Issue #25: Make ORDER BY DESC faster

fbshipit-source-id: 1151e93

storage/rocksdb/ha_rocksdb.h

@@ -169,8 +169,7 @@ class ha_rocksdb: public handler
 
   int create_key_defs(TABLE *table_arg, const char *db_table, uint len);
   int secondary_index_read(int keyno, uchar *buf);
-  void setup_index_scan(rocksdb::ColumnFamilyHandle *cf,
-                        rocksdb::Slice *slice);
+  void setup_index_scan(RDBSE_KEYDEF *keydef, rocksdb::Slice *slice);
   int get_row_by_rowid(uchar *buf, const char *pk_tuple, uint pk_tuple_size);
 
   void update_auto_incr_val();
@@ -316,6 +315,10 @@ class ha_rocksdb: public handler
 
   int index_first(uchar *buf);
   int index_last(uchar *buf);
+private:
+  int index_first_intern(uchar *buf);
+  int index_last_intern(uchar *buf);
+public:
   int index_end();
 
   void unlock_row();

storage/rocksdb/rdb_applyiter.cc

@@ -41,11 +41,13 @@ Apply_changes_iter::~Apply_changes_iter()
 }
 
 
-void Apply_changes_iter::init(Row_table *trx_arg, rocksdb::Iterator *rdb_arg)
+void Apply_changes_iter::init(bool is_reverse_arg, Row_table *trx_arg,
+                              rocksdb::Iterator *rdb_arg)
 {
   delete trx;
   delete rdb;
-  trx= new Row_table_iter(trx_arg);
+  is_reverse= is_reverse_arg;
+  trx= new Row_table_iter(trx_arg, is_reverse);
   rdb= rdb_arg;
   valid= false;
 }
@@ -136,6 +138,9 @@ void Apply_changes_iter::advance(int direction)
       int cmp= direction *
                compare_mem_comparable_keys((const uchar*)trx_key.data(), trx_key.size(),
                                            (const uchar*)rdb_key.data(), rdb_key.size());
+      if (is_reverse)
+        cmp *= -1;
+
       if (!cmp) // keys are equal
       {
         if (trx->is_tombstone())

storage/rocksdb/rdb_applyiter.h

@@ -17,6 +17,8 @@
 class Row_table;
 class Row_table_iter;
 
+class RDBSE_KEYDEF;
+
 /*
   A class that looks like RocksDB's iterator, but internally it takes into
   account the changes made by the transaction.
@@ -33,10 +35,14 @@ class Apply_changes_iter
   /* These are the iterators we're merging. We own them, so should free them */
   Row_table_iter *trx;
   rocksdb::Iterator* rdb;
+
+  /* If true, we're scanning reverse-ordered data */
+  bool is_reverse;
 public:
   Apply_changes_iter();
   ~Apply_changes_iter();
-  void init(Row_table *trx_arg, rocksdb::Iterator *rdb_arg);
+  void init(bool is_reverse_arg, Row_table *trx_arg,
+            rocksdb::Iterator *rdb_arg);
 
   void Next();
   void Prev();

storage/rocksdb/rdb_cf_manager.cc

@@ -30,6 +30,18 @@
 #endif
 #include <inttypes.h>
 
+
+/* Check if ColumnFamily name says it's a reverse-ordered CF */
+bool is_cf_name_reverse(const char *name)
+{
+  /* NULL means the default CF is used.. (TODO: can the default CF be reverse?) */
+  if (name && !strncmp(name, "rev:", 4))
+    return true;
+  else
+    return false;
+}
+
+
 void Column_family_manager::init(std::vector<std::string> *names,
                                  std::vector<rocksdb::ColumnFamilyHandle*> *handles)
 {

storage/rocksdb/rdb_cf_options.h

@@ -21,6 +21,8 @@ namespace rocksdb {
   class ColumnFamilyOptions;
 }
 
+bool is_cf_name_reverse(const char *name);
+
 /*
   Per-column family options configs.
 

storage/rocksdb/rdb_datadic.cc

@@ -962,7 +962,7 @@ bool Table_ddl_manager::init(rocksdb::DB *rdb_dict)
         look at Field* objects and set max_length and other attributes
       */
       tdef->key_descr[keyno]= new RDBSE_KEYDEF(index_number, keyno, NULL,
-                                               false);
+                                               false, false);
 
       /* Keep track of what was the last index number we saw */
       if (max_number < index_number)

storage/rocksdb/rdb_datadic.h

@@ -199,9 +199,11 @@ class RDBSE_KEYDEF
   }
 
   RDBSE_KEYDEF(uint indexnr_arg, uint keyno_arg,
-               rocksdb::ColumnFamilyHandle* cf_handle_arg, bool is_auto_cf_arg) :
+               rocksdb::ColumnFamilyHandle* cf_handle_arg,
+               bool is_reverse_cf_arg, bool is_auto_cf_arg) :
     index_number(indexnr_arg),
     cf_handle(cf_handle_arg),
+    is_reverse_cf(is_reverse_cf_arg),
     is_auto_cf(is_auto_cf_arg),
     pk_part_no(NULL),
     pack_info(NULL),
@@ -219,9 +221,11 @@ class RDBSE_KEYDEF
 
   void setup(TABLE *table);
   void set_cf_handle(rocksdb::ColumnFamilyHandle* cf_handle_arg,
+                     bool is_reverse_cf_arg,
                      bool is_auto_cf_arg)
   {
     cf_handle= cf_handle_arg;
+    is_reverse_cf= is_reverse_cf_arg;
     is_auto_cf= is_auto_cf_arg;
   }
 
@@ -235,7 +239,11 @@ class RDBSE_KEYDEF
   uchar index_number_storage_form[INDEX_NUMBER_SIZE];
 
   rocksdb::ColumnFamilyHandle* cf_handle;
+
 public:
+  /* If true, the column family stores data in the reverse order */
+  bool is_reverse_cf;
+
   bool is_auto_cf;
 private:
 

storage/rocksdb/rdb_debug.cc

@@ -83,23 +83,26 @@ void dump_value(FILE *out, const rocksdb::Slice &val)
 
 void dump_trx_changes(FILE *out, Row_table &changes)
 {
-  Row_table_iter iter(&changes);
-  fprintf(out, "TRX %p\n", current_thd);
-  for (iter.SeekToFirst(); iter.Valid(); iter.Next())
+  for (int reverse= 0; reverse <= 1; reverse++)
+  Row_table_iter iter(&changes, (bool)reverse);
   {
-    if (iter.is_tombstone())
+    fprintf(out, "TRX %p\n", current_thd);
+    for (iter.SeekToFirst(); iter.Valid(); iter.Next())
     {
-      fprintf(out, "DEL ");
-      dump_value(out, iter.key());
-      fputs("\n", out);
-    }
-    else
-    {
-      fprintf(out, "PUT ");
-      dump_value(out, iter.key());
-      fputs(" ", out);
-      dump_value(out, iter.value());
-      fputs("\n", out);
+      if (iter.is_tombstone())
+      {
+        fprintf(out, "DEL ");
+        dump_value(out, iter.key());
+        fputs("\n", out);
+      }
+      else
+      {
+        fprintf(out, "PUT ");
+        dump_value(out, iter.key());
+        fputs(" ", out);
+        dump_value(out, iter.value());
+        fputs("\n", out);
+      }
     }
   }
 }