And.... it was SystemD...

On Fri, Apr 12, 2019 at 8:30 AM Mike Thomsen <[email protected]> wrote:

> When I do lsof -u nifi, it says the nifi user only has 5761 handles
> associated with it.
>
> One warning I saw on StackExchange said that sometimes SystemD subtly
> messes with this stuff on RHEL.
>
> On Fri, Apr 12, 2019 at 8:14 AM Mike Thomsen <[email protected]>
> wrote:
>
>> About 5600-5700 starting fresh. Got to about 6500-6800 before hitting the
>> ceiling.
>>
>> On Fri, Apr 12, 2019 at 7:30 AM Joe Witt <[email protected]> wrote:
>>
>>> mike
>>>
>>> lsof -p <pid>
>>>
>>> with the pid of the actual nifi process is probably better to look at
>>> for nifi resource handling observation.  what is that count.  yes the jars
>>> and such will all be loaded.  you can expect a few thousand off that.
>>>  then there are sockets and content and prov and flowfile....which adds a
>>> bit more.
>>>
>>> you should be able view the lsof input and get a pretty good idea of any
>>> unexpected file handles.
>>>
>>> thanks
>>>
>>> On Fri, Apr 12, 2019, 7:00 AM Mike Thomsen <[email protected]>
>>> wrote:
>>>
>>>> I know you can increase the file handle limit in
>>>> /etc/security/limits.conf, but we're having a really weird issue where a
>>>> CentOS 7.5 box can handle a massive record set just fine and another
>>>> running CentOS 7.6 cannot.
>>>>
>>>> When I run *lsof | wc -l* on the 7.6 box after NiFi has been running
>>>> for a while, it prints out hundreds of thousands to a million as the value.
>>>> Every jar, class file, etc. that is part of the work folder is listed as an
>>>> open file and the content report oddly enough has maybe 10k-15k files at
>>>> the most during the ingestion of the largest pieces. So a limit of say 500k
>>>> open file handles feels like it should be **plenty**.
>>>>
>>>> There's a known bug in some releases of CentOS that causes PAM to kill
>>>> a session if the file handle limit is higher than 1M or unlimited.
>>>>
>>>> Anyone have suggestions on what might be happening here?
>>>>
>>>> Thanks,
>>>>
>>>> Mike
>>>>
>>>

Reply via email to