So this turns out to actually be an old bug.
It just happened to trigger more easily after patch for MDEV-8294.

The race is that there is a small window where SHOW GLOBAL STATUS may count
twice the values of a thread that is exiting. This happens also for normal
threads, not just the slave thread.

Here is a test case for latest 10.0, triggers with the following sleep
patched in:

diff --git a/sql/mysqld.cc b/sql/mysqld.cc

index e05c0b6..75af8f7 100644

--- a/sql/mysqld.cc

+++ b/sql/mysqld.cc

@@ -2759,6 +2759,7 @@ void unlink_thd(THD *thd)

   thd->add_status_to_global();

+my_sleep(2000000);

   mysql_mutex_lock(&LOCK_thread_count);

   thd->unlink();

   /*

CREATE TABLE t1 (a INT PRIMARY KEY);

connect (c1,localhost,root,,);

INSERT INTO t1 VALUES (1);

INSERT INTO t1 VALUES (2);

INSERT INTO t1 VALUES (3);

--let $cid= `SELECT CONNECTION_ID()`

--connection default

eval KILL CONNECTION $cid;

--sleep 1

SHOW GLOBAL STATUS LIKE 'Com_insert';

--sleep 2

SHOW GLOBAL STATUS LIKE 'Com_insert';

DROP TABLE t1;

The first SHOW GLOBAL STATUS counts the thread twice:

CREATE TABLE t1 (a INT PRIMARY KEY);

INSERT INTO t1 VALUES (1);

INSERT INTO t1 VALUES (2);

INSERT INTO t1 VALUES (3);

KILL CONNECTION 4;

SHOW GLOBAL STATUS LIKE 'Com_insert';

Variable_name	Value

Com_insert	6

SHOW GLOBAL STATUS LIKE 'Com_insert';

Variable_name	Value

Com_insert	3

DROP TABLE t1;

The race occurs because there is a window of time between a thread does THD::add_status_to_global() and THD::unlink(). During this window, calc_sum_of_all_status() can run and count the contribution of the thread twice.

The original buildbot test failure can be reproduced easily by injecting the
following sleep in the code:

diff --git a/sql/slave.cc b/sql/slave.cc

index 4129c4c..32e6bd4 100644

--- a/sql/slave.cc

+++ b/sql/slave.cc

@@ -4837,6 +4837,7 @@ log '%s' at position %s, relay log '%s' position: %s%s", RPL_LOG_NAME,

   mysql_cond_broadcast(&rli->stop_cond);

   DBUG_EXECUTE_IF("simulate_slave_delay_at_terminate_bug38694", sleep(5););

   mysql_mutex_unlock(&rli->run_lock);  // tell the world we are done

+my_sleep(500000);

   /*

     Deactivate the parallel replication thread pool, if there are now no more

Monty said he would fix this during this week

Monty, I've pushed a temporary work-around for the test failure in rpl.rpl_000011:

http://lists.askmonty.org/pipermail/commits/2015-June/008039.html

Please remove that work-around from the testcase when you push the fix for this underlying issue.

Fixed by adding a marker if we have added the thread counters to the global counters.
Fix will be pushed into 5.5 tree

Fix pushed into 5.5
When merged to 10.0, the above temporary fix should be removed and this bug closed.

Automated message:
----------------------------
Since this issue has not been updated since 6 weeks, it's time to move it back to Stalled.

This was completed long ago. T bug is fixed, as described in comments. And the temporary work-around was removed as part of commit 1e0f09cacbb05eec95078e10df1ea610736b9c1a.