#261 auto detect index scans to fix perf problem with partitions

mysql-test/suite/tokudb.bugs/r/part_index_scan.result

+set default_storage_engine='tokudb';
+drop table if exists t,t1,t2,t3;
+CREATE TABLE `t` (
+`num` int(10) unsigned auto_increment NOT NULL,
+`val` varchar(32) DEFAULT NULL,
+PRIMARY KEY (`num`)
+);
+INSERT INTO t values (null,null);
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+SELECT count(*) FROM t;
+count(*)
+8388608
+CREATE TABLE `t1` (
+`num` int(10) unsigned NOT NULL,
+`val` varchar(32) DEFAULT NULL,
+PRIMARY KEY (`num`)
+);
+CREATE TABLE `t2` (
+`num` int(10) unsigned NOT NULL,
+`val` varchar(32) DEFAULT NULL,
+PRIMARY KEY (`num`)
+) 
+PARTITION BY HASH (num) PARTITIONS 10;
+CREATE TABLE `t3` (
+`num` int(10) unsigned NOT NULL,
+`val` varchar(32) DEFAULT NULL,
+PRIMARY KEY (`num`)
+)
+PARTITION BY RANGE (num)
+(PARTITION p0 VALUES LESS THAN (1000000),
+PARTITION p1 VALUES LESS THAN (2000000),
+PARTITION p2 VALUES LESS THAN (3000000),
+PARTITION p3 VALUES LESS THAN (4000000),
+PARTITION p4 VALUES LESS THAN (5000000),
+PARTITION p5 VALUES LESS THAN (6000000),
+PARTITION p6 VALUES LESS THAN (7000000),
+PARTITION p7 VALUES LESS THAN (8000000),
+PARTITION px VALUES LESS THAN MAXVALUE);
+insert into t1 select * from t;
+insert into t2 select * from t;
+insert into t3 select * from t;
+select count(*) from t1;
+count(*)
+8388608
+select count(*) from t2;
+count(*)
+8388608
+1
+select count(*) from t3;
+count(*)
+8388608
+1
+select count(*) from t1 where num>7000000;
+count(*)
+1847274
+select count(*) from t2 where num>7000000;
+count(*)
+1847274
+1
+select count(*) from t3 where num>7000000;
+count(*)
+1847274
+1
+drop table if exists t,t1,t2,t3;

mysql-test/suite/tokudb.bugs/t/part_index_scan.test

+# verify that index scans on parititions are not slow
+# due totokudb bulk fetch not being used
+source include/have_tokudb.inc;
+set default_storage_engine='tokudb';
+disable_warnings;
+drop table if exists t,t1,t2,t3;
+
+CREATE TABLE `t` (
+  `num` int(10) unsigned auto_increment NOT NULL,
+  `val` varchar(32) DEFAULT NULL,
+  PRIMARY KEY (`num`)
+);
+
+# put 8M rows into t
+INSERT INTO t values (null,null);
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+INSERT INTO t SELECT null,null FROM t;
+SELECT count(*) FROM t;
+
+CREATE TABLE `t1` (
+  `num` int(10) unsigned NOT NULL,
+  `val` varchar(32) DEFAULT NULL,
+  PRIMARY KEY (`num`)
+);
+
+CREATE TABLE `t2` (
+  `num` int(10) unsigned NOT NULL,
+  `val` varchar(32) DEFAULT NULL,
+  PRIMARY KEY (`num`)
+) 
+PARTITION BY HASH (num) PARTITIONS 10;
+
+CREATE TABLE `t3` (
+  `num` int(10) unsigned NOT NULL,
+  `val` varchar(32) DEFAULT NULL,
+  PRIMARY KEY (`num`)
+)
+PARTITION BY RANGE (num)
+(PARTITION p0 VALUES LESS THAN (1000000),
+ PARTITION p1 VALUES LESS THAN (2000000),
+ PARTITION p2 VALUES LESS THAN (3000000),
+ PARTITION p3 VALUES LESS THAN (4000000),
+ PARTITION p4 VALUES LESS THAN (5000000),
+ PARTITION p5 VALUES LESS THAN (6000000),
+ PARTITION p6 VALUES LESS THAN (7000000),
+ PARTITION p7 VALUES LESS THAN (8000000),
+ PARTITION px VALUES LESS THAN MAXVALUE);
+
+insert into t1 select * from t;
+insert into t2 select * from t;
+insert into t3 select * from t;
+
+# verify that full index scans on partitioned tables t2 and t3 are comparable to a non-partitioned table t1
+let $s = `select to_seconds(now())`;
+select count(*) from t1;
+let $t1 = `select to_seconds(now()) - $s`;
+# echo $t1;
+
+let $s = `select to_seconds(now())`;
+select count(*) from t2;
+let $t2 = `select to_seconds(now()) - $s`;
+# echo $t2;
+let $d = `select abs($t2 - $t1) <= $t1`;
+echo $d;
+
+let $s = `select to_seconds(now())`;
+select count(*) from t3;
+let $t3 = `select to_seconds(now()) - $s`;
+# echo $t3;
+let $d = `select abs($t3 - $t1) <= $t1`;
+echo $d;
+
+let $s = `select to_seconds(now())`;
+select count(*) from t1 where num>7000000;
+let $t1 = `select to_seconds(now()) - $s`;
+# echo $t1;
+
+let $s = `select to_seconds(now())`;
+select count(*) from t2 where num>7000000;
+let $t2 = `select to_seconds(now()) - $s`;
+# echo $t2;
+let $d = `select abs($t2 - $t1) <= $t1`;
+echo $d;
+
+let $s = `select to_seconds(now())`;
+select count(*) from t3 where num>7000000;
+let $t3 = `select to_seconds(now()) - $s`;
+# echo $t3;
+let $d = `select abs($t3 - $t1) <= $t1`;
+echo $d;
+
+enable_warnings;
+drop table if exists t,t1,t2,t3;

storage/tokudb/ha_tokudb.cc

@@ -4504,6 +4504,7 @@ int ha_tokudb::index_init(uint keynr, bool sorted) {
     }
     invalidate_bulk_fetch();
     doing_bulk_fetch = false;
+    maybe_index_scan = false;
     error = 0;
 exit:
     TOKUDB_HANDLER_DBUG_RETURN(error);
@@ -5247,19 +5248,23 @@ cleanup:
 
 int ha_tokudb::get_next(uchar* buf, int direction, DBT* key_to_compare, bool do_key_read) {
     int error = 0;
-    uint32_t flags = SET_PRELOCK_FLAG(0);
-    THD* thd = ha_thd();
-    tokudb_trx_data* trx = (tokudb_trx_data *) thd_get_ha_data(thd, tokudb_hton);;
-    bool need_val;
     HANDLE_INVALID_CURSOR();
 
+    if (maybe_index_scan) {
+        maybe_index_scan = false;
+        if (!range_lock_grabbed) {
+            error = prepare_index_scan();
+        }
+    }
+    
+    if (!error) {
+        uint32_t flags = SET_PRELOCK_FLAG(0);
+
         // we need to read the val of what we retrieve if
         // we do NOT have a covering index AND we are using a clustering secondary
         // key
-    need_val = (do_key_read == 0) && 
-                (tokudb_active_index == primary_key || 
-                 key_is_clustering(&table->key_info[tokudb_active_index])
-                       );
+        bool need_val = (do_key_read == 0) && 
+            (tokudb_active_index == primary_key || key_is_clustering(&table->key_info[tokudb_active_index]));
 
         if ((bytes_used_in_range_query_buff - curr_range_query_buff_offset) > 0) {
             error = read_data_from_range_query_buff(buf, need_val, do_key_read);
@@ -5328,6 +5333,7 @@ int ha_tokudb::get_next(uchar* buf, int direction, DBT* key_to_compare, bool do_
                 error = handle_cursor_error(error, HA_ERR_END_OF_FILE, tokudb_active_index);
             }
         }
+    }
 
     //
     // at this point, one of two things has happened
@@ -5337,12 +5343,15 @@ int ha_tokudb::get_next(uchar* buf, int direction, DBT* key_to_compare, bool do_
     // read the full row by doing a point query into the 
     // main table.
     //
-    
     if (!error && !do_key_read && (tokudb_active_index != primary_key) && !key_is_clustering(&table->key_info[tokudb_active_index])) {
         error = read_full_row(buf);
     }
+
+    if (!error) {
+        tokudb_trx_data* trx = (tokudb_trx_data *) thd_get_ha_data(ha_thd(), tokudb_hton);
         trx->stmt_progress.queried++;
-    track_progress(thd);
+        track_progress(ha_thd());
+    }
 cleanup:
     return error;
 }
@@ -5411,8 +5420,7 @@ int ha_tokudb::index_first(uchar * buf) {
     info.buf = buf;
     info.keynr = tokudb_active_index;
 
-    error = cursor->c_getf_first(cursor, flags,
-            SMART_DBT_CALLBACK(key_read), &info);
+    error = cursor->c_getf_first(cursor, flags, SMART_DBT_CALLBACK(key_read), &info);
     error = handle_cursor_error(error,HA_ERR_END_OF_FILE,tokudb_active_index);
 
     //
@@ -5422,9 +5430,11 @@ int ha_tokudb::index_first(uchar * buf) {
     if (!error && !key_read && (tokudb_active_index != primary_key) && !key_is_clustering(&table->key_info[tokudb_active_index])) {
         error = read_full_row(buf);
     }
+    if (trx) {
         trx->stmt_progress.queried++;
+    }
     track_progress(thd);
-    
+    maybe_index_scan = true;    
 cleanup:
     TOKUDB_HANDLER_DBUG_RETURN(error);
 }
@@ -5454,8 +5464,7 @@ int ha_tokudb::index_last(uchar * buf) {
     info.buf = buf;
     info.keynr = tokudb_active_index;
 
-    error = cursor->c_getf_last(cursor, flags,
-            SMART_DBT_CALLBACK(key_read), &info);
+    error = cursor->c_getf_last(cursor, flags, SMART_DBT_CALLBACK(key_read), &info);
     error = handle_cursor_error(error,HA_ERR_END_OF_FILE,tokudb_active_index);
     //
     // still need to get entire contents of the row if operation done on
@@ -5469,6 +5478,7 @@ int ha_tokudb::index_last(uchar * buf) {
         trx->stmt_progress.queried++;
     }
     track_progress(thd);
+    maybe_index_scan = true;
 cleanup:
     TOKUDB_HANDLER_DBUG_RETURN(error);
 }

storage/tokudb/ha_tokudb.h

@@ -251,6 +251,7 @@ private:
     uint64_t bulk_fetch_iteration;
     uint64_t rows_fetched_using_bulk_fetch;
     bool doing_bulk_fetch;
+    bool maybe_index_scan;
 
     //
     // buffer used to temporarily store a "packed key"