为什么MySQL优化器不使用所有列索引？

Question

Percona MySQL 5.7

table scheeme:

CREATE TABLE Developer.Rate (
  ID bigint(20) UNSIGNED NOT NULL AUTO_INCREMENT,
  TIME datetime NOT NULL,
  BASE varchar(3) NOT NULL,
  QUOTE varchar(3) NOT NULL,
  BID double NOT NULL,
  ASK double NOT NULL,
  PRIMARY KEY (ID),
  INDEX IDX_TIME (TIME),
  UNIQUE INDEX IDX_UK (BASE, QUOTE, TIME)
)
ENGINE = INNODB
ROW_FORMAT = COMPRESSED;

I try to make request for latests data before selected period. The optimazer use no-complete unique key, only 2 columns of 3.

If I do request in common way:

EXPLAIN FORMAT=JSON
SELECT
  BID
FROM 
  Rate
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
ORDER BY 
  `TIME` DESC 
LIMIT 1
;

"Explain" shows that only 2 first columns of index are used: BASE, QUOTE

{
  "query_block": {
    "select_id": 1,
    "cost_info": {
      "query_cost": "10231052.40"
    },
    "ordering_operation": {
      "using_filesort": false,
      "table": {
        "table_name": "Rate",
        "access_type": "ref",
        "possible_keys": [
          "IDX_UK",
          "IDX_TIME"
        ],
        "key": "IDX_UK",
        "used_key_parts": [
          "BASE",
          "QUOTE"
        ],
        "key_length": "22",
        "ref": [
          "const",
          "const"
        ],
        "rows_examined_per_scan": 45966462,
        "rows_produced_per_join": 22983231,
        "filtered": "50.00",
        "cost_info": {
          "read_cost": "1037760.00",
          "eval_cost": "4596646.20",
          "prefix_cost": "10231052.40",
          "data_read_per_join": "1G"
        },
        "used_columns": [
          "ID",
          "TIME",
          "BASE",
          "QUOTE",
          "BID"
        ],
        "attached_condition": "((`Developer`.`Rate`.`BASE` <=> 'EUR') and (`Developer`.`Rate`.`QUOTE` <=> 'USD') and (`Developer`.`Rate`.`TIME` <= <cache>((now() - interval 1 month))))"
      }
    }
  }
}

But if you force the optimizer to use IDX_UK, MySQL uses all 3 columns in the request:

EXPLAIN FORMAT=JSON
SELECT
  BID
FROM 
  Rate FORCE INDEX(IDX_UK)
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
ORDER BY 
  `TIME` DESC 
LIMIT 1

{
  "query_block": {
    "select_id": 1,
    "cost_info": {
      "query_cost": "10231052.40"
    },
    "ordering_operation": {
      "using_filesort": false,
      "table": {
        "table_name": "Rate",
        "access_type": "range",
        "possible_keys": [
          "IDX_UK"
        ],
        "key": "IDX_UK",
        "used_key_parts": [
          "BASE",
          "QUOTE",
          "TIME"
        ],
        "key_length": "27",
        "rows_examined_per_scan": 45966462,
        "rows_produced_per_join": 15320621,
        "filtered": "100.00",
        "index_condition": "((`Developer`.`Rate`.`BASE` = 'EUR') and (`Developer`.`Rate`.`QUOTE` = 'USD') and (`Developer`.`Rate`.`TIME` <= <cache>((now() - interval 1 month))))",
        "cost_info": {
          "read_cost": "1037760.00",
          "eval_cost": "3064124.31",
          "prefix_cost": "10231052.40",
          "data_read_per_join": "818M"
        },
        "used_columns": [
          "ID",
          "TIME",
          "BASE",
          "QUOTE",
          "BID"
        ]
      }
    }
  }
}

Why the optimizer don't use all 3 columns without explicit declaration of index?

Added:

A'm I understanding right, I should to use request like this?

Reuest example:

EXPLAIN FORMAT=JSON
SELECT
  BID
FROM 
  Rate
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
ORDER BY 
  BASE DESC, QUOTE DESC, TIME DESC
LIMIT 1

If I understand it right, the output of Explain vouldn't be better. There are still only 2 columns are used without TIME

Explain Output

{ "query_block": { "select_id": 1, "cost_info": { "query_cost": "10384642.20" }, "ordering_operation": { "using_filesort": false, "table": { "table_name": "Rate", "access_type": "ref", "possible_keys": [ "IDX_UK", "IDX_TIME" ], "key": "IDX_UK", "used_key_parts": [ "BASE", "QUOTE" ], "key_length": "22", "ref": [ "const", "const" ], "rows_examined_per_scan": 46734411, "rows_produced_per_join": 23367205, "filtered": "50.00", "index_condition": "(( Developer . Rate . BASE <=> 'EUR') and ( Developer . Rate . QUOTE <=> 'USD') and ( Developer . Rate . TIME <= ((now() - interval 1 month))))", "cost_info": { "read_cost": "1037760.00", "eval_cost": "4673441.10", "prefix_cost": "10384642.20", "data_read_per_join": "1G" }, "used_columns": [ "ID", "TIME", "BASE", "QUOTE", "BID" ] } } } }

Added 2:

I made these 4 requests:

— 1 —


<code>FLUSH STATUS;
SELECT
  BID
FROM 
  Rate
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
LIMIT 1;
SHOW SESSION STATUS LIKE 'Handler%';</code>

— 2 —

<code>FLUSH STATUS;
SELECT
  BID
FROM 
  Rate FORCE INDEX (IDX_UK)
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
LIMIT 1;
SHOW SESSION STATUS LIKE 'Handler%';
</code>

— 3 —

<code>FLUSH STATUS;
SELECT
  BID
FROM 
  Rate
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
ORDER BY 
  `TIME` DESC 
LIMIT 1;
SHOW SESSION STATUS LIKE 'Handler%';</code>

— 4 —

<code>
FLUSH STATUS;
SELECT
  BID
FROM 
  Rate FORCE INDEX (IDX_UK)
WHERE 
  BASE = 'EUR' 
  AND QUOTE = 'USD' 
  AND `TIME` <= (NOW() - INTERVAL 1 MONTH) 
ORDER BY 
  `TIME` DESC 
LIMIT 1;
SHOW SESSION STATUS LIKE 'Handler%';</code>

The output of session_status is the same in all requests except request 3. In output of request 3: Handler_read_prev = 486474; In output of all ather requests: Handler_read_prev = 0;

Added 3:

I made a copy of the table, removed Id field, promoted UNIQUE key as PRIMARY.

The scheme:

CREATE TABLE Developer.Rate2 (
  TIME datetime NOT NULL,
  BASE varchar(3) NOT NULL,
  QUOTE varchar(3) NOT NULL,
  BID double NOT NULL,
  ASK double NOT NULL,
  PRIMARY KEY (BASE, QUOTE, TIME),
  INDEX IDX_BID_ASK (BID, ASK)
)
ENGINE = INNODB
AVG_ROW_LENGTH = 26
CHARACTER SET utf8
COLLATE utf8_general_ci
ROW_FORMAT = COMPRESSED;

{
  "query_block": {
    "select_id": 1,
    "cost_info": {
      "query_cost": "9673452.20"
    },
    "ordering_operation": {
      "using_filesort": false,
      "table": {
        "table_name": "Rate2",
        "access_type": "range",
        "possible_keys": [
          "PRIMARY"
        ],
        "key": "PRIMARY",
        "used_key_parts": [
          "BASE",
          "QUOTE",
          "TIME"
        ],
        "key_length": "27",
        "rows_examined_per_scan": 48023345,
        "rows_produced_per_join": 16006180,
        "filtered": "100.00",
        "cost_info": {
          "read_cost": "68783.20",
          "eval_cost": "3201236.12",
          "prefix_cost": "9673452.20",
          "data_read_per_join": "732M"
        },
        "used_columns": [
          "TIME",
          "BASE",
          "QUOTE",
          "BID"
        ],
        "attached_condition": "((`Developer`.`Rate2`.`BASE` = 'EUR') and (`Developer`.`Rate2`.`QUOTE` = 'USD') and (`Developer`.`Rate2`.`TIME` <= <cache>((now() - interval 1 month))))"
      }
    }
  }
}

Now the request really works and Explain shows all 3 columns are used. This variant works.

Answer 1

Get rid of ID , it is of no use. Promote your UNIQUE key to be PRIMARY . Now, magically, the query will be faster, and the Question you posed will become moot. (You may also need the DESC trick that lorraine suggested.)

Here's another technique to compare performance:

FLUSH STATUS;
SELECT ...;
SHOW SESSION STATUS LIKE 'Handler%';

I would be interested to see the output from the SHOW for with and without the DESC trick. And with/without the FORCE INDEX you alluded to.

Why faster? Your query was using a secondary index, but it needed bid , which was not 'covered' by the index. To get bid , the PRIMARY KEY needed to be drilled down in the 'data'. By changing it so that the PK is used, this extra drill-down is obviated.

Answer 2

The behavior you describe (ref access instead of range access over more columns) reminds me of Bug#81341 and Bug#87613 . These bugs were fixed in MySQL 5.7.17 and 5.7.21, respectively. Which version are you using?