[MDEV-11675] Lag Free Alter On Slave - Jira

Details

Type: Task
Status: Closed (View Workflow)
Priority: Critical
Resolution: Fixed
Fix Version/s: 10.8.1
Component/s: Replication
Labels:

Description

Desc:- This will split Alter into 2 different commits. START ALTER and COMMIT
/ROLLBACK ALTER , Start Alter will be written in binlog as soon as we get the
locks for the table, alter will proceeds as usual and at the time of writing
binlog if alter is successful we will write COMMIT Alter other wise ROLLBACK
Alter.

Implementation details are in develop list
https://lists.launchpad.net/maria-developers/msg12067.html

Attachments

Issue Links

causes

MDEV-33057 alter table form s3 to different engine on leader break replicas with binlog_alter_two_phase;

Open

MDEV-35431 my_snprintf fixes for 10.11+

Closed

MDEV-35474 Start Alter GTID Error Message Can Use Wrong Server_Id

In Review

MDEV-35665 Potential Buffer Overrun in Gtid_log_event::write()

Closed

is blocked by

MDEV-27528 Cluster nodes become unstable when a transaction is replicated through Async replication from Galera Primary node

Closed

is part of

MDEV-27373 Q1 2022 release merge

Closed

relates to

MDEV-742 LP:803649 - Xa recovery failed on client disconnection

Closed

MDEV-16223 Background ADD INDEX

Closed

MDEV-16281 Implement parallel CREATE INDEX, ALTER TABLE, or bulk load

Open

MDEV-16329 Engine-independent online ALTER TABLE

Closed

MDEV-17567 Atomic DDL

Closed

MDEV-21742 Add recovery to MDEV-11675 and handle corner cases

Closed

MDEV-27528 Cluster nodes become unstable when a transaction is replicated through Async replication from Galera Primary node

Closed

MDEV-7502 Automatic provisioning of slave

Open

MDEV-22777 Assertion `thd->rgi_slave' failed in process_commit_alter / process_rollback_alter

Closed

MDEV-22985 Assertion `!(thd->rgi_slave && thd->rgi_slave->did_mark_start_commit)' failed in ha_rollback_trans

Closed

MDEV-27349 ASAN use-after-poison in Query_log_event::Query_log_event/THD::binlog_query

Closed

MDEV-27628 Lag free ALTER statement breaks replication when executing binlog_alter_two_phase = 1,0 dynamically

Closed

mentioned in: Page Loading...; Page Loading...; Page Loading...

(1 is part of, 12 relates to, 3 mentioned in)

Activity

Ascending order - Click to sort in descending order

Marko Mäkelä added a comment - 2018-10-25 07:18 - edited

A variation of the same idea is in ~~MDEV-16223~~, for initiating ADD INDEX in the background.

This solution would be preferred instead of that one (~~MDEV-16223~~).

Marko Mäkelä added a comment - 2018-10-25 07:18 - edited A variation of the same idea is in MDEV-16223 , for initiating ADD INDEX in the background. This solution would be preferred instead of that one ( MDEV-16223 ).

Andrei Elkin added a comment - 2019-02-04 11:33

Note ~~MDEV-7974~~ is actually to provide an infrastructure to replicate the prepare and commit parts of an object not being limited to XA trx.
Specifically crash-safety (recovery) property of the prepare part are common.

Andrei Elkin added a comment - 2019-02-04 11:33 Note MDEV-7974 is actually to provide an infrastructure to replicate the prepare and commit parts of an object not being limited to XA trx. Specifically crash-safety (recovery) property of the prepare part are common.

Sachin Setiya (Inactive) added a comment - 2019-08-09 10:23

Hi ralf.gebhardt@mariadb.com
There is no update , because we there was no proposal for this project .

Sachin Setiya (Inactive) added a comment - 2019-08-09 10:23 Hi ralf.gebhardt@mariadb.com There is no update , because we there was no proposal for this project .

Ralf Gebhardt added a comment - 2019-08-09 10:39

sachin.setiya.007 I only asked because there are the gsoc labels. I guess that they can be removed, no?

Ralf Gebhardt added a comment - 2019-08-09 10:39 sachin.setiya.007 I only asked because there are the gsoc labels. I guess that they can be removed, no?

Sergei Golubchik added a comment - 2019-08-12 11:11

No-no, they should not be removed. They mean that the task was offered for GSoC 2018 and GSoC 2019.
It's a historical fact, that it was, so the labels should be there, indicating that fact.

Sergei Golubchik added a comment - 2019-08-12 11:11 No-no, they should not be removed. They mean that the task was offered for GSoC 2018 and GSoC 2019. It's a historical fact, that it was, so the labels should be there, indicating that fact.

Sachin Setiya (Inactive) added a comment - 2020-02-06 06:29

So currently I am working on issue , when we write START ALTER gtid_slave_pos table is not updated, Which is normal because thread is hanging/waiting in
mysql_alter_table , So we cant call do_update_pos (we are still inside QUery_log_event::do_apply_event)
So I am thinking of calling ::stmt_done inside mysql_alter_table(after writing START ALTER into binlog), I have stored the log_pos in thd (debatable )
So the issue which I am facing is , we call lock_tables in mysql_alter_table (normal table , which we want to change), and after that writing START ALTER into binlog we will write into gtid_slave_pos , but when ::stmt_done calls lock_tables its assert is failing because thd->lock is not NULL

So I am thinking of something like lock gtid_slave_pos when I acquire lock for normal table and release the lock when it is updated , NOt sure if it is possible

Sachin Setiya (Inactive) added a comment - 2020-02-06 06:29 So currently I am working on issue , when we write START ALTER gtid_slave_pos table is not updated, Which is normal because thread is hanging/waiting in mysql_alter_table , So we cant call do_update_pos (we are still inside QUery_log_event::do_apply_event) So I am thinking of calling ::stmt_done inside mysql_alter_table(after writing START ALTER into binlog), I have stored the log_pos in thd (debatable ) So the issue which I am facing is , we call lock_tables in mysql_alter_table (normal table , which we want to change), and after that writing START ALTER into binlog we will write into gtid_slave_pos , but when ::stmt_done calls lock_tables its assert is failing because thd->lock is not NULL So I am thinking of something like lock gtid_slave_pos when I acquire lock for normal table and release the lock when it is updated , NOt sure if it is possible

Andrei Elkin added a comment - 2021-12-13 22:01

The about to be created preview branch misses out FTWRL fixes that we've discussed today
and it's just too late night for me to complete with them into the branch.
This will be done tomorrow.

Andrei Elkin added a comment - 2021-12-13 22:01 The about to be created preview branch misses out FTWRL fixes that we've discussed today and it's just too late night for me to complete with them into the branch. This will be done tomorrow.

Brandon Nesterenko added a comment - 2021-12-14 15:30

I approve the current patch c592a5e. Note that it will still need the FTWRL fixes.

Brandon Nesterenko added a comment - 2021-12-14 15:30 I approve the current patch c592a5e . Note that it will still need the FTWRL fixes.

Andrei Elkin added a comment - 2021-12-14 20:05

Approving on the behalf of bnestere.

Andrei Elkin added a comment - 2021-12-14 20:05 Approving on the behalf of bnestere .

Brandon Nesterenko added a comment - 2021-12-14 21:24

Confirmed Andrei's approval on my behalf, I had to step away momentarily

Brandon Nesterenko added a comment - 2021-12-14 21:24 Confirmed Andrei's approval on my behalf, I had to step away momentarily

Alice Sherepa added a comment - 2022-01-27 17:21

I think it's fine to push the feature into 10.8.

Alice Sherepa added a comment - 2022-01-27 17:21 I think it's fine to push the feature into 10.8.

People

Assignee:: Andrei Elkin

Reporter:: Sachin Setiya (Inactive)

Votes:: 4 Vote for this issue

Watchers:: 20 Start watching this issue

Dates

Created:: 2016-12-28 08:24

Updated:: 2025-03-11 17:16

Resolved:: 2022-01-31 09:56

Git Integration

Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.

MariaDB Server