Indeed, parallel replication can make STOP SLAVE much longer if sizable transactions are running. The test case below demonstrates that. It updates many rows on master in a table without PK, so that RBR is really slow, then waits till the updates start running on the slave, executes STOP SLAVE, waits till it's finished and checks what's happened to the contents of the table on slave.

If the test is run without parallel replication, STOP SLAVE finishes very fast, and the contents of the table remains unchanged – that is, updates on the slave are interrupted and not committed (rolled back).
If the test is run with parallel replication, even if it's slave-parallel-threads=1, STOP SLAVE takes a long time, and at the end the contents of the table is updated – that is, the slave executes the transaction to the end and commits it before stopping.

I expect it to be a design choice of parallel replication, hopefully Elkin will check it and confirm (or not).

Note: the test case below is for reproducing only, do not put it into the regression suite!

# Optionally run with --mysqld=--slave-parallel-threads=N

--source include/have_innodb.inc

--source include/master-slave.inc

--source include/have_binlog_format_row.inc

create table t1 (i int, c varchar(255)) engine=InnoDB;

begin;

insert into t1 values (1,'foo'),(2,'bar'),(3,'abc'),(4,'def');

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

insert into t1 select * from t1;

commit;

--sync_slave_with_master

--connection master

begin;

update t1 set c = 'foobar';

update t1 set c = 'qux';

commit;

--connection slave

--let $show_statement= SHOW PROCESSLIST

--let $field= State

--let $condition= like 'Update_rows_log_event::find_row%'

--source include/wait_show_condition.inc

send stop slave;

--connection slave1

while (`SELECT COUNT(*) FROM information_schema.processlist where state like '%Killing slave%'`)

{

  select user,command,time,state,info from information_schema.processlist where command like 'slave%' or state like '%slave%';

  --sleep 5

}

select distinct(c) from t1;

# Cleanup

--connection slave

--reap

drop table t1;

--connection master

drop table t1;

I ran the test case above against 10.8.2-debug (commit 12cd3dc78d2a58a15377000a7a8adb92d4fa74fb) and could see the same 10x increase in time with

--mysqld=--slave-parallel-threads=1

as with 10.2.43-debug (commit 941bc7053616d5ca8c9e6538828c4a65802e8c3d).

Hi woosang, sure. It won't be done in time for the release this week, but we are planning on having it done in the next release (planned for June 8). Do they need a custom build to have that sooner?

Hi Andrei!

This is ready for review: PR-2534.

bnestere

The customer is asking if we have tested this fix with the same status of master as they encountered at Feb 2nd 13:33:07.
They need it to acknowledge that the fix will work in case of they encounter the same issue.

IMHO, these are available options.
1. Perform tests with the same status of master as they encountered
2. Provide additional message helps the customer acknowledge that this fix will work for them. (e.g. explain that the fix works regardless of the status of master)
3. Provide the status of master where you tested for this fix.

Please refer to the SR (CS0537360)
When the maxscale tried to perform automatic failover, the master was starting-up after threw signal-6 at that time.
maxcale log entries (attachment : maxscale.log_mbhapp01) :
2023-02-07 13:33:07 notice : [mariadbmon] Performing automatic failover to replace failed master 'server1'.
master log entries (attachment : maria.err_mbsdbp01):
2023-02-07 13:33:13 0 [Note] InnoDB: Read redo log up to LSN=3516540925952

woosangcc

Hi chanjong.yu!

I'm still working on the fix, but generally speaking, it should work with the version of their master, as it is change to the replica-side only.

Adjustments for the first round of review have been made PR 2534

The overall patch looks pretty good.
Thanks for creative and tedious work done!

Pushed into 10.4 as 0a99d457b

Merge conflicts observed through manual cherry picking in 10.5 and 10.9, with fixes in branches bb-10.5-MDEV-13915-mergefix and bb-10.9-MDEV-13915-mergefix, respectively

Please don't do this. And certainly not in a GA release!
STOP SLAVE isn't expected to roll back any transactions, it completes what has started and then stops.
This has always been so, it's the same in non-parallel replication.
If users want to stop the slave faster due to a long-running query, they can kill the specific query.

Edited: Ok, apparently we do that for non-parallel, but only when reading the next event.
In any case, you cannot do this in a GA release. How will 10.4 users get security fixes without also getting this change?

knielsen, salve. First my apologies that you were not explicitly notified on the patch for reviewing. It's not terminally late now I hope, provided you have time.

It turns out 10.4 push was our mistake in the version that the support case needs. We then are free to play safe and revert it from 10.4.

To the technical part of the patch, yes its idea is to make the sequential and parallel STOP:s work consistently,
including concern of non-transactional "transactions" that must complete to "commit".
To my reasoning there's nothing that the user of the patch can find unusual (apart from quick stopping perhaps).
Indeed, the user cannot see what has been scheduled to workers. So any quicker stop may happen without the patch.
In other words, the patch emulates slow scheduling.

The patch is being reverted in 10.4 and rebased now on the top 10.5.
I hope you'll find that acceptable at least for reviewing and your risk assessment.

bnestere and myself can discuss it interactively on zulip if you'll need that.
We found in our team github review is somewhat useful, so feel free to comment in detail on the patch over there.
A new PR that is being arranged to be published soon.

Cheers,
Andrei

Re-opening to address additional findings

Something that is missing from the discussion here is that the main reason STOP SLAVE is slow in parallel replication is not because it doesn't roll back running transactions. The main problem is that in many cases parallel replication will replicate all queued events (@@slave_parallel_max_queued).

I think this is a left-over of when only conservative mode existed. The current STOP SLAVE mechanism is seen in do_gco_wait(), it continues until the current GCO is completed (wait_count > entry->stop_count). But in optimistic mode, the GCO can be very large, potentially all queued events, thus stop is delayed longer than needed.

I think a much simpler solution is to fix this, so that stop_count is initialised to largest_started_sub_id, and compared against rgi->gtid_sub_id.

This will not rollback an existing long-running transaction, but I think that's actually good. Forcing stop immediately will cause massive rollback when many threads are configured (Jean Francois Gagné tested using > 1000 threads), which seems undesirable. And forcing stop does not guarantee fast stop anyway, a long-running statement will not be aborted.

I implemented a much simpler fix for this in the branch knielsen_faster_stop_slave:

https://github.com/MariaDB/server/commits/knielsen_faster_stop_slave

And a better fix for the MDEV-31448 fix included in Brandon's patch for this issue:

https://github.com/MariaDB/server/commits/knielsen_faster_stop_slave
https://github.com/MariaDB/server/commit/6ce9c839997c1fc78c2989540dd21155a96fb419

I now pushed another patch to the branch knielsen_faster_stop_slave:

https://github.com/MariaDB/server/commits/knielsen_faster_stop_slave
https://github.com/MariaDB/server/commit/c20ce0dff404890df59f1fd305bd85be0ada86f8

This implements a STOP SLAVE FORCE option, which can be used to optionally force a quick slave stop by rolling back all active transactions on the next event or row operation, if no non-transactional event groups are blocking it. I included the test cases from Brandon's pull request, they are passing with this patch (replacing STOP SLAVE with STOP SLAVE FORCE).

I think this is the way to go if we want this rollback functionality, leaving the option to not roll back a lot of work needlessly. I also think this is a new feature and should only go to development version (11.1?). Comments welcome, of course.

The simple fix earlier on my branch (without forcing rollback) should be sufficient to solve the user/customers problem. I think that can go in 10.5 if we want (or even 10.4, it should be safe). Some of the other bug fixes on the branch related to stop may also be appropriate for some earlier branch, suggestions wellcome.

I have pushed the fix for slow STOP SLAVE to 10.4 (The STOP SLAVE FORCE feature can go to a development branch later if we wish).
So I suggest to close this bug?

- Kristian.