Chasing a Hung MySQL Transaction: InnoDB History Length Strikes Back

In this blog post, I’ll review how a hung MySQL transaction can cause the InnoDB history length to grow and negatively affect MySQL performance.

Recently I was helping a customer discover why SELECT queries were running slower and slower until the server restarts (which got things back to normal). It took some time to get from that symptom to a final diagnosis. Please follow me on the journey of chasing this strange MySQL behavior!


Changes in the query response time can mean tons of things. We can check everything from the query plan to the disk performance. However, fixing it with a restart is less common. After looking at “show engine innodb status”, I noticed some strange lines:

There was a total of 940 transactions like this.

Another insight was the InnoDB transaction history length graph from Percona Monitoring and Management (PMM):

Hung MySQL Transaction

History length of 6 million and growing clearly indicates a problem.

Problem localized

There have been a number of blog posts describing a similar problem: Peter stated in a blog post: “InnoDB transaction history often hides dangerous ‘debt’“. As the InnoDB transaction history grows, SELECTs need to scan more and more previous versions of the rows, and performance suffers. That explains the issue: SELECT queries get slower and slower until restart. Peter also filed this bug: Major regression having many row versions.

But why does the InnoDB transaction history start growing? There are 940 transactions in this state: ACTIVE 766132 sec. MySQL’s process list shows those transactions in “Sleep” state. It turns out that those transactions were “lost” or “hung”. As we can also see, each of those transactions holds two lock structures and one undo record, so they are not committed and not rolled-back. They are sitting there doing nothing. In this case, with the default isolation level REPEATABLE-READ, InnoDB can’t purge the undo records (transaction history) for other transactions until these “hung” transactions are finished.

The quick solution is simple: kill those connections and InnoDB will roll back those transactions and purge transaction history. After killing those 940 transactions, the graph looked like this:

Hung MySQL Transaction

However, several questions remain:

  1. What are the queries inside of this lost transaction? Where are they coming from? The problem is that neither MySQL’s process list nor InnoDB’s status shows the queries for this transaction, as it is not running those queries right now (the process list is a snapshot of what is happening inside MySQL right at this moment)
  2. Can we fix it so that the “hung” transactions don’t affect other SELECT queries and don’t cause the growth of transaction history?


As it turns out, it is very easy to simulate this issue with sysbench.

Test preparation

To add some load, I’m using sysbench,16 threads (you can open less, it does not really matter here) and a script for a “write-only” load (running for 120 seconds):

Simulate a “hung” transaction

While the above sysbench is running, open another connection to MySQL:

Note: we will need to run the SELECT as a part of this transaction. Do not close the connection.

Watch the history

We can see it is growing. Now it is time to commit or rollback or even kill our original transaction:

As we can see, it has purged the history length.

Finding the queries from the hung transactions

There are a number of options to find the queries from that “hung” transaction. In older MySQL versions, the only way is to enable the general log (or the slow query log). Starting with MySQL 5.6, we can use the Performance Schema. Here are the steps:

  1. Enable performance_schema if not enabled (it is disabled on RDS / Aurora by default).
  2. Enable events_statements_history:
  3. Run the query to find all transaction started 10 seconds ago (change the number of seconds to match your workload):

    Now we can see the list of queries from the old transaction (the MySQL query used was taken with modifications from this blog post: Tracking MySQL query history in long running transactions).

At this point, we can chase this issue at the application level and find out why this transaction was not committed. The typical causes:

  • There is a heavy, non-database-related process inside the application code. For example, the application starts a transaction to get a list of images for analysis and then starts an external application to process those images (machine learning or similar), which can take a very long time.
  • The application got an uncaught exception and exited, but the connection to MySQL was not closed for some reason (i.e., returned to the connection pool).

We can also try to configure the timeouts on MySQL or the application so that the connections are closed after “N” minutes.

Changing the transaction isolation level to fix the InnoDB transaction history issue

Now that we know which transaction is holding up the purge process of InnoDB history, we can find this transaction and make changes so it will not “hang”. We can change the transaction isolation level from REPEATABLE READ (default) to READ COMMITTED. In READ COMMITTED, InnoDB does not need to maintain history length when other transactions have committed changes. (More details about different isolation methods and how they affect InnoDB transactions.) That will work in MySQL 5.6 and later. However this doesn’t work in Amazon Aurora (as of now): even with READ COMMITTED isolation level, the history length still grows.

Here is the list of MySQL versions where changing the isolation level fixes the issue

MySQL Version Transaction isolation InnoDB History Length
MySQL 5.6repeatable readhistory is not purged until “hung” transaction finishes
MySQL 5.6read committed(fixed) history is purged
Aurorarepeatable readhistory is not purged until “hung” transaction finishes
Auroraread committedhistory is not purged until “hung” transaction finishes


Hung transactions can cause the InnoDB history length to grow and (surprisingly, on the first glance) affect the performance of other running select queries. We can use the performance schema to chase the “hung” transaction. Changing the MySQL transaction isolation level can potentially help.

Share this post

Comments (5)

  • Neo


    Great article, I have been there, luckily I was able to troubleshoot at that time by killing long running transaction.

    An observation for readers who might try to simulate this issue by various ways, the Repeatable read txn snapshot is initiated in the lifecycle of the transaction after the first SELECT SQL, i.e. a simple BEGIN; would not block Innodb from purging the history length.

    May 9, 2017 at 12:16 am
  • Pep Pla


    Really interesting. Should help to avoid this problems tuning wait_timeout/interactive_timeout? I assume the “hung” sessions were inactive and could be killed after a reasonable amount of time.

    Obviously this would not fix the root cause (aka hung sessions) but would stop them from creating more harm.


    May 9, 2017 at 5:46 pm
  • Akhil_MS


    That was really a great article.

    I feel there should be a mechanism to identify the dead connections (either from application nor from users) and kill them instantly. So that we can reduce the cost of the impact. what say?


    May 12, 2017 at 5:45 am
  • Adam


    As per my understanding,repeatable read is we can read that data which was available when this transaction started.

    Suppose another transaction started and updates that data ,but still you are selecting old version of data because repeatable read isolation. So if we commit this select transaction or kill , then n then it will purge old version data(innodb History for this transaction).

    If am correct or wrong please reply…..

    January 20, 2019 at 12:50 am
  • treerootprotectors

    Changes in the query response time can mean tons of things. We can check everything from the query plan to the disk performance. However, fixing it with a restart is less common. After looking at “show engine innodb status”, I noticed some strange lines:

    May 23, 2019 at 4:03 am

Comments are closed.

Use Percona's Technical Forum to ask any follow-up questions on this blog topic.