: ' . $progress, |
| 184 |
$progress |
| 185 |
); |
| 186 |
} |
| 187 |
|
| 188 |
/** @return int Unix timestamp of last successful build, or 0 if missing. */ |
| 189 |
private function viewDoneBuiltAt(): int { |
| 190 |
if (!function_exists('get_option')) { |
| 191 |
return 0; |
| 192 |
} |
| 193 |
$built = get_option($this->viewDoneFreshnessOptionName(), 0); |
| 194 |
return is_scalar($built) ? max(0, intval($built)) : 0; |
| 195 |
} |
| 196 |
|
| 197 |
/** |
| 198 |
* Public read-only check used by the AJAX fetch endpoints to gate "serve |
| 199 |
* from cache vs. return pending". True when view_done exists on disk and |
| 200 |
* contains rows. Stale-but-present is serveable: invalidate clears the |
| 201 |
* freshness signal but leaves the table contents intact, so the steady- |
| 202 |
* state warm path serves stale and schedules a background rebuild |
| 203 |
* without blocking. |
| 204 |
* |
| 205 |
* Note: serveability does NOT depend on the freshness/built_at signal. |
| 206 |
* That signal gates whether to schedule a rebuild (stale = schedule), not |
| 207 |
* whether the existing data can be returned. Serving stale-but-present |
| 208 |
* data is correct: the data is at most one freshness window out of date |
| 209 |
* relative to when it was last produced, and the maybeRaiseViewDoneHard |
| 210 |
* StaleNotice() check fires an admin notice when staleness exceeds the |
| 211 |
* upper bound (VIEW_DONE_HARD_STALE_NOTICE_AGE_SECONDS). |
| 212 |
* |
| 213 |
* The ViewUpdater AJAX path uses this to avoid triggering the inline |
| 214 |
* build inside a request: if not serveable, the fetch returns |
| 215 |
* `viewBuildPending: true` and the JS poller hits ajaxAdvanceViewBuild. |
| 216 |
* |
| 217 |
* @return bool |
| 218 |
*/ |
| 219 |
public function viewDoneIsServeable(): bool { |
| 220 |
if ($this->viewDoneIsServeableCache !== null) { |
| 221 |
return $this->viewDoneIsServeableCache; |
| 222 |
} |
| 223 |
if (!$this->viewDoneTableExists()) { |
| 224 |
$this->viewDoneIsServeableCache = false; |
| 225 |
return false; |
| 226 |
} |
| 227 |
// Admin-mutation gate (Phase 4 watermark mechanism, owned by |
| 228 |
// ABJ_404_Solution_DataAccess_AdminMutationGateTrait). Blocks |
| 229 |
// reads while built_watermark < the watermark the admin observed |
| 230 |
// at click-Save time, OR until the sanity window elapses. |
| 231 |
if ($this->adminMutationGateBlocks()) { |
| 232 |
$this->viewDoneIsServeableCache = false; |
| 233 |
return false; |
| 234 |
} |
| 235 |
// Empty view_done is NOT serveable when there has never been a |
| 236 |
// successful build: rendering an empty admin screen during a cold |
| 237 |
// start is a worse UX than a brief pending/loading state that drives |
| 238 |
// the build forward. The has-rows probe also catches the rare "S11 |
| 239 |
// promoted an empty buffer" failure mode where the swap completed |
| 240 |
// but S2 produced no rows (botched build state); without this guard |
| 241 |
// the admin would render blank indefinitely with no rebuild ever |
| 242 |
// scheduled. |
| 243 |
// |
| 244 |
// BUT: when a build has actually completed (data_built_at > 0) and |
| 245 |
// the table is genuinely empty (e.g. a fresh install with no |
| 246 |
// redirects yet, or the admin dropped wp_abj404_redirects via WP-CLI |
| 247 |
// and the recreated table is empty), an empty view_done IS the |
| 248 |
// correct serveable result. Returning false here would loop the JS |
| 249 |
// poller forever on a cold install: every build cycle produces an |
| 250 |
// empty view_done, viewDoneIsServeable() returns false, ViewUpdater |
| 251 |
// returns viewBuildPending, the poller fires another advance, and |
| 252 |
// the cycle repeats with no exit. data_built_at distinguishes |
| 253 |
// "build has never completed" from "build completed and the dataset |
| 254 |
// is genuinely empty". |
| 255 |
if ($this->viewDoneHasRows()) { |
| 256 |
$this->viewDoneIsServeableCache = true; |
| 257 |
return true; |
| 258 |
} |
| 259 |
if ($this->viewDoneDataBuiltAt() > 0) { |
| 260 |
$this->viewDoneIsServeableCache = true; |
| 261 |
return true; |
| 262 |
} |
| 263 |
$this->viewDoneIsServeableCache = false; |
| 264 |
return false; |
| 265 |
} |
| 266 |
|
| 267 |
/** |
| 268 |
* Public accessor for the unix-time the view_done snapshot was last |
| 269 |
* successfully built. Returns 0 when never built or when the freshness |
| 270 |
* option has been cleared by an invalidation. Used by the admin footer |
| 271 |
* (and any diagnostic surface) to render a "Cache view freshness: 5m" |
| 272 |
* indicator without exposing the internal option name. |
| 273 |
* |
| 274 |
* @return int Unix timestamp, or 0. |
| 275 |
*/ |
| 276 |
public function getViewDoneBuiltAtTimestamp(): int { |
| 277 |
return $this->viewDoneBuiltAt(); |
| 278 |
} |
| 279 |
|
| 280 |
/** |
| 281 |
* Invalidate the request-lifetime serveability cache. Called from any |
| 282 |
* code path that mutates view_done (rename/drop/build completion) so a |
| 283 |
* subsequent read in the same request sees fresh state. |
| 284 |
* |
| 285 |
* @return void |
| 286 |
*/ |
| 287 |
private function invalidateViewDoneServeableCache(): void { |
| 288 |
$this->viewDoneIsServeableCache = null; |
| 289 |
} |
| 290 |
|
| 291 |
/** |
| 292 |
* Public hook called from the S11 swap completion path and from the |
| 293 |
* reconcile-promote path when a fresh view_done has just been published. |
| 294 |
* Updates both freshness and data-built-at signals to now, clears the |
| 295 |
* hard-stale admin notice (self-heal), and resets the request-lifetime |
| 296 |
* serveability cache so subsequent reads in the same request see the |
| 297 |
* just-published table. |
| 298 |
* |
| 299 |
* The data-built-at signal is the floor used by maybeRaiseViewDoneHard |
| 300 |
* StaleNotice() to decide when to surface the "data may be out of date" |
| 301 |
* admin notice. Updating it here means the notice can self-clear |
| 302 |
* automatically once the build catches up, so an admin who fixed the |
| 303 |
* underlying cron or host issue does not see a 24h-stale warning for |
| 304 |
* the entire dedup TTL after recovery. |
| 305 |
* |
| 306 |
* @return void |
| 307 |
*/ |
| 308 |
public function markViewDoneBuildCompleted(): void { |
| 309 |
if (function_exists('update_option')) { |
| 310 |
$now = $this->clock()->now(); |
| 311 |
update_option($this->viewDoneFreshnessOptionName(), $now, false); |
| 312 |
update_option($this->viewDoneDataBuiltAtOptionName(), $now, false); |
| 313 |
} |
| 314 |
$this->clearViewDoneHardStaleNotice(); |
| 315 |
// Clear the admin-mutation gate: the fresh build covers any |
| 316 |
// mutation that triggered it, so viewDoneIsServeable() no longer |
| 317 |
// needs to block reads. Leaving it set would force "Loading |
| 318 |
// redirects" for the full 5-minute sanity window after every |
| 319 |
// admin save even though fresh data is on disk. |
| 320 |
$this->clearAdminMutationGateOptions(); |
| 321 |
$this->invalidateViewDoneServeableCache(); |
| 322 |
} |
| 323 |
|
| 324 |
/** |
| 325 |
* Public progress snapshot used by the AJAX fetch endpoints when they |
| 326 |
* return a pending response, and by the build-advance endpoint after each |
| 327 |
* tick. Always safe to call; never queries beyond cheap option reads |
| 328 |
* plus a SHOW TABLES probe. |
| 329 |
* |
| 330 |
* Shape: |
| 331 |
* - status: 'ready' (view_done is serveable) or 'pending' |
| 332 |
* - stage: current sub-stage (0..11) reached so far |
| 333 |
* - of: 11 (total number of build sub-stages) |
| 334 |
* - build_started: unix ts when this resumable build began (0 if none) |
| 335 |
* - progress_text: short human-readable summary (e.g. "stage 2/11") |
| 336 |
* - fingerprint: per-tick mutation counters used by the JS poller |
| 337 |
* to detect within-stage progress (S2/S4/S5 advance |
| 338 |
* high-water ids across multiple ticks while |
| 339 |
* current_stage stays the same). The poller gives |
| 340 |
* up only when this fingerprint stops changing. |
| 341 |
* |
| 342 |
* @return array<string, mixed> |
| 343 |
*/ |
| 344 |
public function getViewBuildProgress(): array { |
| 345 |
$stage = $this->readProgressOption('current_stage', 0); |
| 346 |
$startedAt = $this->readProgressOption('started_at', 0); |
| 347 |
$status = $this->viewDoneIsServeable() ? 'ready' : 'pending'; |
| 348 |
return array( |
| 349 |
'status' => $status, |
| 350 |
'stage' => max(0, $stage), |
| 351 |
'of' => 11, |
| 352 |
'build_started' => max(0, $startedAt), |
| 353 |
// Cheap text derived from option reads only. The row-count rich |
| 354 |
// version (describeBuildProgressForNotice) issues two extra DAO |
| 355 |
// queries (countViewBuildRows + countLiveRedirects) which on a |
| 356 |
// slow host can multiply the gate's response time several-fold. |
| 357 |
// Callers that want the rich text can call describeBuildProgressForNotice |
| 358 |
// directly; AJAX gate / poll responses use the cheap form. |
| 359 |
'progress_text' => $stage > 0 ? ('stage ' . $stage . '/11') : 'not yet started', |
| 360 |
'fingerprint' => $this->getViewBuildProgressFingerprint(), |
| 361 |
); |
| 362 |
} |
| 363 |
|
| 364 |
/** |
| 365 |
* Public bounded build-advance entry point used by ajaxAdvanceViewBuild. |
| 366 |
* Runs at most one resumable tick of the staged build (10s/stage budget; |
| 367 |
* yields mid-stage on S2/S4/S5). Idempotent: safe to call concurrently. |
| 368 |
* Competing callers fail to acquire the build lock and just return the |
| 369 |
* current progress. Returns the same shape as getViewBuildProgress() |
| 370 |
* with an additional `locked` bool that is true when this call did not |
| 371 |
* acquire the lock (another worker is already advancing the build). |
| 372 |
* |
| 373 |
* Errors during a tick propagate as exceptions; the caller (AJAX handler) |
| 374 |
* surfaces them. This is intentionally NOT silent: a failing build that |
| 375 |
* never advances would otherwise leave the JS poller spinning forever. |
| 376 |
* |
| 377 |
* @param bool $forceRebuild Diagnostic mode (?abj404_force_view_rebuild=1): |
| 378 |
* - skip the viewDoneIsServeable() short-circuit so we always run the |
| 379 |
* staged build under the caller's request context, making every |
| 380 |
* staged_build_s* sub-stage event visible in the AJAX debug log; |
| 381 |
* - wait up to 30s for the build lock so a sibling cron / tab build |
| 382 |
* finishes and we can take ownership of the next build cleanly; |
| 383 |
* - re-invalidate inside the locked region so we run a fresh build |
| 384 |
* rather than the data the prior lock holder just produced; |
| 385 |
* - reset the per-request once-guard so a force-rebuild always proceeds |
| 386 |
* even if a sibling code path already ran a build in this request. |
| 387 |
* |
| 388 |
* @return array<string, mixed> |
| 389 |
*/ |
| 390 |
public function advanceViewBuildOnce(bool $forceRebuild = false): array { |
| 391 |
if ($forceRebuild) { |
| 392 |
// Allow the build to run even if a sibling read path on the same |
| 393 |
// request already entered the once-guard; the diagnostic flow |
| 394 |
// explicitly wants to rerun. |
| 395 |
self::$viewBuildAlreadyRanThisRequest = false; |
| 396 |
} |
| 397 |
if (!$forceRebuild && $this->viewDoneIsServeable()) { |
| 398 |
return $this->getViewBuildProgress(); |
| 399 |
} |
| 400 |
// 10s wait when forced so a cron build mid-flight can release the |
| 401 |
// lock before we take it. Without this, force-rebuild would return |
| 402 |
// locked=true, the JS poller would back off, the cron build would |
| 403 |
// finish in the background under no AJAX context, and the next |
| 404 |
// poll would see view_done as fresh -- no stage diagnostics ever |
| 405 |
// reach the debug log. 10s leaves comfortable headroom inside a |
| 406 |
// 30s PHP request: the typical cron build completes in seconds, |
| 407 |
// and if it doesn't we still return locked=true and the JS poller |
| 408 |
// can retry on the next page load. |
| 409 |
$lockTimeoutSeconds = $forceRebuild ? 10 : 0; |
| 410 |
if (!$this->acquireViewBuildLock($lockTimeoutSeconds)) { |
| 411 |
// Force-rebuild lock losses are interesting: a 10s wait that |
| 412 |
// still failed means another build held the lock longer than |
| 413 |
// expected (cron stuck, sibling tab mid-S2/S4/S5, dead session |
| 414 |
// holding GET_LOCK). Always-locked is one of the symptoms |
| 415 |
// Bruno/Troy report when their build never finishes, so log |
| 416 |
// every miss with the path so we can tell which caller blocked. |
| 417 |
$this->logger->debugMessage(sprintf( |
| 418 |
'[staged] advanceViewBuildOnce: lock not acquired ' |
| 419 |
. '(forceRebuild=%s, waited up to %ds)', |
| 420 |
$forceRebuild ? 'true' : 'false', $lockTimeoutSeconds |
| 421 |
)); |
| 422 |
$progress = $this->getViewBuildProgress(); |
| 423 |
$progress['locked'] = true; |
| 424 |
return $progress; |
| 425 |
} |
| 426 |
try { |
| 427 |
if ($forceRebuild) { |
| 428 |
// Whatever the prior lock holder produced (cron, sibling tab, |
| 429 |
// a finished S11 swap) we discard inside the locked region |
| 430 |
// so the rebuild happens fresh under the caller's request |
| 431 |
// context. The inside-lock cleanup helper drops the buffer, |
| 432 |
// clears progress + S1 prefix capture, and clears the |
| 433 |
// active-build started-watermark stamp; it also flips the |
| 434 |
// per-request serveability cache so getViewBuildProgress() |
| 435 |
// at the end reflects the rebuilt state. Direct call to |
| 436 |
// the runner-owned inside-lock helper avoids reacquiring |
| 437 |
// the lock (we already hold it). |
| 438 |
$this->runForceRestartCleanupInsideLock(); |
| 439 |
// Force-rebuild also clears any per-stage permanent skip |
| 440 |
// markers and the build-halted gate from a prior host |
| 441 |
// failure. The admin pressed "rebuild" explicitly, so |
| 442 |
// re-attempting denied DDL is the intended action. |
| 443 |
$this->clearStagedBuildDegradedState(); |
| 444 |
} else { |
| 445 |
// Same runner-startup reconciliation as the cron entry: an |
| 446 |
// AJAX advance picking up after a previous crash must |
| 447 |
// preserve the buffer S2-S10 already built rather than |
| 448 |
// throwing it away to start over. Force-rebuild skips |
| 449 |
// this because the admin explicitly asked to rebuild |
| 450 |
// from scratch. |
| 451 |
$reconcileResult = $this->reconcileStagedTablesAtRunnerStartup(); |
| 452 |
if ($reconcileResult === 'promoted') { |
| 453 |
return $this->getViewBuildProgress(); |
| 454 |
} |
| 455 |
} |
| 456 |
$isComplete = $this->runStagedBuildOnce(); |
| 457 |
} finally { |
| 458 |
$this->releaseViewBuildLock(); |
| 459 |
} |
| 460 |
if ($isComplete) { |
| 461 |
return $this->getViewBuildProgress(); |
| 462 |
} |
| 463 |
// Yielded mid-stage; schedule a background tick so cron pushes forward |
| 464 |
// even if the JS poller stops. Cron respects the foreground lease. |
| 465 |
$this->scheduleViewDoneRebuild(); |
| 466 |
return $this->getViewBuildProgress(); |
| 467 |
} |
| 468 |
|
| 469 |
/** |
| 470 |
* Synchronous fallback that advances the staged view-build by one |
| 471 |
* tick on plugin admin page-load when WP-Cron is broken. Pairs with |
| 472 |
* the cron-stuck admin notice (c374): the notice tells the admin |
| 473 |
* cron is broken; this fallback unblocks the page in the meantime |
| 474 |
* so they do not stare at "Carregando redirecionamentos..." (the |
| 475 |
* Portuguese localization Bruno reported) forever while they fix |
| 476 |
* cron. |
| 477 |
* |
| 478 |
* Without this, hosts where DISABLE_WP_CRON is set in wp-config.php |
| 479 |
* AND no external system cron replaces it leave the staged build |
| 480 |
* stuck: the AJAX JS poller would advance it, but the poller only |
| 481 |
* fires after the page renders, and the fetch path hard-gates on |
| 482 |
* view_done being serveable. The admin sees the loading message |
| 483 |
* indefinitely on every page-load. |
| 484 |
* |
| 485 |
* Gates (cheap, in order): |
| 486 |
* 1. getCronStuckHours() < 24: cron is healthy enough; no fallback |
| 487 |
* needed. The 24h floor matches the cron-stuck notice (c374) |
| 488 |
* so the two signals fire together, not piecewise. |
| 489 |
* 2. viewDoneIsServeable() === true: the build is already done, |
| 490 |
* so there is nothing to advance. Free option-read check. |
| 491 |
* 3. abj404_page_load_fallback_advance transient set: a sibling |
| 492 |
* sub-request already ran the fallback inside the 60s window; |
| 493 |
* avoid burning a second stage of inline work in the same |
| 494 |
* admin burst. |
| 495 |
* |
| 496 |
* Bounding: |
| 497 |
* - A short-lived filter is registered on the per-stage budget |
| 498 |
* hook (abj404_view_build_per_stage_budget_seconds) so the |
| 499 |
* advance call inside the lock cannot spend the full 10s |
| 500 |
* default per-stage budget. The filter is removed in finally |
| 501 |
* so the next AJAX advance / cron tick sees the normal budget. |
| 502 |
* |
| 503 |
* Lock semantics: |
| 504 |
* - Delegates to advanceViewBuildOnce(false), which acquires the |
| 505 |
* build lock with a 0s timeout. A sibling cron / AJAX advance |
| 506 |
* already in flight returns immediately with locked=true and |
| 507 |
* this method reports reason='locked' without doing further |
| 508 |
* work. The build progress under the existing lock holder is |
| 509 |
* still being made; the admin's next page-load (after the gate |
| 510 |
* window) will try again. |
| 511 |
* |
| 512 |
* Caller contract (admin_init wrapper in 404-solution.php): |
| 513 |
* - Only call when is_admin() is true. |
| 514 |
* - Only call when the current user has the plugin-admin |
| 515 |
* capability so unauthenticated requests cannot trigger build |
| 516 |
* work. |
| 517 |
* - Wrap in try/catch; a failure here must not break admin page |
| 518 |
* rendering. Log at warning level so it does not generate dev |
| 519 |
* email reports per the self-healing philosophy. |
| 520 |
* |
| 521 |
* @return array{ran:bool, reason:string, progress:array<string,mixed>} |
| 522 |
*/ |
| 523 |
public function runPageLoadFallbackAdvance(): array { |
| 524 |
// Cron is healthy: nothing for the fallback to do. Free check |
| 525 |
// (one wp_get_ready_cron_jobs call) so we can run it first. |
| 526 |
if ($this->getCronStuckHours() < 24) { |
| 527 |
return array( |
| 528 |
'ran' => false, |
| 529 |
'reason' => 'cron_healthy', |
| 530 |
'progress' => $this->getViewBuildProgress(), |
| 531 |
); |
| 532 |
} |
| 533 |
|
| 534 |
// Build already serveable: returning before any further work |
| 535 |
// keeps the fallback's steady-state cost at zero on hosts that |
| 536 |
// recover, which is the desirable shape (admin returns to a |
| 537 |
// working page without page-load latency). |
| 538 |
if ($this->viewDoneIsServeable()) { |
| 539 |
return array( |
| 540 |
'ran' => false, |
| 541 |
'reason' => 'not_needed', |
| 542 |
'progress' => $this->getViewBuildProgress(), |
| 543 |
); |
| 544 |
} |
| 545 |
|
| 546 |
// Transient gate: a single admin click can produce many |
| 547 |
// sub-requests (prefetch, refresh, multiple browser tabs). |
| 548 |
// Cap inline advances to one per 60s so the page-load impact |
| 549 |
// cannot compound. 60s is short enough that an attentive admin |
| 550 |
// sees real per-load progress, and long enough that bursts of |
| 551 |
// navigation do not stack inline build work. |
| 552 |
$haveTransientApi = function_exists('get_transient') && function_exists('set_transient'); |
| 553 |
$gateKey = ABJ_404_Solution_ViewBuildConfig::PAGE_LOAD_FALLBACK_GATE_KEY; |
| 554 |
if ($haveTransientApi && get_transient($gateKey) !== false) { |
| 555 |
return array( |
| 556 |
'ran' => false, |
| 557 |
'reason' => 'gate_active', |
| 558 |
'progress' => $this->getViewBuildProgress(), |
| 559 |
); |
| 560 |
} |
| 561 |
if ($haveTransientApi) { |
| 562 |
// Set the gate BEFORE running the advance so any failure or |
| 563 |
// long-running stage still suppresses the next sub-request. |
| 564 |
// Without this, a slow advance that times out partway would |
| 565 |
// be retried by the very next sub-request, compounding the |
| 566 |
// page-load impact instead of bounding it. |
| 567 |
set_transient( |
| 568 |
$gateKey, |
| 569 |
1, |
| 570 |
(int)ABJ_404_Solution_ViewBuildConfig::PAGE_LOAD_FALLBACK_GATE_SECONDS |
| 571 |
); |
| 572 |
} |
| 573 |
|
| 574 |
// Compress the per-stage budget for just this advance. The |
| 575 |
// production filter machinery is the existing hook |
| 576 |
// abj404_view_build_per_stage_budget_seconds; clamping via |
| 577 |
// min() rather than overwriting preserves any operator-set |
| 578 |
// smaller-budget filter for hosts that have already tuned |
| 579 |
// down. Priority 100 runs after most operator filters so the |
| 580 |
// fallback's ceiling dominates. |
| 581 |
$budgetSeconds = (float)ABJ_404_Solution_ViewBuildConfig::PAGE_LOAD_FALLBACK_BUDGET_SECONDS; |
| 582 |
$budgetFilter = static function ($incoming) use ($budgetSeconds) { |
| 583 |
$value = is_scalar($incoming) ? (float)$incoming : $budgetSeconds; |
| 584 |
return min($value, $budgetSeconds); |
| 585 |
}; |
| 586 |
$filterRegistered = false; |
| 587 |
if (function_exists('add_filter')) { |
| 588 |
add_filter('abj404_view_build_per_stage_budget_seconds', $budgetFilter, 100); |
| 589 |
$filterRegistered = true; |
| 590 |
} |
| 591 |
|
| 592 |
try { |
| 593 |
// forceRebuild=false: this is the self-healing path. The |
| 594 |
// explicit ?abj404_force_view_rebuild=1 admin recovery is a |
| 595 |
// separate gesture that intentionally clears degraded gates |
| 596 |
// and waits 30s for the lock. Page-load fallback should |
| 597 |
// never escalate to those semantics. |
| 598 |
$progress = $this->advanceViewBuildOnce(false); |
| 599 |
} finally { |
| 600 |
if ($filterRegistered && function_exists('remove_filter')) { |
| 601 |
remove_filter('abj404_view_build_per_stage_budget_seconds', $budgetFilter, 100); |
| 602 |
} |
| 603 |
} |
| 604 |
|
| 605 |
$reason = !empty($progress['locked']) ? 'locked' : 'advanced'; |
| 606 |
return array( |
| 607 |
'ran' => true, |
| 608 |
'reason' => $reason, |
| 609 |
'progress' => $progress, |
| 610 |
); |
| 611 |
} |
| 612 |
|
| 613 |
/** |
| 614 |
* COUNT(*) sibling to runRedirectsForViewStaged. Used by |
| 615 |
* getRedirectsForViewCount when filterText is non-empty (the |
| 616 |
* filterText-empty path already uses the optimized COUNT against |
| 617 |
* the live redirects table, which stays fast). Same build/serve flow |
| 618 |
* as the row path; the build is shared via the per-request guard. |
| 619 |
* |
| 620 |
* @param string $sub |
| 621 |
* @param array<string, mixed> $tableOptions |
| 622 |
* @return int |
| 623 |
*/ |
| 624 |
public function runRedirectsForViewCountStaged(string $sub, array $tableOptions): int { |
| 625 |
$this->stagedQueryTimeoutSeconds = isset($tableOptions['_abj404_query_timeout']) |
| 626 |
&& is_numeric($tableOptions['_abj404_query_timeout']) |
| 627 |
? max(0, intval($tableOptions['_abj404_query_timeout'])) : 0; |
| 628 |
$builtAt = $this->viewDoneBuiltAt(); |
| 629 |
$isFresh = $builtAt > 0 |
| 630 |
&& (time() - $builtAt) < ABJ_404_Solution_ViewBuildConfig::VIEW_DONE_FRESHNESS_TTL_SECONDS |
| 631 |
&& $this->viewDoneIsServeable(); |
| 632 |
|
| 633 |
if (!$this->viewDoneIsServeable()) { |
| 634 |
// view_done is missing on disk or empty (no usable data). |
| 635 |
// Schedule a background rebuild and signal pending; never run |
| 636 |
// the staged build inline inside a request. The fetch AJAX gate |
| 637 |
// prevents this from being reached under normal traffic; non- |
| 638 |
// AJAX callers retry on next request. |
| 639 |
$this->scheduleViewDoneRebuild(); |
| 640 |
$progress = $this->describeBuildProgressForNotice(); |
| 641 |
throw new ABJ_404_Solution_ViewBuildPendingException( |
| 642 |
'Staged view-count build pending; background rebuild scheduled. Progress: ' . $progress, |
| 643 |
$progress |
| 644 |
); |
| 645 |
} |
| 646 |
|
| 647 |
if (!$isFresh) { |
| 648 |
// Stale or invalidated but data on disk is serveable: return the |
| 649 |
// stale count and kick off a background rebuild. Hard-stale |
| 650 |
// notice fires when the data on disk exceeds the upper bound. |
| 651 |
$this->scheduleViewDoneRebuild(); |
| 652 |
$this->maybeRaiseViewDoneHardStaleNotice(); |
| 653 |
} |
| 654 |
|
| 655 |
$sql = $this->buildViewDoneCountQuery($sub, $tableOptions); |
| 656 |
$result = $this->queryAndGetResults($sql, $this->stagedQueryOptions()); |
| 657 |
$rows = is_array($result['rows'] ?? null) ? $result['rows'] : array(); |
| 658 |
if (empty($rows)) { |
| 659 |
return 0; |
| 660 |
} |
| 661 |
$row = is_array($rows[0]) ? $rows[0] : array(); |
| 662 |
$raw = $row['cnt'] ?? reset($row); |
| 663 |
return is_scalar($raw) ? intval($raw) : 0; |
| 664 |
} |
| 665 |
|
| 666 |
/** |
| 667 |
* Hook target for `wp_schedule_single_event('abj404_rebuildViewDone')`. |
| 668 |
* Rebuilds inline (under the build lock) so the next admin request |
| 669 |
* sees fresh data. Called from PluginLogic during cron registration. |
| 670 |
* |
| 671 |
* @return void |
| 672 |
*/ |
| 673 |
public function rebuildViewDoneInBackground(): void { |
| 674 |
if ($this->foregroundViewBuildLeaseActive()) { |
| 675 |
$this->logger->debugMessage( |
| 676 |
'[staged] rebuildViewDoneInBackground: deferring; ' |
| 677 |
. 'foreground build lease active. Rescheduled.' |
| 678 |
); |
| 679 |
$this->scheduleViewDoneRebuild(ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_FOREGROUND_LEASE_SECONDS); |
| 680 |
return; |
| 681 |
} |
| 682 |
if (!$this->acquireViewBuildLock()) { |
| 683 |
$this->logger->debugMessage( |
| 684 |
'[staged] rebuildViewDoneInBackground: lock not acquired ' |
| 685 |
. '(another worker is building); skipping this cron tick.' |
| 686 |
); |
| 687 |
return; |
| 688 |
} |
| 689 |
try { |
| 690 |
// Runner-startup reconciliation: clean up inconsistent staged- |
| 691 |
// table state from a previous run that crashed mid-S11 or was |
| 692 |
// OOM-killed before the swap option write. Runs BEFORE |
| 693 |
// runStagedBuildOnce() so its halt-gate / once-guard short- |
| 694 |
// circuits cannot suppress the cleanup, and it can short- |
| 695 |
// circuit the rebuild itself when it manages to recover the |
| 696 |
// previous run's buffer in place. |
| 697 |
$reconcileResult = $this->reconcileStagedTablesAtRunnerStartup(); |
| 698 |
if ($reconcileResult === 'promoted') { |
| 699 |
// The previous run's view_build was renamed to view_done |
| 700 |
// in place; freshness is recorded; view_done is now |
| 701 |
// serveable. No need to re-run the staged build this tick. |
| 702 |
return; |
| 703 |
} |
| 704 |
$isComplete = $this->runStagedBuildOnce(); |
| 705 |
if (!$isComplete) { |
| 706 |
// Build yielded mid-stage; schedule another tick to continue. |
| 707 |
$this->scheduleViewDoneRebuild(); |
| 708 |
} |
| 709 |
} catch (Throwable $e) { |
| 710 |
// Log at warning level, not error: a failed background rebuild |
| 711 |
// leaves the plugin functional (the prior view_done snapshot, |
| 712 |
// if any, is still served). Per CLAUDE.md self-healing rules, |
| 713 |
// infrastructure failures should not generate dev email reports. |
| 714 |
$this->logger->warn('[staged] background rebuild yielded an error: ' . $e->getMessage()); |
| 715 |
} finally { |
| 716 |
$this->releaseViewBuildLock(); |
| 717 |
} |
| 718 |
} |
| 719 |
|
| 720 |
/** |
| 721 |
* Reconcile staged-build table state from a previous run that ended |
| 722 |
* in an inconsistent place, before this run's stages execute. Called |
| 723 |
* from rebuildViewDoneInBackground() AFTER the build lock is |
| 724 |
* acquired (so we cannot race a sibling worker on the same site) |
| 725 |
* and BEFORE runStagedBuildOnce() (so the staged orchestrator sees |
| 726 |
* a clean starting state regardless of which entry path took the |
| 727 |
* lock). |
| 728 |
* |
| 729 |
* Cases handled: |
| 730 |
* |
| 731 |
* 1. Orphan `{wp_abj404_view_deleteme}` from a prior crashed S11 |
| 732 |
* swap (or a critical-stage halt that left the previous run's |
| 733 |
* deleteme on disk). Drop it. Always safe; deleteme is a |
| 734 |
* transient by design. |
| 735 |
* |
| 736 |
* 2. `{wp_abj404_view_build}` exists, `{wp_abj404_view_done}` does |
| 737 |
* NOT, AND no resumable build progress is recorded: the |
| 738 |
* previous run completed S2-S10 but crashed before the S11 |
| 739 |
* RENAME swap published the buffer. Promote the buffer in |
| 740 |
* place via `RENAME TABLE view_build TO view_done`, mark |
| 741 |
* fresh, clear progress. Preserves the work of S2-S10 instead |
| 742 |
* of throwing it away. |
| 743 |
* |
| 744 |
* 3. Both `{wp_abj404_view_build}` and `{wp_abj404_view_done}` |
| 745 |
* exist, AND no resumable build progress is recorded: the |
| 746 |
* previous run halted between stages with both tables on |
| 747 |
* disk. Treat view_done as the live one and drop view_build |
| 748 |
* so the next fresh build starts from a known empty buffer. |
| 749 |
* |
| 750 |
* Resumable progress = `started_at` within |
| 751 |
* VIEW_BUILD_RESUME_TTL_SECONDS and either a completed stage |
| 752 |
* (`current_stage` > 0) or a durable started-stage marker. When a |
| 753 |
* resumable build is in flight we leave view_build alone so the |
| 754 |
* next tick can continue from the persisted high-water id (cases |
| 755 |
* 2 and 3 are skipped; case 1 still runs). |
| 756 |
* |
| 757 |
* Reconciliation actions are best-effort: when DROP / RENAME is |
| 758 |
* denied by the host (privilege loss between runs), surface a |
| 759 |
* deduplicated admin notice naming the specific tables and |
| 760 |
* recommending manual cleanup. The build then falls through to |
| 761 |
* runStagedBuildOnce() which will hit its own host-failure |
| 762 |
* classifier. |
| 763 |
* |
| 764 |
* @return string One of: |
| 765 |
* 'none' - no reconciliation needed. |
| 766 |
* 'cleaned' - orphan tables dropped; build can proceed. |
| 767 |
* 'promoted' - view_build was renamed to view_done; |
| 768 |
* view_done is fresh; rebuild can be |
| 769 |
* skipped this tick. |
| 770 |
* 'failed' - reconciliation could not complete |
| 771 |
* (privilege denied); admin notice set. |
| 772 |
*/ |
| 773 |
public function reconcileStagedTablesAtRunnerStartup(): string { |
| 774 |
$tempDeletemeTable = $this->viewDeletemeTableName(); |
| 775 |
$tempBuildTable = $this->viewBuildTableName(); |
| 776 |
$doneTable = $this->viewDoneTableName(); |
| 777 |
|
| 778 |
$action = 'none'; |
| 779 |
$haveDeleteme = $this->stagedTableExists($tempDeletemeTable); |
| 780 |
|
| 781 |
if ($haveDeleteme) { |
| 782 |
$r = $this->queryAndGetResults( |
| 783 |
'DROP TABLE IF EXISTS `' . $tempDeletemeTable . '`', |
| 784 |
array('log_errors' => false) |
| 785 |
); |
| 786 |
$err = isset($r['last_error']) && is_string($r['last_error']) ? trim($r['last_error']) : ''; |
| 787 |
if ($err === '' || !$this->stagedTableExists($tempDeletemeTable)) { |
| 788 |
$this->logger->infoMessage(sprintf( |
| 789 |
'[staged] reconcile: dropped orphan view_deleteme `%s` from a previous failed run', |
| 790 |
$tempDeletemeTable |
| 791 |
)); |
| 792 |
$action = 'cleaned'; |
| 793 |
} else { |
| 794 |
$this->logger->warn(sprintf( |
| 795 |
'[staged] reconcile: orphan view_deleteme `%s` could not be dropped: %s', |
| 796 |
$tempDeletemeTable, substr($err, 0, 200) |
| 797 |
)); |
| 798 |
$this->setStagedBuildHaltNotice('orphan_deleteme', sprintf( |
| 799 |
'An orphan staged-build buffer `%s` from a previous failed run could not be removed (privilege denied?): %s. Manual cleanup: drop the buffer table `%s` from your database (e.g. via phpMyAdmin or your hosting MySQL console).', |
| 800 |
$tempDeletemeTable, substr($err, 0, 200), $tempDeletemeTable |
| 801 |
)); |
| 802 |
$action = 'failed'; |
| 803 |
} |
| 804 |
} |
| 805 |
|
| 806 |
// Resumable build in flight? Leave $tempBuildTable / view_done alone |
| 807 |
// so the next tick can continue from the persisted high-water |
| 808 |
// id; orphan deleteme cleanup above already ran and is enough. |
| 809 |
$startedAt = $this->readProgressOption('started_at', 0); |
| 810 |
$currentStage = $this->readProgressOption('current_stage', 0); |
| 811 |
$lastStartedStage = $this->readProgressOption('last_started_stage', 0); |
| 812 |
$resumeWindowOk = $startedAt > 0 |
| 813 |
&& (time() - $startedAt) < ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_RESUME_TTL_SECONDS; |
| 814 |
if ($resumeWindowOk && ($currentStage > 0 || $lastStartedStage > 0)) { |
| 815 |
return $action; |
| 816 |
} |
| 817 |
|
| 818 |
$haveBuild = $this->stagedTableExists($tempBuildTable); |
| 819 |
$haveDone = $this->stagedTableExists($doneTable); |
| 820 |
|
| 821 |
// Case 2: view_build exists, view_done missing. Promote the |
| 822 |
// buffer in place rather than re-running S1-S11 from scratch |
| 823 |
// -- but only when an integrity probe says the buffer is |
| 824 |
// plausibly complete. Without the integrity check we could |
| 825 |
// publish a partially-built buffer (S2 stopped halfway, or a |
| 826 |
// force-rebuild cleared progress while a redirect edit had |
| 827 |
// also added rows we never picked up). |
| 828 |
if ($haveBuild && !$haveDone) { |
| 829 |
if (!$this->bufferIntegrityPassesForPromote($tempBuildTable)) { |
| 830 |
$this->logger->infoMessage(sprintf( |
| 831 |
'[staged] reconcile: not promoting view_build `%s` (integrity probe failed); dropping for fresh rebuild', |
| 832 |
$tempBuildTable |
| 833 |
)); |
| 834 |
$this->queryAndGetResults('DROP TABLE IF EXISTS `' . $tempBuildTable . '`', |
| 835 |
array('log_errors' => false)); |
| 836 |
return 'cleaned'; |
| 837 |
} |
| 838 |
$sql = 'RENAME TABLE `' . $tempBuildTable . '` TO `' . $doneTable . '`'; |
| 839 |
$r = $this->queryAndGetResults($sql, array('log_errors' => true)); |
| 840 |
$err = isset($r['last_error']) && is_string($r['last_error']) ? trim($r['last_error']) : ''; |
| 841 |
if ($err === '' && $this->stagedTableExists($doneTable)) { |
| 842 |
$this->logger->infoMessage(sprintf( |
| 843 |
'[staged] reconcile: promoted view_build to view_done ' |
| 844 |
. '(`%s` -> `%s`); previous run crashed before S11 swap', |
| 845 |
$tempBuildTable, $doneTable |
| 846 |
)); |
| 847 |
// Same as the S11 swap completion: update both freshness |
| 848 |
// signals, clear hard-stale notice, reset serveability cache. |
| 849 |
$this->markViewDoneBuildCompleted(); |
| 850 |
$this->clearAllProgressOptions(); |
| 851 |
return 'promoted'; |
| 852 |
} |
| 853 |
$this->logger->warn(sprintf( |
| 854 |
'[staged] reconcile: could not promote view_build to view_done: %s', |
| 855 |
substr($err, 0, 200) |
| 856 |
)); |
| 857 |
$this->setStagedBuildHaltNotice('promote_build_failed', sprintf( |
| 858 |
'A staged-build buffer `%s` exists from a previous run but could not be promoted to `%s` (privilege denied?): %s. Manual cleanup: rename the buffer `%s` to `%s`, or remove the buffer `%s` from your database (e.g. via phpMyAdmin or your hosting MySQL console).', |
| 859 |
$tempBuildTable, $doneTable, substr($err, 0, 200), |
| 860 |
$tempBuildTable, $doneTable, $tempBuildTable |
| 861 |
)); |
| 862 |
return 'failed'; |
| 863 |
} |
| 864 |
|
| 865 |
// Case 3: both tables exist. view_done is the live one; the |
| 866 |
// orphan $tempBuildTable is from a halted previous run. Drop it so |
| 867 |
// the next fresh build starts from a known empty buffer. |
| 868 |
if ($haveBuild && $haveDone) { |
| 869 |
$r = $this->queryAndGetResults( |
| 870 |
'DROP TABLE IF EXISTS `' . $tempBuildTable . '`', |
| 871 |
array('log_errors' => false) |
| 872 |
); |
| 873 |
$err = isset($r['last_error']) && is_string($r['last_error']) ? trim($r['last_error']) : ''; |
| 874 |
if ($err === '' || !$this->stagedTableExists($tempBuildTable)) { |
| 875 |
// WARN (not INFO) so this signal survives a site with DEBUG |
| 876 |
// disabled. Carries the four progress fields support needs |
| 877 |
// to distinguish "build keeps restarting at S1" from |
| 878 |
// "build invalidated on every redirect edit / cron tick" |
| 879 |
// from "multi-tab/cron lock contention orphaning each |
| 880 |
// partial build" without asking for another debug zip. |
| 881 |
$lastCompletedStage = $this->readProgressOption('last_completed_stage', 0); |
| 882 |
$age = $startedAt > 0 ? max(0, time() - $startedAt) : 0; |
| 883 |
$this->logger->warn(sprintf( |
| 884 |
'[staged] reconcile: dropped orphan view_build `%s` (view_done is live; previous run halted before swap); ' |
| 885 |
. 'current_stage=%d last_started_stage=%d last_completed_stage=%d started_at=%d age=%ds', |
| 886 |
$tempBuildTable, |
| 887 |
$currentStage, |
| 888 |
$lastStartedStage, |
| 889 |
$lastCompletedStage, |
| 890 |
$startedAt, |
| 891 |
$age |
| 892 |
)); |
| 893 |
return 'cleaned'; |
| 894 |
} |
| 895 |
$this->logger->warn(sprintf( |
| 896 |
'[staged] reconcile: orphan view_build `%s` could not be dropped: %s', |
| 897 |
$tempBuildTable, substr($err, 0, 200) |
| 898 |
)); |
| 899 |
$this->setStagedBuildHaltNotice('orphan_build', sprintf( |
| 900 |
'A staged-build buffer `%s` from a previous run still exists alongside the live view_done, but could not be removed: %s. Manual cleanup: drop the buffer `%s` from your database (e.g. via phpMyAdmin or your hosting MySQL console).', |
| 901 |
$tempBuildTable, substr($err, 0, 200), $tempBuildTable |
| 902 |
)); |
| 903 |
return 'failed'; |
| 904 |
} |
| 905 |
|
| 906 |
return $action; |
| 907 |
} |
| 908 |
|
| 909 |
/** |
| 910 |
* Integrity probe used by the case-2 promote branch of |
| 911 |
* reconcileStagedTablesAtRunnerStartup(). Returns true only when the |
| 912 |
* buffer is plausibly complete: row count matches the live redirects |
| 913 |
* table within a small tolerance (one redirect could have been |
| 914 |
* added during the build window). Returns false on any probe error |
| 915 |
* so a transient DB hiccup never publishes a buffer of unknown |
| 916 |
* shape as the live snapshot. |
| 917 |
* |
| 918 |
* Row-count parity is a coarse check (it cannot detect stale POST |
| 919 |
* resolutions when wp_posts has changed mid-flight). Promote is |
| 920 |
* already an opportunistic recovery; if we're wrong, the next |
| 921 |
* invalidate-driven rebuild will replace view_done. |
| 922 |
* |
| 923 |
* @param string $bufferTable |
| 924 |
* @return bool |
| 925 |
*/ |
| 926 |
private function bufferIntegrityPassesForPromote(string $bufferTable): bool { |
| 927 |
$bufferRows = $this->countViewBuildRows(); |
| 928 |
if ($bufferRows <= 0) { |
| 929 |
return false; |
| 930 |
} |
| 931 |
$liveRows = $this->countLiveRedirects(); |
| 932 |
if ($liveRows <= 0) { |
| 933 |
// No redirects in the live table -- treat any buffer as |
| 934 |
// unsafe to publish (a bug pruned all redirects, or the |
| 935 |
// count probe itself errored). |
| 936 |
return false; |
| 937 |
} |
| 938 |
// Allow the buffer to differ from live by up to one row in |
| 939 |
// either direction so an admin who created or deleted a single |
| 940 |
// redirect during the build window does not block promotion. |
| 941 |
$diff = abs($bufferRows - $liveRows); |
| 942 |
return $diff <= 1; |
| 943 |
} |
| 944 |
|
| 945 |
// invalidateViewDone() was deleted in Phase 4 of the staged view-build |
| 946 |
// watermark refactor (see docs/refactor-staged-view-build-watermark.md). |
| 947 |
// The god-method conflated three concepts (logical invalidation, runner |
| 948 |
// lifecycle, reader policy) and was the seam through which external code |
| 949 |
// destroyed runner-owned state. Replacements, by intent: |
| 950 |
// |
| 951 |
// - Source data changed (admin/REST/CLI/AJAX/cron mutation): call |
| 952 |
// bumpMutationWatermark() (DataAccessTrait_MutationWatermarkSeam). |
| 953 |
// For admin form actions that also need the strict admin-visibility |
| 954 |
// gate, call markViewDoneInvalidatedByAdminMutation() |
| 955 |
// (DataAccessTrait_AdminMutationGate) which composes the watermark |
| 956 |
// bump with the observed-watermark gate option. |
| 957 |
// - Discard + restart the in-flight build (admin "rebuild now", |
| 958 |
// diagnostic ?abj404_force_view_rebuild=1 paths): call |
| 959 |
// forceRestartViewBuild() (DataAccessTrait_ViewBuildForceRestart). |
| 960 |
// - Schedule a cron rebuild without other side effects: call |
| 961 |
// scheduleViewDoneRebuild() (DataAccessTrait_ViewBuildLockAndCron). |
| 962 |
// |
| 963 |
// The semantic-forbidden-operation lint in StagedBuildOwnershipLintTest |
| 964 |
// (lint d) enforces that no future code introduces a new caller of |
| 965 |
// ->invalidateViewDone( anywhere in includes/. |
| 966 |
|
| 967 |
// progressOptionName / readProgressOption / writeProgressOption / |
| 968 |
// clearAllProgressOptions live on the sibling |
| 969 |
// ABJ_404_Solution_DataAccess_ViewBuildHelpersTrait. They read and |
| 970 |
// write the progress option registry declared in that trait. |
| 971 |
|
| 972 |
/** |
| 973 |
* Read the configured per-batch row count. Honors: |
| 974 |
* - define('ABJ404_VIEW_BUILD_BATCH_SIZE', N) for tests/operators |
| 975 |
* - apply_filters('abj404_view_build_batch_size', N) for site overrides |
| 976 |
* |
| 977 |
* @return int Always >= 1. |
| 978 |
*/ |
| 979 |
private function viewBuildBatchSize(): int { |
| 980 |
$size = ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_DEFAULT_BATCH_SIZE; |
| 981 |
if (defined('ABJ404_VIEW_BUILD_BATCH_SIZE')) { |
| 982 |
$size = intval(ABJ404_VIEW_BUILD_BATCH_SIZE); |
| 983 |
} |
| 984 |
if (function_exists('apply_filters')) { |
| 985 |
$filtered = apply_filters('abj404_view_build_batch_size', $size); |
| 986 |
if (is_scalar($filtered)) { |
| 987 |
$size = intval($filtered); |
| 988 |
} |
| 989 |
} |
| 990 |
return max(1, $size); |
| 991 |
} |
| 992 |
|
| 993 |
|
| 994 |
/** |
| 995 |
* Wall-clock budget after which a batched stage (S2 / S4 / S5) yields |
| 996 |
* to the next request rather than starting another batch. This is NOT a |
| 997 |
* query cancellation: any in-flight INSERT or UPDATE-JOIN runs to its |
| 998 |
* own MySQL timeout. It only stops the loop from issuing more batches |
| 999 |
* once the request is close to the real ceiling. |
| 1000 |
* |
| 1001 |
* Default: derive from PHP's max_execution_time minus a 2s response |
| 1002 |
* cushion. WP-CLI / cron with no PHP time limit (max_execution_time = 0) |
| 1003 |
* fall back to ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_PER_STAGE_BUDGET_SECONDS |
| 1004 |
* since unbounded loops would still be undesirable in those contexts. |
| 1005 |
* |
| 1006 |
* Explicit overrides (define / filter) win over auto-detection so |
| 1007 |
* operators can tune it for their host. The minimum floor of 0.1s is |
| 1008 |
* preserved so tests that set a tiny override still complete a batch. |
| 1009 |
* |
| 1010 |
* @return float Seconds; always > 0. |
| 1011 |
*/ |
| 1012 |
private function viewBuildPerStageBudgetSeconds(): float { |
| 1013 |
$explicitOverride = false; |
| 1014 |
$budget = (float)ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_PER_STAGE_BUDGET_SECONDS; |
| 1015 |
|
| 1016 |
if (defined('ABJ404_VIEW_BUILD_PER_STAGE_BUDGET_SECONDS')) { |
| 1017 |
$budget = (float)ABJ404_VIEW_BUILD_PER_STAGE_BUDGET_SECONDS; |
| 1018 |
$explicitOverride = true; |
| 1019 |
} |
| 1020 |
|
| 1021 |
// When set_time_limit() is in disable_functions, the build cannot |
| 1022 |
// extend its time mid-request. Widen the cushion (4s vs. 2s) and |
| 1023 |
// cap below the default budget so each tick yields earlier and the |
| 1024 |
// next cron tick resumes inside its own fresh request budget. |
| 1025 |
$setTimeLimitAvailable = $this->probeSetTimeLimitAvailability(); |
| 1026 |
|
| 1027 |
if (!$explicitOverride) { |
| 1028 |
$maxExec = (int)ini_get('max_execution_time'); |
| 1029 |
if ($maxExec >= 5) { |
| 1030 |
$cushion = $setTimeLimitAvailable ? 2 : 4; |
| 1031 |
$budget = (float)max(1, $maxExec - $cushion); |
| 1032 |
} |
| 1033 |
} |
| 1034 |
if (!$explicitOverride && !$setTimeLimitAvailable) { |
| 1035 |
$tightCap = max( |
| 1036 |
1.0, |
| 1037 |
(float)ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_PER_STAGE_BUDGET_SECONDS - 4.0 |
| 1038 |
); |
| 1039 |
$budget = min($budget, $tightCap); |
| 1040 |
} |
| 1041 |
|
| 1042 |
if (function_exists('apply_filters')) { |
| 1043 |
$filtered = apply_filters('abj404_view_build_per_stage_budget_seconds', $budget); |
| 1044 |
if (is_scalar($filtered)) { |
| 1045 |
$budget = (float)$filtered; |
| 1046 |
} |
| 1047 |
} |
| 1048 |
return $budget > 0.1 ? $budget : 0.1; |
| 1049 |
} |
| 1050 |
|
| 1051 |
/** |
| 1052 |
* Release the build lock and immediately reacquire it (non-blocking). |
| 1053 |
* Called between every stage of the staged view build so the lock hold |
| 1054 |
* is bounded by the duration of a single stage rather than the cumulative |
| 1055 |
* 11-stage wall time. |
| 1056 |
* |
| 1057 |
* Why this exists: production error reports #14 (ajasha.de, S4), #15 |
| 1058 |
* (greyleafmedia.com, S11), #16 (p2p-game.com, S5), and #17 (remiancelin.fr, |
| 1059 |
* S11), all plugin 4.1.18 via wp-cron, traced to a connection-drop race. |
| 1060 |
* GET_LOCK was acquired once and held across the full S1-S11 run. On |
| 1061 |
* shared hosting the MySQL connection dropped mid-hold (wait_timeout or |
| 1062 |
* pool eviction); wpdb auto-reconnected with no lock; a second cron tick |
| 1063 |
* acquired the (now freed) lock, ran reconcile Case 3, and dropped |
| 1064 |
* view_build out from under the first worker. The first worker then |
| 1065 |
* queried a missing table: "Table doesn't exist" (MariaDB) or "Can't find |
| 1066 |
* .frm file" (MySQL 5.7). |
| 1067 |
* |
| 1068 |
* Per-stage release bounds the connection-drop exposure window to a |
| 1069 |
* single stage, and a sibling worker that takes the lock between our |
| 1070 |
* stages sees the persisted current_stage / started_at and resumes from |
| 1071 |
* where we left off rather than running reconcile against a partial build. |
| 1072 |
* |
| 1073 |
* @return bool true when the lock was reacquired and the caller should |
| 1074 |
* continue with the next stage; false when a sibling worker |
| 1075 |
* took the lock in the gap -- the caller must yield this |
| 1076 |
* tick. RELEASE_LOCK / delete_option of a lock we no longer |
| 1077 |
* hold is a no-op, so the caller's outer try/finally release |
| 1078 |
* is harmless even on the yield path. |
| 1079 |
*/ |
| 1080 |
private function releaseAndReacquireBetweenStages(): bool { |
| 1081 |
$this->releaseViewBuildLock(); |
| 1082 |
if (!$this->acquireViewBuildLock(0)) { |
| 1083 |
$this->logger->infoMessage( |
| 1084 |
'[staged] runStagedBuildOnce: released build lock between stages; ' |
| 1085 |
. 'another worker took it during the gap. Yielding this tick; ' |
| 1086 |
. 'the next cron / AJAX advance will resume from the persisted current_stage.' |
| 1087 |
); |
| 1088 |
return false; |
| 1089 |
} |
| 1090 |
return true; |
| 1091 |
} |
| 1092 |
|
| 1093 |
/** |
| 1094 |
* Verify `$wpdb->prefix` has not changed since S1 captured it. When the |
| 1095 |
* snapshot and the live prefix disagree, surface a deduplicated admin |
| 1096 |
* notice, log the mismatch with both prefixes for post-mortem, and |
| 1097 |
* return true so the orchestrator halts before S2-S11 run any DML |
| 1098 |
* against a different blog's tables (Codex finding #8: a mu-plugin |
| 1099 |
* calling `switch_to_blog()` between cron ticks would otherwise let |
| 1100 |
* the build write across prefixes silently). |
| 1101 |
* |
| 1102 |
* Idempotent: calling on the matching path is cheap (one option read or |
| 1103 |
* one in-memory string compare) and never mutates state. |
| 1104 |
* |
| 1105 |
* @param int $aboutToRunStage 1..11; included in the notice for context. |
| 1106 |
* @return bool True on mismatch -- caller should `return false` from |
| 1107 |
* runStagedBuildOnce immediately. False when prefix matches |
| 1108 |
* (or no capture exists) -- caller proceeds with the stage. |
| 1109 |
*/ |
| 1110 |
private function haltIfPrefixChangedSinceStageOne(int $aboutToRunStage): bool { |
| 1111 |
if ($this->verifyPrefixUnchangedSinceStageOne()) { |
| 1112 |
return false; |
| 1113 |
} |
| 1114 |
global $wpdb; |
| 1115 |
$current = (isset($wpdb->prefix) && is_string($wpdb->prefix)) ? $wpdb->prefix : ''; |
| 1116 |
$captured = $this->capturedPrefixForLog(); |
| 1117 |
$msg = sprintf( |
| 1118 |
'Multisite blog context changed during view rebuild; rebuild ' |
| 1119 |
. 'aborted to prevent cross-blog data corruption. ' |
| 1120 |
. 'captured_prefix=%s current_prefix=%s aborted_at_stage=%d', |
| 1121 |
$captured, |
| 1122 |
$current, |
| 1123 |
$aboutToRunStage |
| 1124 |
); |
| 1125 |
$this->setStagedBuildHaltNotice('multisite_prefix_changed', $msg); |
| 1126 |
$this->logger->warn('[staged] ' . $msg); |
| 1127 |
// Do NOT clear progress / captured prefix here: the original blog's |
| 1128 |
// resume on a future request will see its (untouched) progress and |
| 1129 |
// prefix capture, verify cleanly, and continue. Clearing here would |
| 1130 |
// be writing through the WRONG blog's options table anyway. |
| 1131 |
return true; |
| 1132 |
} |
| 1133 |
|
| 1134 |
/** |
| 1135 |
* Run the staged build from wherever we left off, atomically swap into |
| 1136 |
* view_done when all stages have completed. |
| 1137 |
* |
| 1138 |
* Resumable: each stage records progress in WP options so the next |
| 1139 |
* request (driven by WP-Cron or by the JS poll re-issuing the page |
| 1140 |
* request) can continue where this request left off. S2/S4/S5 are |
| 1141 |
* additionally batched within a single request and yield mid-stage |
| 1142 |
* when the per-stage budget is exhausted; the next request resumes |
| 1143 |
* from the persisted high-water id. |
| 1144 |
* |
| 1145 |
* Process-local guard prevents re-entrance within a single request. |
| 1146 |
* |
| 1147 |
* @return bool true when the build fully completed and view_done is |
| 1148 |
* now fresh; false when the request yielded mid-stage and |
| 1149 |
* another request is needed to finish. |
| 1150 |
*/ |
| 1151 |
private function runStagedBuildOnce(): bool { |
| 1152 |
if (self::$viewBuildAlreadyRanThisRequest) { |
| 1153 |
// Already either ran to completion or yielded earlier in this |
| 1154 |
// request, do not re-enter. Caller should not block on this. |
| 1155 |
return $this->viewDoneIsFresh(); |
| 1156 |
} |
| 1157 |
self::$viewBuildAlreadyRanThisRequest = true; |
| 1158 |
$this->registerViewBuildShutdownDiagnostics(); |
| 1159 |
|
| 1160 |
// Probe the PHP runtime once per request: surfaces a low-memory |
| 1161 |
// admin notice and gates the per-stage budget into a tighter |
| 1162 |
// cron-tick mode when set_time_limit() is in disable_functions. |
| 1163 |
$this->probePhpEnvironmentForBuild(); |
| 1164 |
// Probe filesystem-side host constraints (open_basedir, upload_tmp_dir, |
| 1165 |
// tmpdir disk-free) once per request. Read-and-warn-only: surfaces a |
| 1166 |
// deduplicated admin notice if anything is out of range; never blocks. |
| 1167 |
$this->probeFilesystemEnvironmentForBuild(); |
| 1168 |
|
| 1169 |
// Build is in the dedup window after a critical-stage permanent |
| 1170 |
// host failure. Re-running would just produce the same denied |
| 1171 |
// DDL again. Cron ticks during the window are no-ops; an explicit |
| 1172 |
// force rebuild clears the gate via clearStagedBuildDegradedState(). |
| 1173 |
if ($this->isBuildHaltedForHostFailure()) { |
| 1174 |
return $this->viewDoneIsFresh(); |
| 1175 |
} |
| 1176 |
|
| 1177 |
// Decide: resume or restart from scratch? |
| 1178 |
$startedAt = $this->readProgressOption('started_at', 0); |
| 1179 |
$bufferExists = $this->stagedTableExists($this->viewBuildTableName()); |
| 1180 |
$isResuming = $startedAt > 0 |
| 1181 |
&& (time() - $startedAt) < ABJ_404_Solution_ViewBuildConfig::VIEW_BUILD_RESUME_TTL_SECONDS |
| 1182 |
&& $bufferExists; |
| 1183 |
|
| 1184 |
// Single line per advance request that pins down WHICH path was |
| 1185 |
// taken and why. On a build that takes hours across many requests, |
| 1186 |
// this is the entry point for any "stuck at stage N" investigation: |
| 1187 |
// a single grep for [staged] in the debug log shows whether each |
| 1188 |
// request was resuming, restarting, or skipping due to the per-request |
| 1189 |
// guard. |
| 1190 |
$currentStage = $this->readProgressOption('current_stage', 0); |
| 1191 |
if (!$isResuming) { |
| 1192 |
$reason = ($startedAt <= 0) |
| 1193 |
? 'no prior started_at' |
| 1194 |
: (!$bufferExists |
| 1195 |
? 'buffer table missing (prior crash or fresh install)' |
| 1196 |
: ('prior build older than resume TTL (' |
| 1197 |
. (time() - $startedAt) . 's elapsed)')); |
| 1198 |
// INFO (not DEBUG) so this signal survives a site that has |
| 1199 |
// disabled DEBUG. Without it, a stuck-at-S1 redirects page on |
| 1200 |
// such a site leaves no log evidence of whether each request |
| 1201 |
// is restarting fresh or resuming. |
| 1202 |
$this->logger->infoMessage(sprintf( |
| 1203 |
'[staged] runStagedBuildOnce: fresh start (%s); current_stage=%d', |
| 1204 |
$reason, $currentStage |
| 1205 |
)); |
| 1206 |
// Scrap any partial state. An abandoned partial build older |
| 1207 |
// than the resume TTL is not safe to continue; wp_posts / |
| 1208 |
// wp_terms / wp_options state may have drifted. Also wipes |
| 1209 |
// the Phase-2 active stamp (kept outside the progress |
| 1210 |
// registry so it survives the S11 happy-path clear). |
| 1211 |
$this->performFreshStartCleanup(); |
| 1212 |
} else { |
| 1213 |
// INFO (not DEBUG): see fresh-start branch above. The pair |
| 1214 |
// (fresh start vs. resuming) is the entry point for any |
| 1215 |
// stuck-build investigation and must survive DEBUG-off sites. |
| 1216 |
$this->logger->infoMessage(sprintf( |
| 1217 |
'[staged] runStagedBuildOnce: resuming (started_at=%d, %ds ago); current_stage=%d', |
| 1218 |
$startedAt, time() - $startedAt, $currentStage |
| 1219 |
)); |
| 1220 |
// Resuming: drop only the leftover deleteme from a prior crashed |
| 1221 |
// RENAME swap. Keep the buffer + progress options intact. |
| 1222 |
$this->dropDeletemeTable(); |
| 1223 |
} |
| 1224 |
|
| 1225 |
// Set our own per-query timeout for every staged-build query |
| 1226 |
// run during this advance call. Sized below the host's session |
| 1227 |
// max_statement_time so our hint fires first, producing a |
| 1228 |
// classifiable "max_statement_time exceeded" error we can react |
| 1229 |
// to (Path B: shrink the batch). Without this, MariaDB's silent |
| 1230 |
// server-level kill produces a less-classifiable connection or |
| 1231 |
// generic-query error. |
| 1232 |
$this->stagedQueryTimeoutSeconds = (int)round($this->intelligentStagedQueryTimeoutSeconds()); |
| 1233 |
|
| 1234 |
$stage = $this->readProgressOption('current_stage', 0); |
| 1235 |
|
| 1236 |
if ($stage < 1) { |
| 1237 |
// Capture $wpdb->prefix BEFORE the S1 callback so subsequent |
| 1238 |
// stage entries can detect a mid-build switch_to_blog(). |
| 1239 |
$this->capturePrefixAtBuildStart(); |
| 1240 |
$this->stampStartedWatermarksAtS1Entry(); |
| 1241 |
// Probe sql_mode + max_allowed_packet for THIS connection. The |
| 1242 |
// probe persists in `view_build_state` and (best-effort) clears |
| 1243 |
// STRICT_TRANS_TABLES / ONLY_FULL_GROUP_BY for the build session |
| 1244 |
// so any future query the build adds inherits non-strict |
| 1245 |
// semantics. The S2 INSERT is already strict-safe via REGEXP- |
| 1246 |
// guarded CAST so this is belt-and-suspenders for new code. |
| 1247 |
$this->probeSqlModeForBuild(); |
| 1248 |
// Probe operational + DDL-safety MySQL session variables once at |
| 1249 |
// S1 entry. Read-and-warn-only: surfaces a single consolidated |
| 1250 |
// admin notice when any variable is out of range; never blocks. |
| 1251 |
$this->probeSessionVariablesAtS1Entry(); |
| 1252 |
$this->logger->debugMessage(sprintf( |
| 1253 |
'[staged] runStagedBuildOnce: capturing prefix at S1 entry: prefix=%s', |
| 1254 |
$this->capturedPrefixForLog() |
| 1255 |
)); |
| 1256 |
$this->markBuildStage('staged_build_s1_create'); |
| 1257 |
$r = $this->runTimedViewBuildStage(1, 'staged_build_s1_create', function () { |
| 1258 |
$this->stageCreateBuildTable(); |
| 1259 |
}); |
| 1260 |
if ($r === false || $r === 'halted') { |
| 1261 |
return false; // host killed S1 or halt set; next tick gated on halt window |
| 1262 |
} |
| 1263 |
// Stamp started_at on the very first stage so the resume-TTL |
| 1264 |
// clock starts from buffer creation. |
| 1265 |
if ($this->readProgressOption('started_at', 0) === 0) { |
| 1266 |
$this->writeProgressOption('started_at', time()); |
| 1267 |
} |
| 1268 |
$this->writeProgressOption('current_stage', 1); |
| 1269 |
$stage = 1; |
| 1270 |
} |
| 1271 |
|
| 1272 |
if ($stage < 2) { |
| 1273 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1274 |
if ($this->haltIfPrefixChangedSinceStageOne(2)) { return false; } |
| 1275 |
if ($this->gateAbortIfMutationWatermarkAdvanced(2)) { return false; } |
| 1276 |
$r = $this->runTimedViewBuildStage(2, 'staged_build_s2_insert', function () { |
| 1277 |
return $this->stageInsertRedirectsBatched(); |
| 1278 |
}); |
| 1279 |
if ($r === false || $r === 'halted') { |
| 1280 |
return false; // budget exhausted, kill, or halt; resume / no-op next request |
| 1281 |
} |
| 1282 |
$this->writeProgressOption('current_stage', 2); |
| 1283 |
$stage = 2; |
| 1284 |
} |
| 1285 |
|
| 1286 |
if ($stage < 3) { |
| 1287 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1288 |
if ($this->haltIfPrefixChangedSinceStageOne(3)) { return false; } |
| 1289 |
if ($this->gateAbortIfMutationWatermarkAdvanced(3)) { return false; } |
| 1290 |
if ($this->isStageMarkedSkipped(3)) { |
| 1291 |
// Permanent host-side denial recorded on a prior tick. |
| 1292 |
// Advance current_stage past S3 without touching the SQL. |
| 1293 |
$this->writeProgressOption('current_stage', 3); |
| 1294 |
$stage = 3; |
| 1295 |
} else { |
| 1296 |
$this->markBuildStage('staged_build_s3_index_fd'); |
| 1297 |
// Non-batched: kill-streak escape valve extends the per-query |
| 1298 |
// timeout above the host's session limit on retry. Without |
| 1299 |
// this, a CREATE INDEX that exceeds max_statement_time on |
| 1300 |
// big buffers loops with the same timeout forever. |
| 1301 |
$r = $this->runNonBatchedStageWithKillStreakEscape( |
| 1302 |
3, 'staged_build_s3_index_fd', 's3_kill_streak', |
| 1303 |
function () { $this->stageAddPreJoinIndexes(); } |
| 1304 |
); |
| 1305 |
if ($r === false || $r === 'halted') { |
| 1306 |
return false; |
| 1307 |
} |
| 1308 |
$this->writeProgressOption('current_stage', 3); |
| 1309 |
$stage = 3; |
| 1310 |
} |
| 1311 |
} |
| 1312 |
|
| 1313 |
if ($stage < 4) { |
| 1314 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1315 |
if ($this->haltIfPrefixChangedSinceStageOne(4)) { return false; } |
| 1316 |
if ($this->gateAbortIfMutationWatermarkAdvanced(4)) { return false; } |
| 1317 |
$r = $this->runTimedViewBuildStage(4, 'staged_build_s4_update_posts', function () { |
| 1318 |
return $this->stageUpdatePostsBatched(); |
| 1319 |
}); |
| 1320 |
if ($r === false || $r === 'halted') { |
| 1321 |
return false; |
| 1322 |
} |
| 1323 |
$this->writeProgressOption('current_stage', 4); |
| 1324 |
$stage = 4; |
| 1325 |
} |
| 1326 |
|
| 1327 |
if ($stage < 5) { |
| 1328 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1329 |
if ($this->haltIfPrefixChangedSinceStageOne(5)) { return false; } |
| 1330 |
if ($this->gateAbortIfMutationWatermarkAdvanced(5)) { return false; } |
| 1331 |
$r = $this->runTimedViewBuildStage(5, 'staged_build_s5_update_terms', function () { |
| 1332 |
return $this->stageUpdateTermsBatched(); |
| 1333 |
}); |
| 1334 |
if ($r === false || $r === 'halted') { |
| 1335 |
return false; |
| 1336 |
} |
| 1337 |
$this->writeProgressOption('current_stage', 5); |
| 1338 |
$stage = 5; |
| 1339 |
} |
| 1340 |
|
| 1341 |
if ($stage < 6) { |
| 1342 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1343 |
if ($this->haltIfPrefixChangedSinceStageOne(6)) { return false; } |
| 1344 |
if ($this->gateAbortIfMutationWatermarkAdvanced(6)) { return false; } |
| 1345 |
$this->markBuildStage('staged_build_s6_update_home'); |
| 1346 |
$r = $this->runTimedViewBuildStage(6, 'staged_build_s6_update_home', function () { |
| 1347 |
$this->stageUpdateHome(); |
| 1348 |
}); |
| 1349 |
if ($r === false || $r === 'halted') { |
| 1350 |
return false; |
| 1351 |
} |
| 1352 |
$this->writeProgressOption('current_stage', 6); |
| 1353 |
$stage = 6; |
| 1354 |
} |
| 1355 |
|
| 1356 |
if ($stage < 7) { |
| 1357 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1358 |
if ($this->haltIfPrefixChangedSinceStageOne(7)) { return false; } |
| 1359 |
if ($this->gateAbortIfMutationWatermarkAdvanced(7)) { return false; } |
| 1360 |
$this->markBuildStage('staged_build_s7_update_external'); |
| 1361 |
$r = $this->runTimedViewBuildStage(7, 'staged_build_s7_update_external', function () { |
| 1362 |
$this->stageUpdateExternal(); |
| 1363 |
}); |
| 1364 |
if ($r === false || $r === 'halted') { |
| 1365 |
return false; |
| 1366 |
} |
| 1367 |
$this->writeProgressOption('current_stage', 7); |
| 1368 |
$stage = 7; |
| 1369 |
} |
| 1370 |
|
| 1371 |
if ($stage < 8) { |
| 1372 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1373 |
if ($this->haltIfPrefixChangedSinceStageOne(8)) { return false; } |
| 1374 |
if ($this->gateAbortIfMutationWatermarkAdvanced(8)) { return false; } |
| 1375 |
$this->markBuildStage('staged_build_s8_update_special'); |
| 1376 |
$r = $this->runTimedViewBuildStage(8, 'staged_build_s8_update_special', function () { |
| 1377 |
$this->stageUpdateSpecial(); |
| 1378 |
}); |
| 1379 |
if ($r === false || $r === 'halted') { |
| 1380 |
return false; |
| 1381 |
} |
| 1382 |
$this->writeProgressOption('current_stage', 8); |
| 1383 |
$stage = 8; |
| 1384 |
} |
| 1385 |
|
| 1386 |
if ($stage < 9) { |
| 1387 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1388 |
if ($this->haltIfPrefixChangedSinceStageOne(9)) { return false; } |
| 1389 |
if ($this->gateAbortIfMutationWatermarkAdvanced(9)) { return false; } |
| 1390 |
if ($this->isStageMarkedSkipped(9)) { |
| 1391 |
$this->writeProgressOption('current_stage', 9); |
| 1392 |
$stage = 9; |
| 1393 |
} else { |
| 1394 |
// Non-batched: temp-table aggregate over wp_abj404_logs_hits + |
| 1395 |
// UPDATE JOIN against the buffer. Kill-streak escape valve |
| 1396 |
// extends the per-query timeout on retry so a logs_hits scan |
| 1397 |
// that doesn't fit in the host's max_statement_time can |
| 1398 |
// eventually finish. |
| 1399 |
$s9Result = $this->runNonBatchedStageWithKillStreakEscape( |
| 1400 |
9, 'staged_build_s9_update_hits', 's9_kill_streak', |
| 1401 |
function () { |
| 1402 |
if ($this->logsHitsTableExists()) { |
| 1403 |
$this->markBuildStage('staged_build_s9_update_hits'); |
| 1404 |
$this->stageUpdateHits(); |
| 1405 |
return null; |
| 1406 |
} |
| 1407 |
$this->markBuildStage('staged_build_s9_update_hits', 'skipped; logs hits table unavailable'); |
| 1408 |
return 'skipped'; |
| 1409 |
} |
| 1410 |
); |
| 1411 |
if ($s9Result === false || $s9Result === 'halted') { |
| 1412 |
return false; |
| 1413 |
} |
| 1414 |
// Skipped or not, advance past S9. |
| 1415 |
$this->writeProgressOption('current_stage', 9); |
| 1416 |
$stage = 9; |
| 1417 |
} |
| 1418 |
} |
| 1419 |
|
| 1420 |
if ($stage < 10) { |
| 1421 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1422 |
if ($this->haltIfPrefixChangedSinceStageOne(10)) { return false; } |
| 1423 |
if ($this->gateAbortIfMutationWatermarkAdvanced(10)) { return false; } |
| 1424 |
if ($this->isStageMarkedSkipped(10)) { |
| 1425 |
$this->writeProgressOption('current_stage', 10); |
| 1426 |
$stage = 10; |
| 1427 |
} else { |
| 1428 |
$this->markBuildStage('staged_build_s10_index_sort'); |
| 1429 |
// Non-batched: same kill-streak escape valve as S3. A |
| 1430 |
// CREATE INDEX that exceeds the host's max_statement_time on |
| 1431 |
// big buffers needs an extended retry timeout to complete. |
| 1432 |
$r = $this->runNonBatchedStageWithKillStreakEscape( |
| 1433 |
10, 'staged_build_s10_index_sort', 's10_kill_streak', |
| 1434 |
function () { $this->stageAddSortIndexes(); } |
| 1435 |
); |
| 1436 |
if ($r === false || $r === 'halted') { |
| 1437 |
return false; |
| 1438 |
} |
| 1439 |
$this->writeProgressOption('current_stage', 10); |
| 1440 |
$stage = 10; |
| 1441 |
} |
| 1442 |
} |
| 1443 |
|
| 1444 |
if ($stage < 11) { |
| 1445 |
if (!$this->releaseAndReacquireBetweenStages()) { return false; } |
| 1446 |
if ($this->haltIfPrefixChangedSinceStageOne(11)) { return false; } |
| 1447 |
if ($this->gateAbortIfMutationWatermarkAdvanced(11)) { return false; } |
| 1448 |
$this->markBuildStage('staged_build_s11_swap'); |
| 1449 |
if (!$this->runS11SwapWithPreRenameWatermarkRecheck()) { return false; } |
| 1450 |
$this->publishBuiltWatermarkFromActiveBuildStartedWatermark(); |
| 1451 |
$this->markViewDoneBuildCompleted(); |
| 1452 |
$this->clearAllProgressOptions(); |
| 1453 |
} |
| 1454 |
|
| 1455 |
return true; |
| 1456 |
} |
| 1457 |
|
| 1458 |
|
| 1459 |
// The following helpers all live on sibling traits so this file stays |
| 1460 |
// focused on the orchestrator. They're listed here as a navigation aid: |
| 1461 |
// |
| 1462 |
// ABJ_404_Solution_DataAccess_ViewBuildHelpersTrait: |
| 1463 |
// - runStagedSqlFile / runStagedSqlFileTolerantOfDuplicateKey |
| 1464 |
// - describeStagedSqlFailure / describeBuildProgressForNotice |
| 1465 |
// - stagedQueryOptions |
| 1466 |
// - viewDoneTableExists / stagedTableExists |
| 1467 |
// - viewDoneIsFresh |
| 1468 |
// - acquireViewBuildLock / releaseViewBuildLock |
| 1469 |
// - scheduleViewDoneRebuild |
| 1470 |
// |
| 1471 |
// ABJ_404_Solution_DataAccess_ViewBuildStageCallbacksTrait: |
| 1472 |
// - dropTransientStagedTables / dropDeletemeTable |
| 1473 |
// - stageCreateBuildTable (S1) |
| 1474 |
// - stageInsertRedirectsBatched (S2) / runInsertBatch |
| 1475 |
// - stageAddPreJoinIndexes (S3) |
| 1476 |
// - stageUpdatePostsBatched (S4) / stageUpdateTermsBatched (S5) / runIdRangeBatchedUpdate |
| 1477 |
// - stageUpdateHome (S6) / stageUpdateExternal (S7) / stageUpdateSpecial (S8) |
| 1478 |
// - stageUpdateHits (S9) |
| 1479 |
// - stageAddSortIndexes (S10) |
| 1480 |
// - stageRenameSwap (S11) |
| 1481 |
// - countLiveRedirects / countViewBuildRows / maxBuildBufferId / humanBatchProgress |
| 1482 |
} |
| 1483 |
|
