Speed up KEYS for common case of prefix matching by implementing in terms of SCAN

CHANGELOG.md

@@ -9,6 +9,9 @@ causing all writes to stall). From now on, similar kind of corruption should onl
 take out a single node, and not spread to the entire cluster.
 - Add command ``RAFT-JOURNAL-SCAN`` to make searching through the contents of the
 raft journal easier.
+- ``KEYS`` is now implemented in terms of ``SCAN``, making prefix scanning of the
+keyspace just as efficient as with ``SCAN``. (Note: The use of ``KEYS`` is still
+generally discouraged due to potentially huge response size, don't use in production)
 
 ## 0.3.8 (2019-05-27)
 - Prevent elections from hanging on the TCP timeout when one of the member hosts

src/StateMachine.cc

@@ -1700,24 +1700,24 @@ rocksdb::Status StateMachine::exists(StagingArea &stagingArea, const ReqIterator
 rocksdb::Status StateMachine::keys(StagingArea &stagingArea, std::string_view pattern, std::vector<std::string> &result) {
   result.clear();
 
-  bool allkeys = (pattern.length() == 1 && pattern[0] == '*');
-  IteratorPtr iter(stagingArea.getIterator());
+  std::string oldCursor;
+  std::string newCursor;
 
-  std::string searchPrefix(1, char(InternalKeyType::kDescriptor));
-  for(iter->Seek(searchPrefix); iter->Valid(); iter->Next()) {
-    std::string rkey = iter->key().ToString();
-    if(rkey.size() == 0 || rkey[0] != char(InternalKeyType::kDescriptor)) break;
+  while(true) {
+    rocksdb::Status st = scan(stagingArea, oldCursor, pattern, std::numeric_limits<size_t>::max()-1, newCursor, result);
+    if(!st.ok()) return st;
 
-    if(allkeys || stringmatchlen(pattern.data(), pattern.length(), rkey.data()+1, rkey.length()-1, 0)) {
-      result.push_back(rkey.substr(1));
+    if(newCursor.empty()) {
+      break;
     }
+
+    oldCursor = newCursor;
   }
 
   return rocksdb::Status::OK();
 }
 
 rocksdb::Status StateMachine::scan(StagingArea &stagingArea, std::string_view cursor, std::string_view pattern, size_t count, std::string &newcursor, std::vector<std::string> &results) {
-  results.clear();
 
   // Any hits *must* start with patternPrefix. This will allow us in many
   // circumstances to eliminate checking large parts of the keyspace, without

test/state-machine.cc

@@ -654,30 +654,37 @@ TEST_F(State_Machine, scan) {
   ASSERT_EQ(keys, make_vec("key1", "key2"));
   ASSERT_EQ(newcursor, "key3");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "key*", 2, newcursor, keys));
   ASSERT_EQ(keys, make_vec("key3", "key4"));
   ASSERT_EQ(newcursor, "key5");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "key*", 2, newcursor, keys));
   ASSERT_EQ(keys, make_vec("key5", "key6"));
   ASSERT_EQ(newcursor, "");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan("", "*key1", 2, newcursor, keys));
   ASSERT_EQ(keys, make_vec("key1"));
   ASSERT_EQ(newcursor, "key3");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "*key1", 2, newcursor, keys));
   ASSERT_TRUE(keys.empty());
   ASSERT_EQ(newcursor, "key5");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "*key1", 2, newcursor, keys));
   ASSERT_TRUE(keys.empty());
   ASSERT_EQ(newcursor, "otherkey1");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "*key1", 2, newcursor, keys));
   ASSERT_EQ(keys, make_vec("otherkey1"));
   ASSERT_EQ(newcursor, "otherkey3");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "*key1", 2, newcursor, keys));
   ASSERT_TRUE(keys.empty());
   ASSERT_EQ(newcursor, "");
@@ -687,20 +694,24 @@ TEST_F(State_Machine, scan) {
   ASSERT_OK(stateMachine()->set("abc", "8"));
   ASSERT_OK(stateMachine()->set("abcd", "8"));
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan("", "ab?", 3, newcursor, keys));
   ASSERT_EQ(keys, make_vec("aba", "abb", "abc"));
   ASSERT_EQ(newcursor, "abcd");
 
+  keys.clear();
   ASSERT_OK(stateMachine()->scan(newcursor, "ab?", 3, newcursor, keys));
   ASSERT_TRUE(keys.empty());
   ASSERT_EQ(newcursor, "");
 
   // Using a non-sense cursor
+  keys.clear();
   ASSERT_OK(stateMachine()->scan("zz", "ab?", 100, newcursor, keys));
   ASSERT_TRUE(keys.empty());
   ASSERT_EQ(newcursor, "");
 
   // Match only a single key
+  keys.clear();
   ASSERT_OK(stateMachine()->scan("", "abc", 100, newcursor, keys));
   ASSERT_EQ(keys, make_vec("abc"));
   ASSERT_EQ(newcursor, "");