Good point. I will see if I can get the necessary access rights on this
machine to run tcpdump.
Thanks for the suggestion,
Harald.
On 09.07.2014 00:32, Steve McKay wrote:
Sure sounds like a socket bug, doesn't it? I turn to tcpdump when Solr starts
behaving strangely in a socket-related way. Knowing exactly what's happening at
the transport level is worth a month of guessing and poking.
On Jul 8, 2014, at 3:53 AM, Harald Kirsch <harald.kir...@raytion.com> wrote:
Hi all,
This is what happens when I run a regular wget query to log the current number
of documents indexed:
2014-07-08:07:23:28 QTime=20 numFound="5720168"
2014-07-08:07:24:28 QTime=12 numFound="5721126"
2014-07-08:07:25:28 QTime=19 numFound="5721126"
2014-07-08:07:27:18 QTime=50071 numFound="5721126"
2014-07-08:07:29:08 QTime=50058 numFound="5724494"
2014-07-08:07:30:58 QTime=50033 numFound="5730710"
2014-07-08:07:31:58 QTime=13 numFound="5730710"
2014-07-08:07:33:48 QTime=50065 numFound="5734069"
2014-07-08:07:34:48 QTime=16 numFound="5737742"
2014-07-08:07:36:38 QTime=50037 numFound="5737742"
2014-07-08:07:37:38 QTime=12 numFound="5738190"
2014-07-08:07:38:38 QTime=23 numFound="5741208"
2014-07-08:07:40:29 QTime=50034 numFound="5742067"
2014-07-08:07:41:29 QTime=12 numFound="5742067"
2014-07-08:07:42:29 QTime=17 numFound="5742067"
2014-07-08:07:43:29 QTime=20 numFound="5745497"
2014-07-08:07:44:29 QTime=13 numFound="5745981"
2014-07-08:07:45:29 QTime=23 numFound="5746420"
As you can see, the QTime is just over 50 seconds at irregular intervals.
This happens independent of whether I am indexing documents with around 20 dps
or not. First I thought about a dependence on the auto-commit of 5 minutes, but
the the 50 seconds hits are too irregular.
Furthermore, and this is *really strange*: when hooking strace on the solr
process, the 50 seconds QTimes disappear completely and consistently --- a real
Heisenbug.
Nevertheless, strace shows that there is a socket timeout of 50 seconds defined
in calls like this:
[pid 1253] 09:09:37.857413 poll([{fd=96, events=POLLIN|POLLERR}], 1, 50000) = 1
([{fd=96, revents=POLLIN}]) <0.000040>
where the fd=96 is the result of
[pid 25446] 09:09:37.855235 accept(122, {sa_family=AF_INET, sin_port=htons(57236),
sin_addr=inet_addr("ip address of local host")}, [16]) = 96 <0.000054>
where again fd=122 is the TCP port on which solr was started.
My hunch is that this is communication between the cores of solr.
I tried to search the internet for such a strange connection between socket
timeouts and strace, but could not find anything (the stackoverflow entry from
yesterday is my own :-(
This smells a bit like a race condition/deadlock kind of thing which is broken
up by timing differences introduced by stracing the process.
Any hints appreciated.
For completeness, here is my setup:
- solr-4.8.1,
- cloud version running
- 10 shards on 10 cores in one instance
- hosted on SUSE Linux Enterprise Server 11 (x86_64), VERSION 11, PATCHLEVEL 2
- hosted on a vmware, 4 CPU cores, 16 GB RAM
- single digit million docs indexed, exact number does not matter
- zero query load
Harald.
