Re: [Maria-developers] 0520a00: MDEV-7964 - delete_dynamic() takes 0.12% in OLTP RO
Hi, Sergey! On Apr 16, svoj@mariadb.org wrote:
revision-id: 0520a009c7139c940ab92345054bf0dba7808b11 parent(s): eb059cd41e5553597f81fd90ad55c07d76d317ce committer: Sergey Vojtovich branch nick: mariadb timestamp: 2015-04-16 18:38:26 +0400 message:
MDEV-7964 - delete_dynamic() takes 0.12% in OLTP RO ... diff --git a/sql/sql_base.cc b/sql/sql_base.cc index 0e9f668..d36de7f 100644 --- a/sql/sql_base.cc +++ b/sql/sql_base.cc @@ -4240,9 +4233,17 @@ thr_lock_type read_lock_type_for_table(THD *thd, DBUG_RETURN(true); }
- if (! (flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK) && - schema_set.insert(table)) - DBUG_RETURN(TRUE); + /* Scoped locks: Take intention exclusive locks on all involved schemas. */ + if (!(flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK)) + { + MDL_request *schema_request= new (thd->mem_root) MDL_request; + if (schema_request == NULL) + DBUG_RETURN(TRUE); + schema_request->init(MDL_key::SCHEMA, table->db, "", + MDL_INTENTION_EXCLUSIVE, + MDL_TRANSACTION); + mdl_requests.push_front(schema_request); + }
As far as I understand, the Hash_set was used here to make only one MDL request per schema. Now if you'll have 10 tables all in one schema, you'll make push 10 MDL requests instead of one. Which may be fine, if MDL subsystem places only one lock in this case (iirc, it does). But what is cheaper - collapse identical schemas into one mdl request here (with Hash_set) or let mdl sort it out later?
mdl_requests.push_front(&table->mdl_request); }
Regards, Sergei
Hi Sergei, On Fri, May 08, 2015 at 09:21:48PM +0200, Sergei Golubchik wrote:
Hi, Sergey!
On Apr 16, svoj@mariadb.org wrote:
revision-id: 0520a009c7139c940ab92345054bf0dba7808b11 parent(s): eb059cd41e5553597f81fd90ad55c07d76d317ce committer: Sergey Vojtovich branch nick: mariadb timestamp: 2015-04-16 18:38:26 +0400 message:
MDEV-7964 - delete_dynamic() takes 0.12% in OLTP RO ... diff --git a/sql/sql_base.cc b/sql/sql_base.cc index 0e9f668..d36de7f 100644 --- a/sql/sql_base.cc +++ b/sql/sql_base.cc @@ -4240,9 +4233,17 @@ thr_lock_type read_lock_type_for_table(THD *thd, DBUG_RETURN(true); }
- if (! (flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK) && - schema_set.insert(table)) - DBUG_RETURN(TRUE); + /* Scoped locks: Take intention exclusive locks on all involved schemas. */ + if (!(flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK)) + { + MDL_request *schema_request= new (thd->mem_root) MDL_request; + if (schema_request == NULL) + DBUG_RETURN(TRUE); + schema_request->init(MDL_key::SCHEMA, table->db, "", + MDL_INTENTION_EXCLUSIVE, + MDL_TRANSACTION); + mdl_requests.push_front(schema_request); + }
As far as I understand, the Hash_set was used here to make only one MDL request per schema. Now if you'll have 10 tables all in one schema, you'll make push 10 MDL requests instead of one. It was probably supposed to, but since Hash_set is non-unique it didn't do anything useful. That is we would get 10 MDL request with old code too.
Which may be fine, if MDL subsystem places only one lock in this case (iirc, it does). But what is cheaper - collapse identical schemas into one mdl request here (with Hash_set) or let mdl sort it out later? MDL doesn't eliminate them, but rather clones compatible requests. Performance wise it will be slower in certain cases, but since this is DDL is performance that important?
Thanks, Sergey
Hi, Sergey! On May 09, Sergey Vojtovich wrote:
@@ -4240,9 +4233,17 @@ thr_lock_type read_lock_type_for_table(THD *thd, DBUG_RETURN(true); }
- if (! (flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK) && - schema_set.insert(table)) - DBUG_RETURN(TRUE); + /* Scoped locks: Take intention exclusive locks on all involved schemas. */ + if (!(flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK)) + { + MDL_request *schema_request= new (thd->mem_root) MDL_request; + if (schema_request == NULL) + DBUG_RETURN(TRUE); + schema_request->init(MDL_key::SCHEMA, table->db, "", + MDL_INTENTION_EXCLUSIVE, + MDL_TRANSACTION); + mdl_requests.push_front(schema_request); + }
As far as I understand, the Hash_set was used here to make only one MDL request per schema. Now if you'll have 10 tables all in one schema, you'll make push 10 MDL requests instead of one. It was probably supposed to, but since Hash_set is non-unique it didn't do anything useful. That is we would get 10 MDL request with old code too.
Oh, ok.
Which may be fine, if MDL subsystem places only one lock in this case (iirc, it does). But what is cheaper - collapse identical schemas into one mdl request here (with Hash_set) or let mdl sort it out later? MDL doesn't eliminate them, but rather clones compatible requests. Performance wise it will be slower in certain cases, but since this is DDL is performance that important?
If it isn't, why are you eliminating Hash_set? Regards, Sergei
Hi Sergei, On Sat, May 09, 2015 at 09:43:49AM +0200, Sergei Golubchik wrote:
Hi, Sergey!
On May 09, Sergey Vojtovich wrote:
@@ -4240,9 +4233,17 @@ thr_lock_type read_lock_type_for_table(THD *thd, DBUG_RETURN(true); }
- if (! (flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK) && - schema_set.insert(table)) - DBUG_RETURN(TRUE); + /* Scoped locks: Take intention exclusive locks on all involved schemas. */ + if (!(flags & MYSQL_OPEN_SKIP_SCOPED_MDL_LOCK)) + { + MDL_request *schema_request= new (thd->mem_root) MDL_request; + if (schema_request == NULL) + DBUG_RETURN(TRUE); + schema_request->init(MDL_key::SCHEMA, table->db, "", + MDL_INTENTION_EXCLUSIVE, + MDL_TRANSACTION); + mdl_requests.push_front(schema_request); + }
As far as I understand, the Hash_set was used here to make only one MDL request per schema. Now if you'll have 10 tables all in one schema, you'll make push 10 MDL requests instead of one. It was probably supposed to, but since Hash_set is non-unique it didn't do anything useful. That is we would get 10 MDL request with old code too.
Oh, ok.
Which may be fine, if MDL subsystem places only one lock in this case (iirc, it does). But what is cheaper - collapse identical schemas into one mdl request here (with Hash_set) or let mdl sort it out later? MDL doesn't eliminate them, but rather clones compatible requests. Performance wise it will be slower in certain cases, but since this is DDL is performance that important?
If it isn't, why are you eliminating Hash_set? Because Hash_set was created for DML too. :(
Thanks, Sergey
Hi, Sergey! On May 09, Sergey Vojtovich wrote:
Which may be fine, if MDL subsystem places only one lock in this case (iirc, it does). But what is cheaper - collapse identical schemas into one mdl request here (with Hash_set) or let mdl sort it out later? MDL doesn't eliminate them, but rather clones compatible requests. Performance wise it will be slower in certain cases, but since this is DDL is performance that important?
If it isn't, why are you eliminating Hash_set?
Because Hash_set was created for DML too. :(
I see. I believe it's rather unusual to have DDL that involves many tables, so fine, ok to push. Regards, Sergei
participants (2)
-
Sergei Golubchik -
Sergey Vojtovich