add distributed hosts crawling using yggo nodes manifest

README.md

@@ -150,7 +150,7 @@ GET m=SphinxQL
 * [ ] Implement database auto backup on crawl process completing
 * [x] Add transactions to prevent data loss on DB crashes
 * [x] JSON API
-* [ ] Distributed index data sharing between the nodes trough service API
+* [x] Distributed index data sharing between the nodes trough service API
 * [x] An idea to make unique gravatars for sites without favicons, because simpler to ident, comparing to ipv6
 * [ ] An idea to make some visitors counters, like in good old times?
 

config/app.php.txt

@@ -121,8 +121,10 @@ define('CRAWL_STOP_DISK_QUOTA_MB_LEFT', 500);
  * Usually up to 20 pages per minute,
  * to prevent websites overload by sending GET crawling requests
  *
+ * Set 0 to disable
+ *
  */
-define('CRAWL_PAGE_LIMIT', 10);
+define('CRAWL_PAGE_LIMIT', 20);
 
 /*
  * Images (URI) processing limit in the crawler.php queue
@@ -133,8 +135,27 @@ define('CRAWL_PAGE_LIMIT', 10);
  * Usually up to 20 pages per minute,
  * to prevent websites overload by sending GET crawling requests
  *
+ * Set 0 to disable
+ *
+ */
+define('CRAWL_IMAGE_LIMIT', 10);
+
+/*
+ * Manifest (URI) processing limit in the crawler.php queue
+ *
+ * Used to collect distributed data index
+ * that match CRAWL_URL_REGEXP & CRAWL_MANIFEST_API_VERSION
+ *
+ * This option related to CRAWL_MANIFEST_SECONDS_OFFSET value
+ * and the crontab task frequency (https://github.com/YGGverse/YGGo#crontab)
+ *
+ * Usually up to 20 pages per minute,
+ * to prevent websites overload by sending GET crawling requests
+ *
+ * Set 0 to disable
+ *
  */
-define('CRAWL_IMAGE_LIMIT', 20);
+define('CRAWL_MANIFEST_LIMIT', 10);
 
 /*
  * Renew page index by timing offset provided
@@ -162,6 +183,19 @@ define('CRAWL_PAGE_SECONDS_OFFSET', 60*60*24*30*12);
  */
 define('CRAWL_IMAGE_SECONDS_OFFSET', 60*60*24*30*12);
 
+/*
+ * Renew manifests index by timing offset provided
+ *
+ * This option works with CRAWL_MANIFEST_LIMIT step queue
+ *
+ * Pay attention, that CRAWL_MANIFEST_LIMIT + CRAWL_MANIFEST_SECONDS_OFFSET pair
+ * must have enough value to crawl all manifests collected in the DB index
+ *
+ * or the crawler can stuck in queue
+ *
+ */
+define('CRAWL_MANIFEST_SECONDS_OFFSET', 60*60*24*30);
+
 /*
  * Only URL addresses match this rule will be auto-crawled
  *
@@ -242,6 +276,12 @@ define('CRAWL_ROBOTS_POSTFIX_RULES', null); // string|null
  */
 define('CRAWL_MANIFEST', true);
 
+/*
+ * Manifest API version compatibility
+ *
+ */
+define('CRAWL_MANIFEST_API_VERSION', 0.4);
+
 /*
  * Set default auto-crawl status for new manifest added
  *

crontab/cleaner.php

@@ -22,11 +22,18 @@ $db = new MySQL(DB_HOST, DB_PORT, DB_NAME, DB_USERNAME, DB_PASSWORD);
 $timeStart = microtime(true);
 
 $hostsTotal         = $db->getTotalHosts();
+$manifestsTotal     = $db->getTotalManifests();
 $hostsUpdated       = 0;
 $hostsPagesDeleted  = 0;
 $hostsImagesDeleted = 0;
+$manifestsDeleted   = 0;
 
-// Get host queue
+// Begin update
+$db->beginTransaction();
+
+try {
+
+  // Get cleaner queue
   foreach ($db->getCleanerQueue(CLEAN_HOST_LIMIT, time() - CLEAN_HOST_SECONDS_OFFSET) as $host) {
 
     // Parse host info
@@ -41,11 +48,6 @@ foreach ($db->getCleanerQueue(CLEAN_HOST_LIMIT, time() - CLEAN_HOST_SECONDS_OFFS
       $hostRobots = null;
     }
 
-  // Begin update
-  $db->beginTransaction();
-
-  try {
-
     // Update host data
     $hostsUpdated += $db->updateHostRobots($host->hostId, $hostRobots, time());
 
@@ -118,6 +120,58 @@ foreach ($db->getCleanerQueue(CLEAN_HOST_LIMIT, time() - CLEAN_HOST_SECONDS_OFFS
       // Delete host image
       $hostsImagesDeleted += $db->deleteHostImage($hostImage->hostImageId);
     }
+  }
+
+  // Clean up deprecated manifests
+  foreach ($db->getManifests() as $manifest) {
+
+    $delete = false;
+
+    $curl = new Curl($manifest->url);
+
+    // Skip processing non 200 code
+    if (200 != $curl->getCode()) {
+
+      continue; // Wait for reconnect
+    }
+
+    // Skip processing without returned data
+    if (!$remoteManifest = $curl->getContent()) {
+
+      $delete = true;
+    }
+
+    // Skip processing on json encoding error
+    if (!$remoteManifest = @json_decode($remoteManifest)) {
+
+      $delete = true;
+    }
+
+    // Skip processing on required fields missed
+    if (empty($remoteManifest->status) ||
+        empty($remoteManifest->result->config->crawlUrlRegexp) ||
+        empty($remoteManifest->result->api->version)) {
+
+      $delete = true;
+    }
+
+    // Skip processing on API version not compatible
+    if ($remoteManifest->result->api->version !== CRAWL_MANIFEST_API_VERSION) {
+
+      $delete = true;
+    }
+
+    // Skip processing on crawlUrlRegexp does not match CRAWL_URL_REGEXP condition
+    if ($remoteManifest->result->config->crawlUrlRegexp !== CRAWL_URL_REGEXP) {
+
+      $delete = true;
+    }
+
+    if ($delete) {
+
+      $manifestsDeleted += $db->deleteManifest($manifest->manifestId);
+    }
+  }
 
   $db->commit();
 
@@ -127,11 +181,12 @@ foreach ($db->getCleanerQueue(CLEAN_HOST_LIMIT, time() - CLEAN_HOST_SECONDS_OFFS
 
   $db->rollBack();
 }
-}
 
 // Debug
 echo 'Hosts total: ' . $hostsTotal . PHP_EOL;
 echo 'Hosts updated: ' . $hostsUpdated . PHP_EOL;
 echo 'Hosts pages deleted: ' . $hostsPagesDeleted . PHP_EOL;
 echo 'Hosts images deleted: ' . $hostsImagesDeleted . PHP_EOL;
-echo 'Execution time: ' . microtime(true) - $timeStart . PHP_EOL;
+echo 'Manifests total: ' . $manifestsTotal . PHP_EOL;
+echo 'Manifests deleted: ' . $manifestsDeleted . PHP_EOL;
+echo 'Execution time: ' . microtime(true) - $timeStart . PHP_EOL . PHP_EOL;

crontab/crawler.php

@@ -29,8 +29,10 @@ $timeStart = microtime(true);
 
 $hostPagesProcessed  = 0;
 $hostImagesProcessed = 0;
+$manifestsProcessed  = 0;
 $hostPagesIndexed    = 0;
 $hostImagesIndexed   = 0;
+$manifestsIndexed    = 0;
 $hostPagesAdded      = 0;
 $hostImagesAdded     = 0;
 $hostsAdded          = 0;
@@ -38,6 +40,185 @@ $hostsAdded          = 0;
 // Connect database
 $db = new MySQL(DB_HOST, DB_PORT, DB_NAME, DB_USERNAME, DB_PASSWORD);
 
+$db->beginTransaction();
+
+try {
+
+  // Process manifests crawl queue
+  foreach ($db->getManifestCrawlQueue(CRAWL_MANIFEST_LIMIT, time() - CRAWL_MANIFEST_SECONDS_OFFSET) as $queueManifest) {
+
+    $curl = new Curl($queueManifest->url);
+
+    // Update manifest index anyway, with the current time and http code
+    $manifestsProcessed += $db->updateManifestCrawlQueue($queueManifest->manifestId, time(), $curl->getCode());
+
+    // Skip processing non 200 code
+    if (200 != $curl->getCode()) {
+
+      continue;
+    }
+
+    // Skip processing without returned data
+    if (!$remoteManifest = $curl->getContent()) {
+
+      continue;
+    }
+
+    // Skip processing on json encoding error
+    if (!$remoteManifest = @json_decode($remoteManifest)) {
+
+      continue;
+    }
+
+    // Skip processing on required fields missed
+    if (empty($remoteManifest->status) ||
+        empty($remoteManifest->result->config->crawlUrlRegexp) ||
+        empty($remoteManifest->result->api->version) ||
+        empty($remoteManifest->result->api->hosts)) {
+
+      continue;
+    }
+
+    // Skip processing on API version not compatible
+    if ($remoteManifest->result->api->version !== CRAWL_MANIFEST_API_VERSION) {
+
+      continue;
+    }
+
+    // Skip processing on host API not available
+    if (!$remoteManifest->result->api->hosts) {
+
+      continue;
+    }
+
+    // Skip processing on crawlUrlRegexp does not match CRAWL_URL_REGEXP condition
+    if ($remoteManifest->result->config->crawlUrlRegexp !== CRAWL_URL_REGEXP) {
+
+      continue;
+    }
+
+    // Skip processing on host link does not match condition
+    if (false === preg_match(CRAWL_URL_REGEXP, $remoteManifest->result->api->hosts)) {
+
+      continue;
+    }
+
+    // Begin hosts collection
+    $curl = new Curl($remoteManifest->result->api->hosts);
+
+    // Skip processing non 200 code
+    if (200 != $curl->getCode()) {
+
+      continue;
+    }
+
+    // Skip processing without returned data
+    if (!$remoteManifestHosts = $curl->getContent()) {
+
+      continue;
+    }
+
+    // Skip processing on json encoding error
+    if (!$remoteManifestHosts = @json_decode($remoteManifestHosts)) {
+
+      continue;
+    }
+
+    // Skip processing on required fields missed
+    if (empty($remoteManifestHosts->status) ||
+        empty($remoteManifestHosts->result)) {
+
+      continue;
+    }
+
+    // Begin hosts processing
+    foreach ($remoteManifestHosts->result as $remoteManifestHost) {
+
+      // Skip processing on required fields missed
+      if (empty($remoteManifestHost->scheme) ||
+          empty($remoteManifestHost->name)) {
+
+        continue;
+      }
+
+      $hostURL = $remoteManifestHost->scheme . '://' .
+                 $remoteManifestHost->name .
+                (!empty($remoteManifestHost->port) ? ':' . $remoteManifestHost->port : false);
+
+      // Validate formatted link
+      if (filter_var($hostURL, FILTER_VALIDATE_URL) && preg_match(CRAWL_URL_REGEXP, $hostURL)) {
+
+        // Host exists
+        if ($host = $db->getHost(crc32($hostURL))) {
+
+          $hostStatus        = $host->status;
+          $hostPageLimit     = $host->crawlPageLimit;
+          $hostImageLimit    = $host->crawlImageLimit;
+          $hostId            = $host->hostId;
+          $hostRobots        = $host->robots;
+          $hostRobotsPostfix = $host->robotsPostfix;
+
+        // Register new host
+        } else {
+
+          // Get robots.txt if exists
+          $curl = new Curl($hostURL . '/robots.txt', CRAWL_CURLOPT_USERAGENT);
+
+          if (200 == $curl->getCode() && false !== stripos($curl->getContent(), 'user-agent:')) {
+            $hostRobots = $curl->getContent();
+          } else {
+            $hostRobots = CRAWL_ROBOTS_DEFAULT_RULES;
+          }
+
+          $hostRobotsPostfix = CRAWL_ROBOTS_POSTFIX_RULES;
+
+          $hostStatus    = CRAWL_HOST_DEFAULT_STATUS;
+          $hostPageLimit = CRAWL_HOST_DEFAULT_PAGES_LIMIT;
+          $hostImageLimit= CRAWL_HOST_DEFAULT_IMAGES_LIMIT;
+
+          $hostId        = $db->addHost($remoteManifestHosts->result->scheme,
+                                        $remoteManifestHosts->result->name,
+                                        $remoteManifestHosts->result->port,
+                                        crc32($hostURL),
+                                        time(),
+                                        null,
+                                        $hostPageLimit,
+                                        $hostImageLimit,
+                                        (string) CRAWL_HOST_DEFAULT_META_ONLY,
+                                        (string) $hostStatus,
+                                        $hostRobots,
+                                        $hostRobotsPostfix);
+
+          if ($hostId) {
+
+            $hostsAdded++;
+
+          } else {
+
+            continue;
+          }
+        }
+
+        // Init robots parser
+        $robots = new Robots(($hostRobots ? (string) $hostRobots : (string) CRAWL_ROBOTS_DEFAULT_RULES) . PHP_EOL . ($hostRobotsPostfix ? (string) $hostRobotsPostfix : (string) CRAWL_ROBOTS_POSTFIX_RULES));
+
+        // Save home page info
+        // Until page API not implemented, save at least home page to have ability to crawl
+        // @TODO
+        if ($hostStatus && // host enabled
+            $robots->uriAllowed('/') && // page allowed by robots.txt rules
+            $hostPageLimit > $db->getTotalHostPages($hostId) && // pages quantity not reached host limit
+            !$db->getHostPage($hostId, crc32('/'))) {  // page not exists
+
+            if ($db->addHostPage($hostId, crc32('/'), '/', time())) {
+
+              $hostPagesAdded++;
+            }
+        }
+      }
+    }
+  }
+
   // Process images crawl queue
   foreach ($db->getHostImageCrawlQueue(CRAWL_IMAGE_LIMIT, time() - CRAWL_IMAGE_SECONDS_OFFSET) as $queueHostImage) {
 
@@ -204,10 +385,6 @@ foreach ($db->getHostPageCrawlQueue(CRAWL_PAGE_LIMIT, time() - CRAWL_PAGE_SECOND
         // Validate formatted src link
         if (filter_var($imageSrc, FILTER_VALIDATE_URL) && preg_match(CRAWL_URL_REGEXP, $imageSrc)) {
 
-        $db->beginTransaction();
-
-        try {
-
           // Parse formatted src link
           $hostImageURL = Parser::hostURL($imageSrc);
           $hostImageURI = Parser::uri($imageSrc);
@@ -310,15 +487,6 @@ foreach ($db->getHostPageCrawlQueue(CRAWL_PAGE_LIMIT, time() - CRAWL_PAGE_SECOND
 
               $db->updateHostImageRank($hostId, crc32($hostImageURI->string), 1);
           }
-
-          $db->commit();
-
-        } catch(Exception $e) {
-
-          var_dump($e);
-
-          $db->rollBack();
-        }
         }
       }
     }
@@ -370,10 +538,6 @@ foreach ($db->getHostPageCrawlQueue(CRAWL_PAGE_LIMIT, time() - CRAWL_PAGE_SECOND
       // Validate formatted link
       if (filter_var($href, FILTER_VALIDATE_URL) && preg_match(CRAWL_URL_REGEXP, $href)) {
 
-      $db->beginTransaction();
-
-      try {
-
         // Parse formatted link
         $hostURL     = Parser::hostURL($href);
         $hostPageURI = Parser::uri($href);
@@ -454,6 +618,9 @@ foreach ($db->getHostPageCrawlQueue(CRAWL_PAGE_LIMIT, time() - CRAWL_PAGE_SECOND
 
             $db->updateHostPageRank($hostId, crc32($hostPageURI->string), 1);
         }
+      }
+    }
+  }
 
   $db->commit();
 
@@ -463,9 +630,6 @@ foreach ($db->getHostPageCrawlQueue(CRAWL_PAGE_LIMIT, time() - CRAWL_PAGE_SECOND
 
   $db->rollBack();
 }
-    }
-  }
-}
 
 // Debug
 echo 'Pages processed: ' . $hostPagesProcessed . PHP_EOL;
@@ -474,5 +638,7 @@ echo 'Pages added: ' . $hostPagesAdded . PHP_EOL;
 echo 'Images processed: ' . $hostImagesProcessed . PHP_EOL;
 echo 'Images indexed: ' . $hostImagesIndexed . PHP_EOL;
 echo 'Images added: ' . $hostImagesAdded . PHP_EOL;
+echo 'Manifests processed: ' . $manifestsProcessed . PHP_EOL;
+echo 'Manifests indexed: ' . $manifestsIndexed . PHP_EOL;
 echo 'Hosts added: ' . $hostsAdded . PHP_EOL;
-echo 'Total time: ' . microtime(true) - $timeStart . PHP_EOL;
+echo 'Total time: ' . microtime(true) - $timeStart . PHP_EOL . PHP_EOL;

library/mysql.php

@@ -29,6 +29,24 @@ class MySQL {
   }
 
   // Manifest
+  public function getTotalManifests() {
+
+    $query = $this->_db->prepare('SELECT COUNT(*) AS `total` FROM `manifest`');
+
+    $query->execute();
+
+    return $query->fetch()->total;
+  }
+
+  public function getManifests() {
+
+    $query = $this->_db->prepare('SELECT * FROM `manifest`');
+
+    $query->execute();
+
+    return $query->fetchAll();
+  }
+
   public function getManifest(int $crc32url) {
 
     $query = $this->_db->prepare('SELECT * FROM `manifest` WHERE `crc32url` = ? LIMIT 1');
@@ -47,6 +65,15 @@ class MySQL {
     return $this->_db->lastInsertId();
   }
 
+  public function deleteManifest(int $manifestId) {
+
+    $query = $this->_db->prepare('DELETE FROM `manifest` WHERE `manifestId` = ? LIMIT 1');
+
+    $query->execute([$manifestId]);
+
+    return $query->rowCount();
+  }
+
   // Host
   public function getAPIHosts(string $apiHostFields) {
 
@@ -570,4 +597,28 @@ class MySQL {
 
     return $query->rowCount();
   }
+
+  public function getManifestCrawlQueue(int $limit, int $timeFrom) {
+
+    $query = $this->_db->prepare('SELECT * FROM `manifest`
+
+                                           WHERE (`timeUpdated` IS NULL OR `timeUpdated` < ? ) AND `status` <> 0
+
+                                           ORDER BY RAND()
+
+                                           LIMIT ' . (int) $limit);
+
+    $query->execute([$timeFrom]);
+
+    return $query->fetchAll();
+  }
+
+  public function updateManifestCrawlQueue(int $manifestId, int $timeUpdated, int $httpCode) {
+
+    $query = $this->_db->prepare('UPDATE `manifest` SET `timeUpdated` = ?, `httpCode` = ? WHERE `manifestId` = ? LIMIT 1');
+
+    $query->execute([$timeUpdated, $httpCode, $manifestId]);
+
+    return $query->rowCount();
+  }
 }