Speed up `db.<collection>.any()` for the RocksDB storage engine (#5934)

CHANGELOG

@@ -1,6 +1,21 @@
 v3.3.13 (XXXX-XX-XX)
 --------------------
 
+* put an upper bound on the number of documents to be scanned when using
+  `db.<collection>.any()` in the RocksDB storage engine
+
+  previous versions of ArangoDB did a scan of a random amount of documents in
+  the collection, up to the total number of documents available. this produced
+  a random selection with a good quality, but needed to scan half the number
+  of documents in the collection on average.
+
+  The new version will only scan up to 500 documents, so it produces a less
+  random result, but will be a lot faster especially for large collections.
+  
+  The implementation of `any()` for the MMFiles engine remains unchanged. The
+  MMFiles engine will pick a random document from the entire range of the 
+  in-memory primary index without performing scans.
+
 * return an empty result set instead of an "out of memory" exception when
   querying the geo index with invalid (out of range) coordinates
 

arangod/RocksDBEngine/RocksDBIterators.cpp

@@ -193,26 +193,52 @@ RocksDBAnyIndexIterator::RocksDBAnyIndexIterator(
   TRI_ASSERT(_iterator);
 
   _total = col->numberDocuments(trx);
-  uint64_t off = RandomGenerator::interval(_total - 1);
+  _forward = RandomGenerator::interval(uint16_t(1)) ? true : false;
+
+  //initial seek
   if (_total > 0) {
-    if (off <= _total / 2) {
+    uint64_t steps = RandomGenerator::interval(_total - 1) % 500;
-      _iterator->Seek(_bounds.start());
+    auto initialKey = RocksDBKey();
-      while (_iterator->Valid() && off-- > 0) {
+    initialKey.constructDocument(
-        _iterator->Next();
+      static_cast<RocksDBCollection*>(col->getPhysical())->objectId(),
+      LocalDocumentId(RandomGenerator::interval(UINT64_MAX))
+    );
+    _iterator->Seek(initialKey.string());
+
+    if (checkIter()) {
+      if (_forward) {
+        while (steps-- > 0) {
+          _iterator->Next();
+          if(!checkIter()) { break; }
+        }
+      } else {
+        while (steps-- > 0) {
+          _iterator->Prev();
+          if(!checkIter()) { break; }
+        }
       }
-    } else {
-      off = _total - (off + 1);
-      _iterator->SeekForPrev(_bounds.end());
-      while (_iterator->Valid() && off-- > 0) {
-        _iterator->Prev();
-      }
-    }
-    if (!_iterator->Valid() || outOfRange()) {
-      _iterator->Seek(_bounds.start());
     }
   }
 }
 
+bool RocksDBAnyIndexIterator::checkIter(){
+  if ( /* not  valid */            !_iterator->Valid() ||
+       /* out of range forward */  ( _forward && _cmp->Compare(_iterator->key(), _bounds.end())   > 0) ||
+       /* out of range backward */ (!_forward && _cmp->Compare(_iterator->key(), _bounds.start()) < 0)  ) {
+
+    if (_forward) {
+      _iterator->Seek(_bounds.start());
+    } else {
+      _iterator->SeekForPrev(_bounds.end());
+    }
+
+    if (!_iterator->Valid()) {
+      return false;
+    }
+  }
+  return true;
+}
+
 bool RocksDBAnyIndexIterator::next(LocalDocumentIdCallback const& cb, size_t limit) {
   TRI_ASSERT(_trx->state()->isRunning());
 

arangod/RocksDBEngine/RocksDBIterators.h

@@ -87,14 +87,14 @@ class RocksDBAnyIndexIterator final : public IndexIterator {
 
  private:
   bool outOfRange() const;
-  static uint64_t newOffset(LogicalCollection* collection,
+  bool checkIter();
-                            transaction::Methods* trx);
 
   rocksdb::Comparator const* _cmp;
   std::unique_ptr<rocksdb::Iterator> _iterator;
   RocksDBKeyBounds const _bounds;
   uint64_t _total;
   uint64_t _returned;
+  bool _forward;
 };
 
 /// @brief iterates over the primary index and does lookups