The context is the following:

• 100 concurrent content creators
• eZ Publish 3.10 in cluster mode
• 77 siteaccesses all related between them
• A very large amount of view cache files in database

The problem was the following:
When someone publishes a content, some deadlocks arised and some view cache files were not (or not correctly) updated.

The problem is that the function _deleteByDirListInner in ./kernel/classes/clusterfilehandlers/dbbackends/mysql.dicod.php
uses an UPDATE query with a LIKE clause which locks too much rows in the same time.

The solution is to select all the rows and the update them one by one.
This generate a high amount of SQL queries but avoid deadlocks.
The function _deleteByDirListInner is attached to this issue.

I do not provide a diff file since I worked on a custom cluster an a diff file will not work on the standard 3.10 cluster.