[jira] [Created] (DAFFODIL-2637) Daffodil opening too many schema files

["Steve Lawrence (Jira)" <ji...@apache.org>]

Steve Lawrence created DAFFODIL-2637:
----------------------------------------

             Summary: Daffodil opening too many schema files
                 Key: DAFFODIL-2637
                 URL: https://issues.apache.org/jira/browse/DAFFODIL-2637
             Project: Daffodil
          Issue Type: Bug
            Reporter: Steve Lawrence


When running sbt test in a schema project (specifically P8) with about 100 schema files, each of which imports some common schema files (e.g. common.dfdl.xsd, strings.dfdl.xsd) for shared use, Daffodil is holding open hundreds of file descriptors to these same common files. In some cases, this can hit a resource limit and causes failures with "Too many open file" errors.

A good way to visualize this is start an sbt console:
{code}
sbt -mem 4096
{code}

Then in another terminal, find the process ID of the running sbt process:
{code}
pgrep -f sbt
{code}

Start this watch command to see the all the files that are open, including the count of duplicates, replacing $PID with the pid you just found:
{code}
watch -n 0 'readlink /proc/$PID/fd/* | sort | uniq -c | sort -n -r'
{code}

Then trigger the tests to run back in the sbt console:
{code}
sbt> test
{code}

As the tests are run, the top of the watch screen will show many open file descriptors to those common files, e.g.
{code}
    187 /path/to/common.dfdl.xsd
     45 /path/to/strings.dfdl.xsd
     ...
      4 anon_inode:inotify
      3 /dev/urandom
      3 /dev/random
      3 /dev/pts/0
      1 socket:[2233641]
      1 socket:[2233640]
{code}

In some cases, I've seen these common files get to over 1000 instances.

It's not clear where this is happening, it could be schema compilation, it should be somewhere in the TDML runner, maybe somewhere with validating infosets, etc. We need to track down what is opening all these duplicate files and see if it can be fixed so they can be shared. Or if that's difficult, maybe we can at least close these files earlier so that we don't have so many open file descriptors at once.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)