I ended up just nuking the index on the replica with less docs and
restarting it - which triggered the snap pull from the leader. So now
I'm in sync and have better processes in place to capture the
information if it happens again, which given some of the queries my UI
team develops, is highly likely ;-)
Also, all our input data to Solr lives in Hive so I'm doing some id
-to- id comparisons of what is in Solr vs. what is in Hive to find any
discrepancies.
Again, sorry about the loss of the logs. This is a tough scenario to
try to re-create as it was a perfect storm of high indexing throughput
and a rogue query.
Tim
On Mon, Apr 22, 2013 at 2:41 PM, Mark Miller <markrmil...@gmail.com> wrote:
> What do you know about the # of docs you *should*? Do you have that mean when
> taking the bad replica out of the equation?
>
> - Mark
>
> On Apr 22, 2013, at 4:33 PM, Mark Miller <markrmil...@gmail.com> wrote:
>
>> Bummer on the log loss :(
>>
>> Good info though. Somehow that replica became active without actually
>> syncing? This is heavily tested (though not with OOM's I suppose), so I'm a
>> little surprised, but it's hard to speculate how it happened without the
>> logs. Specially, the logs from the node that is off would be great - we
>> would see what it did when it recovered and why it might think it was in
>> sync :(
>>
>> - Mark
>>
>> On Apr 22, 2013, at 2:19 PM, Timothy Potter <thelabd...@gmail.com> wrote:
>>
>>> nm - can't read my own output - the leader had more docs than the replica
>>> ;-)
>>>
>>> On Mon, Apr 22, 2013 at 11:42 AM, Timothy Potter <thelabd...@gmail.com>
>>> wrote:
>>>> Have a little more info about this ... the numDocs for *:* fluctuates
>>>> between two values (difference of 324 docs) depending on which nodes I
>>>> hit (distrib=true)
>>>>
>>>> 589,674,416
>>>> 589,674,092
>>>>
>>>> Using distrib=false, I found 1 shard with a mis-match:
>>>>
>>>> shard15: {
>>>> leader = 32,765,254
>>>> replica = 32,764,930 diff:324
>>>> }
>>>>
>>>> Interesting that the replica has more docs than the leader.
>>>>
>>>> Unfortunately, due to some bad log management scripting on my part,
>>>> the logs were lost when these instances got re-started, which really
>>>> bums me out :-(
>>>>
>>>> For now, I'm going to assume the replica with more docs is the one I
>>>> want to keep and will replicate the full index over to the other one.
>>>> Sorry about losing the logs :-(
>>>>
>>>> Tim
>>>>
>>>>
>>>>
>>>>
>>>> On Sat, Apr 20, 2013 at 10:23 AM, Timothy Potter <thelabd...@gmail.com>
>>>> wrote:
>>>>> Thanks for responding Mark. I'll collect the information you asked
>>>>> about and open a JIRA once I have a little more understanding of what
>>>>> happened. Hopefully I can piece together some story after going over
>>>>> the logs.
>>>>>
>>>>> As for replica / leader, I suspect some leaders went down but
>>>>> fail-over to new leaders seemed to work fine. We lost about 9 nodes at
>>>>> once and continued to serve queries, which is awesome.
>>>>>
>>>>> On Sat, Apr 20, 2013 at 10:11 AM, Mark Miller <markrmil...@gmail.com>
>>>>> wrote:
>>>>>> Yeah, thats no good.
>>>>>>
>>>>>> You might hit each node with distrib=false to get the doc counts.
>>>>>>
>>>>>> Which ones have what you think are the right counts and which the wrong
>>>>>> - eg is it all replicas that are off, or leaders as well?
>>>>>>
>>>>>> You say several replicas - do you mean no leaders went down?
>>>>>>
>>>>>> You might look closer at the logs for a node that has it's count off.
>>>>>>
>>>>>> Finally, I guess I'd try and track it in a JIRA issue.
>>>>>>
>>>>>> - Mark
>>>>>>
>>>>>> On Apr 19, 2013, at 6:37 PM, Timothy Potter <thelabd...@gmail.com> wrote:
>>>>>>
>>>>>>> We had a rogue query take out several replicas in a large 4.2.0 cluster
>>>>>>> today, due to OOM's (we use the JVM args to kill the process on OOM).
>>>>>>>
>>>>>>> After recovering, when I execute the match all docs query (*:*), I get a
>>>>>>> different count each time.
>>>>>>>
>>>>>>> In other words, if I execute q=*:* several times in a row, then I get a
>>>>>>> different count back for numDocs.
>>>>>>>
>>>>>>> This was not the case prior to the failure as that is one thing we
>>>>>>> monitor
>>>>>>> for.
>>>>>>>
>>>>>>> I think I should be worried ... any ideas on how to troubleshoot this?
>>>>>>> One
>>>>>>> thing to mention is that several of my replicas had to do full
>>>>>>> recoveries
>>>>>>> from the leader when they came back online. Indexing was happening when
>>>>>>> the
>>>>>>> replicas failed.
>>>>>>>
>>>>>>> Thanks.
>>>>>>> Tim
>>>>>>
>>
>
