Joe,
   We're running:

   - OpenJDK Java 1.8.0_242
   - NiFi 1.11.0
   - CentOS Linux 7.7.1908


   We're seeing this across a dozen NiFi's with the same setup.  To
reproduce the issue, Generate Flow Files 100GB across a couple million
files -> Site to Site -> Receive data -> Merge Content.  We had no issues
with this stack:

   - OpenJDK Java 1.8.0_232.
   - NiFi 1.9.2
   - CentOS Linux 7.7.1908

   Can your team setup a similar stack and test?

Ryan

On Thu, Feb 6, 2020 at 4:15 PM Joe Witt <[email protected]> wrote:

> received a direct reply - Elli cannot share.
>
> I think unless someone else is able to replicate the behavior there isn't
> much more we can tackle on this.
>
> Thanks
>
> On Thu, Feb 6, 2020 at 4:10 PM Joe Witt <[email protected]> wrote:
>
> > Yes Elli it is possible.  Can we please get those lsof outputs in a JIRA?
> > As well as more details about configuration?
> >
> > Thanks
> >
> > On Thu, Feb 6, 2020 at 2:44 PM Andy LoPresto <[email protected]>
> wrote:
> >
> >> I have no input on the specific issue you’re encountering, but a pattern
> >> we have seen to reduce the overhead of multiple remote input ports being
> >> required is to use a “central” remote input port and immediately follow
> it
> >> with a RouteOnAttribute to distribute specific flowfiles to the
> appropriate
> >> downstream flow / process group. Whatever sends data to this port can
> use
> >> an UpdateAttribute to add some “tracking/routing” attribute on the
> >> flowfiles before being sent. Inserting Merge/Split will likely affect
> your
> >> timing due to waiting for bins to fill, depending on your volume. S2S is
> >> pretty good at transmitting data on-demand with low overhead on one
> port;
> >> it’s when you have many remote input ports that there is substantial
> >> overhead.
> >>
> >>
> >> Andy LoPresto
> >> [email protected]
> >> [email protected]
> >> PGP Fingerprint: 70EC B3E5 98A6 5A3F D3C4  BACE 3C6E F65B 2F7D EF69
> >>
> >> > On Feb 6, 2020, at 2:34 PM, Elli Schwarz <[email protected]
> .INVALID>
> >> wrote:
> >> >
> >> > We ran that command - it appears the site-to-sites that are causing
> the
> >> issue. We had a lot of remote process groups that weren't even being
> used
> >> (no data was being sent to that part of the dataflow), yet when running
> the
> >> lsof command they each had a large number of open files - almost 2k! -
> >> showing CLOSE_WAIT. Again, there were no flowfiles being sent to them,
> so
> >> can it be some kind of bug that keeping a remote process group open is
> >> somehow opening files and not closing them? (BTW, the reason we had to
> >> upgrade from 1.9.2 to 1.11.0 was because we had upgraded our Java
> version
> >> and that cause an IllegalBlockingModeException - is it possible that
> >> whatever fixed that problem is now causing an issue with open files?)
> >> >
> >> > We now disabled all of the unused remote process groups. We still have
> >> several remote process groups that we are using so if this is the issue
> it
> >> might be difficult to avoid, but at least we decreased the number of
> remote
> >> process groups we have. Another approach we are trying is a merge
> content
> >> before we send to the Nifi having the most issues, to have fewer flow
> files
> >> sent at once site to site, and then splitting them after they are
> received.
> >> > Thank you!
> >> >
> >> >    On Thursday, February 6, 2020, 2:19:48 PM EST, Mike Thomsen <
> >> [email protected]> wrote:
> >> >
> >> > Can you share a description of your flows in terms of average flowfile
> >> size, queue size, data velocity, etc.?
> >> > Thanks,
> >> > Mike
> >> >
> >> > On Thu, Feb 6, 2020 at 1:59 PM Elli Schwarz <
> [email protected]>
> >> wrote:
> >> >
> >> >  We seem to be experiencing the same problems. We recently upgraded
> >> several of our Nifis from 1.9.2 to 1.11.0, and now many of them are
> failing
> >> with "too many open files". Nothing else changed other than the upgrade,
> >> and our data volume is the same as before. The only solution we've been
> >> able to come up with is to run a script to check for this condition and
> >> restart the Nifi. Any other ideas?
> >> > Thank you!
> >> >
> >> >     On Sunday, February 2, 2020, 9:11:34 AM EST, Mike Thomsen <
> >> [email protected]> wrote:
> >> >
> >> >  Without further details, this is what I did to see if it was
> something
> >> > other than the usual issue of having not enough file handlers
> available.
> >> > Something like a legitimate case of someone forgetting to close file
> >> > objects or something in the code itself.
> >> >
> >> > 1. Setup a 8core/32GB VM on AWS w/ Amazon AMI.
> >> > 2. Pushed 1.11.1RC1
> >> > 3. Pushed the RAM settings to 6/12GB
> >> > 4. Disabled flowfile archiving because I only allocated 8GB of
> storage.
> >> > 5. Setup a flow that used 2 generateflow instances to generate massive
> >> > amounts of garbage data using all available cores. (All queues were
> >> setup
> >> > to hold 250k flow files)
> >> > 6. Kicked it off and let it run for probably about 20 minutes.
> >> >
> >> > No apparent problem with closing and releasing resources here.
> >> >
> >> > On Sat, Feb 1, 2020 at 8:00 AM Joe Witt <[email protected]> wrote:
> >> >
> >> >> these are usually very easy to find.
> >> >>
> >> >> run lsof -p pid.  and share results
> >> >>
> >> >>
> >> >> thanks
> >> >>
> >> >> On Sat, Feb 1, 2020 at 7:56 AM Mike Thomsen <[email protected]>
> >> >> wrote:
> >> >>
> >> >>>
> >> >>>
> >> >>
> >>
> https://stackoverflow.com/questions/59991035/nifi-1-11-opening-more-than-50k-files/60017064#60017064
> >> >>>
> >> >>> No idea if this is valid or not. I asked for clarification to see if
> >> >> there
> >> >>> might be a specific processor or something that is triggering this.
> >> >>>
> >> >>
> >> >
> >>
> >>
>

Reply via email to