Skip decoding completely during count and improve SK count perf by 20x

Summary:
8.0 is taking a new approach with doing counts - MySQL will call handler::records / records_with_index which will then use rnd_init+rnd_next / index_init+index_next to iterate through everything. If you try count(*) you'll see MyRocks 8.0 is actually much slower when iterating the table with secondary keys than 5.6 (the optimizer picks secondary key for whatever reason).

This is because MySQL no longer calls `handler::extra(KEY_READ)` so we no longer get the hints to only decode the key - instead we'll see that the SK is not covering so we fall back to PK, which dramatically slows down by *a lot*.

This fixes it by adding a m_iteration_only = true hint -  with this hint we'll just quickly iterate through everything without any decoding nor falling back to PK.

Note even with this fix, if you compare MyRocks 8.0 with InnoDB 8.0, you'll see that InnoDB 8.0 is also much faster - this is because InnoDB overrides the implementation and iterate the records directly in parallel (default to 4). This is something we can do as well.

Reviewed By: Pushapgl

Differential Revision: D26272733

Author: yizhang82

storage/rocksdb/ha_rocksdb.cc

@@ -50,6 +50,7 @@
 #include "my_bit.h"
 #include "my_stacktrace.h"
 #include "my_sys.h"
+#include "scope_guard.h"
 #include "sql/binlog.h"
 #include "sql/debug_sync.h"
 #include "sql-common/json_dom.h"
@@ -7108,6 +7109,7 @@ ha_rocksdb::ha_rocksdb(my_core::handlerton *const hton,
       m_pack_buffer(nullptr),
       m_lock_rows(RDB_LOCK_NONE),
       m_keyread_only(false),
+      m_iteration_only(false),
       m_insert_with_update(false),
       m_dup_key_found(false),
       mrr_rowid_reader(nullptr),
@@ -9085,6 +9087,11 @@ int ha_rocksdb::secondary_index_read(const int keyno, uchar *const buf) {
 
     /* Check if we've ran out of records of this index */
     if (m_key_descr_arr[keyno]->covers_key(key)) {
+      if (m_iteration_only) {
+        table->m_status = 0;
+        return 0;
+      }
+
       int rc = 0;
 
       // TODO: We could here check if we have ran out of range we're scanning
@@ -9756,6 +9763,22 @@ int ha_rocksdb::get_row_by_rowid(uchar *const buf, const char *const rowid,
   DBUG_RETURN(rc);
 }
 
+int ha_rocksdb::records(ha_rows *num_rows) {
+  m_iteration_only = true;
+  auto iteration_guard =
+      create_scope_guard([this]() { m_iteration_only = false; });
+  int count = handler::records(num_rows);
+  return count;
+}
+
+int ha_rocksdb::records_from_index(ha_rows *num_rows, uint index) {
+  m_iteration_only = true;
+  auto iteration_guard =
+      create_scope_guard([this]() { m_iteration_only = false; });
+  int count = handler::records_from_index(num_rows, index);
+  return count;
+}
+
 /**
   @return
     HA_EXIT_SUCCESS  OK
@@ -11301,7 +11324,7 @@ int ha_rocksdb::rnd_next(uchar *const buf) {
 int ha_rocksdb::rnd_next_with_direction(uchar *const buf, bool move_forward) {
   DBUG_ENTER_FUNC();
 
-  int rc;
+  int rc = 0;
   THD *thd = ha_thd();
 
   table->m_status = STATUS_NOT_FOUND;
@@ -11349,6 +11372,11 @@ int ha_rocksdb::rnd_next_with_direction(uchar *const buf, bool move_forward) {
       break;
     }
 
+    if (m_iteration_only) {
+      table->m_status = 0;
+      break;
+    }
+
     if (m_lock_rows != RDB_LOCK_NONE) {
       /*
         Lock the row we've just read.

storage/rocksdb/ha_rocksdb.h

@@ -272,6 +272,9 @@ class ha_rocksdb : public my_core::handler {
   /* true means we're doing an index-only read. false means otherwise. */
   bool m_keyread_only;
 
+  /* We only iterate but don't need to decode anything */
+  bool m_iteration_only;
+
   bool m_skip_scan_it_next_call;
 
   /* true means we are accessing the first row after a snapshot was created */
@@ -1005,6 +1008,10 @@ class ha_rocksdb : public my_core::handler {
   void build_decoder();
   void check_build_decoder();
 
+ protected:
+  int records(ha_rows *num_rows) override;
+  int records_from_index(ha_rows *num_rows, uint index) override;
+
  public:
   virtual void rpl_before_delete_rows() override;
   virtual void rpl_after_delete_rows() override;