[MDEV-19593] Split create_schema_table() into virtual methods in Type_handler - Jira

Details

Type: Task
Status: Closed (View Workflow)
Priority: Major
Resolution: Fixed
Fix Version/s: 10.5.0
Component/s: Data types, Information Schema
Labels:
None

Description

The code in create_schema_table() has switch's and if's testing MYSQL_TYPE_XXX.
This is not friendly to data type plugins.

Additionally, when schema tables are created in two steps:

On the first step, Items are created from ST_FIELD_INFO
On the second step, these Items create Fields
which is slow.

Let's do the following:

Add a new method in Type_handler:

   virtual Field *make_schema_field(TABLE *table,

                                    const Record_addr &addr,

                                    const ST_FIELD_INFO &def,

                                    bool show_field) const

Change the logic to create Fields directly from ST_FIELD_INFO.
For that purpose, and to be able to reuse the code in a convenient way, wrap create_tmp_table() into a class Create_tmp_table with approximately these methods:

class Create_tmp_table

public:

...

  // The former beginning of create_tmp_table()

  TABLE *start(THD *thd,

               TMP_TABLE_PARAM *param,

               const LEX_CSTRING *table_alias);

  // The former loop of create_tmp_table(), adding fields into the table

  bool add_fields(THD *thd, TABLE *table,

                  TMP_TABLE_PARAM *param, List<Item> &fields);

  // A new code adding fields directly from ST_FIELD_INFO

  bool add_schema_fields(THD *thd, TABLE *table,

                         TMP_TABLE_PARAM *param,

                         const ST_SCHEMA_TABLE &schema_table,

                         const MY_BITMAP &bitmap);

  // The forder end of create_tmp_table(), after the loop adding fields.

  bool finalize(THD *thd, TABLE *table, TMP_TABLE_PARAM *param,

                bool do_not_open, bool keep_row_order);

};

Remove method Item::create_field_for_schema()
Remove class Item_return_date_time
Remove class Item_blob

Attachments

Issue Links

blocks

MDEV-4912 Data type plugin API version 1

Closed

Activity

Alexander Barkov added a comment - 2019-08-17 00:55 - edited

Results of this change look good in many aspects.

Code change made summary

The old code made a really hard job on every INFORMATION_SCHEMA query:

converted every ST_FIELD_INFO into Item
then converted every new Item to Field
then put every new Field into List<Field>
and only after that called create_tmp_table() with the List<Field>

The new code just creates a table directly from the array of ST_FIELD_INFO without such heavy conversion

Unnecessary classes removed

These classes were used only in I_S table creation, in the ST_FIELD_INFO->Item->Field chain:

Item_return_date_time
Item_blob

After this change these classes were not needed any more, so they were removed.
Note, it should be also easy to remove class Item_return_int (in a separate change).

Creation of INFORMATION_SCHEMA tables is now faster

I_S tables are now created directly from an array of ST_FIELD_INFO, no such heavy conversion involved any more.

This example code runs 1% faster after the change:

mysql --socket=/tmp/mysql.sock -s --database=test << QEND

DELIMITER $$

CREATE OR REPLACE PROCEDURE p1()

BEGIN

  FOR i IN 1..1000000

DO

    BEGIN

      SELECT * FROM INFORMATION_SCHEMA.COLUMNS LIMIT 0;

    END;

  END FOR;

END;

$$

DELIMITER ;

QEND

time mysql --socket=/tmp/mysql.sock -s --database=test --execute="CALL p1()"

The code is now data type friendly

I_S columns of new data types now work out of the box, without a need of new wrapper classes like the former Item_blob, Item_return_date_time or Item_return_int. A new data type developer just needs to override Type_handler_xxx::create_schema_field(), which is one line method with a single new Field_xxx() statement.

User defined plugins are now able to have fields of their own data type in their pluggable INFORMATION_SCHEMA tables.
We'll be adding built-in SQL standard data types soon, such as TIMESTAMP WITH TIME ZONE, INTERVAL, so adding I_S columns of these types is also trivial.

It's now easy to extend creating of temporary tables

The new class Create_tmp_table now understands two input data formats:

List<Item>
ST_SCHEMA_TABLE

Adding new formats is now trivial and does not need heavy conversion like ST_FIELD_INFO->Item->Field. One just needs to add a proper add_fields() method.

Examples:

it's now trivial to create a temporary table directly from List<Item>. Currently we convert List<Item> into List<Field> in such cases. Direct table creation from List<Item> can make the code faster.
On ALTER, we convert every Field into Create_field. Such conversion is not really needed. We could create an altered table from a mixture of old Fields with new Create_fields, which would improve performance.

Alexander Barkov added a comment - 2019-08-17 00:55 - edited Results of this change look good in many aspects. Code change made summary The old code made a really hard job on every INFORMATION_SCHEMA query: converted every ST_FIELD_INFO into Item then converted every new Item to Field then put every new Field into List<Field> and only after that called create_tmp_table() with the List<Field> The new code just creates a table directly from the array of ST_FIELD_INFO without such heavy conversion Unnecessary classes removed These classes were used only in I_S table creation, in the ST_FIELD_INFO->Item->Field chain: Item_return_date_time Item_blob After this change these classes were not needed any more, so they were removed. Note, it should be also easy to remove class Item_return_int (in a separate change). Creation of INFORMATION_SCHEMA tables is now faster I_S tables are now created directly from an array of ST_FIELD_INFO, no such heavy conversion involved any more. This example code runs 1% faster after the change: mysql --socket=/tmp/mysql.sock -s --database=test << QEND DELIMITER $$ CREATE OR REPLACE PROCEDURE p1() BEGIN FOR i IN 1..1000000 DO BEGIN SELECT * FROM INFORMATION_SCHEMA.COLUMNS LIMIT 0; END ; END FOR ; END ; $$ DELIMITER ; QEND time mysql --socket=/tmp/mysql.sock -s --database=test --execute="CALL p1()" The code is now data type friendly I_S columns of new data types now work out of the box, without a need of new wrapper classes like the former Item_blob, Item_return_date_time or Item_return_int. A new data type developer just needs to override Type_handler_xxx::create_schema_field(), which is one line method with a single new Field_xxx() statement. User defined plugins are now able to have fields of their own data type in their pluggable INFORMATION_SCHEMA tables. We'll be adding built-in SQL standard data types soon, such as TIMESTAMP WITH TIME ZONE, INTERVAL, so adding I_S columns of these types is also trivial. It's now easy to extend creating of temporary tables The new class Create_tmp_table now understands two input data formats: List<Item> ST_SCHEMA_TABLE Adding new formats is now trivial and does not need heavy conversion like ST_FIELD_INFO->Item->Field. One just needs to add a proper add_fields() method. Examples: it's now trivial to create a temporary table directly from List<Item>. Currently we convert List<Item> into List<Field> in such cases. Direct table creation from List<Item> can make the code faster. On ALTER, we convert every Field into Create_field. Such conversion is not really needed. We could create an altered table from a mixture of old Fields with new Create_fields, which would improve performance.

MariaDB Server

Split create_schema_table() into virtual methods in Type_handler