lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Erick Erickson <erickerick...@gmail.com>
Subject Re: distribution of leader and replica in SolrCloud
Date Wed, 10 May 2017 15:17:13 GMT
Bernd:

Short form: Worrying about which node is the leader is wasting your
time. Details below:

Why do you care what nodes the leaders are on? There has to be some
concern you have about co-locating the leaders on the same node or you
wouldn't be spending the time on it. Please articulate that concern.
This really sounds like an XY problem, you're asking about X (how to
insure leader location) when you're concerned about Y. What's the Y?

Worrying about placing leaders is really a waste of time at the scale
you're talking. Which replica is the leader changes anyway when nodes
go up and down or depending on the order you start your Solr
instances. There's no way to specify it with rule based node placement
during collection creation because it's not worth the effort.

You may be misunderstanding what the leader does. It's biggest
responsibility is that it forwards the raw documents to the other
replicas when indexing. The reason updates are sent to the leader is
to insure ordering (i.e. if the same doc is sent to two nodes in the
cloud for indexing by two different clients at the same time, one of
them has to "win" deterministically). But other than this trivial
check (and forwarding the docs of course, but that's almost entirely
I/O) the leader is just like any other replica in the shard.

The leader does _not_ index the document and the forward the results
to the followers, all replicas on a shard index each document
independently.

When querying, the leader has no additional duties at all, it's just
another replica serving queries.

So all the added duties of a leader amount to is forwarding the raw
docs to the followers, collecting the responses and returning the
status of the indexing request to the caller. During querying, the
leader may or may not participate.

I doubt you'll even be able to measure the increased load on a node
even if all the leaders are located on it. As far as cluster
robustness is concerned, again where the leaders are placed is
irrelevant. The only time I've seen any problems with all the leaders
being on a single physical node, there were several hundred shards.

Best,
Erick

On Wed, May 10, 2017 at 5:15 AM, Rick Leir <rleir@leirtech.com> wrote:
> Myself, I am still in the old camp. For critical machines, I want to know that it is
my machine, with my disks, and what software is installed exactly. But maybe the cloud provider's
fast network is more important? Cheers--Rick
>
> On May 10, 2017 6:13:27 AM EDT, Bernd Fehling <bernd.fehling@uni-bielefeld.de>
wrote:
>>Hi Rick,
>>
>>yes I have distributed 5 virtual server accross 5 physical machines.
>>So each virtual server is on a separate physical machine.
>>
>>Splitting each virtual server (64GB RAM) into two (32GB RAM), which
>>then
>>will be 10 virtual server accross 5 physical machines, is no option
>>because there is no gain against hardware failure of a physical
>>machine.
>>
>>So I rather go with two Solr instances per 64GB virtual server as first
>>try.
>>
>>Currently I'm still trying to solve the Rule-based Replica Placement.
>>There seams to be no way to report if a node is a "leader" or has the
>>"role="leader".
>>
>>Do you know how to create a rule like:
>>--> "do not create the replica on the same host where his leader
>>exists"
>>
>>Regards,
>>Bernd
>>
>>
>>Am 10.05.2017 um 10:54 schrieb Rick Leir:
>>> Bernd,
>>>
>>> Yes, cloud, ahhh. As you say, the world changed.  Do you have any
>>hint from the cloud provider as to which physical machine your virtual
>>server
>>> is on? If so, you can hopefully distribute your replicas across
>>physical machines. This is not just for reliability: in a sharded
>>system, each
>>> query will cause activity in several virtual servers and you would
>>prefer that they are on separate physical machines, not competing for
>>> resources. Maybe, for Solr, you should choose a provider which can
>>lease you the whole physical machine. You would prefer a 256G machine
>>over
>>> several shards on 64G virtual machines.
>>>
>>> And many cloud providers assume that servers are mostly idle, so they
>>cram too many server containers into a machine. Then, very
>>occasionally,
>>> you get OOM even though you did not exceed your advertised RAM. This
>>is a topic for some other forum, where should I look?
>>>
>>> With AWS you can choose to locate your virtual machine in
>>US-west-Oregon or US-east-i-forget or a few other locations, but that
>>is a very coarse
>>> division. Can you choose physical machine?
>>>
>>> With Google, it might be dynamic?
>>> cheers -- Rick
>>>
>>>
>>> On 2017-05-09 03:44 AM, Bernd Fehling wrote:
>>>> I would name your solution more a work around as any similar
>>solution of this kind.
>>>> The issue SOLR-6027 is now 3 years open and the world has changed.
>>>> Instead of racks full of blades where you had many dedicated bare
>>metal servers
>>>> you have now huge machines with 256GB RAM and many CPUs.
>>Virtualization has taken place.
>>>> To get under these conditions some independance from the physical
>>hardware you have
>>>> to spread the shards across several physical machines with virtual
>>servers.
>>>> >From my point of view it is a good solution to have 5 virtual 64GB
>>servers
>>>> on 5 different huge physical machines and start 2 instances on each
>>virtual server.
>>>> If I would split up each 64GB virtual server into two 32GB virtual
>>server there would
>>>> be no gain. We don't have 10 huge machines (no security win) and we
>>have to admin
>>>> and control 10 virtual servers instead of 5 (plus zookeeper
>>servers).
>>>>
>>>> It is state of the art that you don't have to care about the servers
>>within
>>>> the cloud. This is the main sense of a cloud.
>>>> The leader should always be aware who are the members of his cloud,
>>how to reach
>>>> them (IP address) and how are the users of the cloud (collections)
>>distributed
>>>> across the cloud.
>>>>
>>>> It would be great if a solution of issue SOLR-6027 would lead to
>>some kind of
>>>> "automatic mode" for server distribution, without any special
>>configuring.
>>>>
>>>> Regards,
>>>> Bernd
>>>>
>>>>
>>>> Am 08.05.2017 um 17:47 schrieb Erick Erickson:
>>>>> Also, you can specify custom placement rules, see:
>>>>>
>>https://cwiki.apache.org/confluence/display/solr/Rule-based+Replica+Placement
>>>>>
>>>>> But Shawn's statement is the nub of what you're seeing, by default
>>>>> multiple JVMs on the same physical machine are considered separate
>>>>> Solr instances.
>>>>>
>>>>> Also note that if you want to, you can specify a nodeSet when you
>>>>> create the nodes, and in particular the special value EMPTY.
>>That'll
>>>>> create a collection with no replicas and you can ADDREPLICA to
>>>>> precisely place each one if you require that level of control.
>>>>>
>>>>> Best,
>>>>> Erick
>>>>>
>>>>> On Mon, May 8, 2017 at 7:44 AM, Shawn Heisey <apache@elyograg.org>
>>wrote:
>>>>>> On 5/8/2017 5:38 AM, Bernd Fehling wrote:
>>>>>>> boss ------ shard1 ----- server2:7574
>>>>>>>         |             |-- server2:8983 (leader)
>>>>>> The reason that this happened is because you've got two nodes
>>running on
>>>>>> every server.  From SolrCloud's perspective, there are ten
>>distinct
>>>>>> nodes, not five.
>>>>>>
>>>>>> SolrCloud doesn't notice the fact that different nodes are running
>>on
>>>>>> the same server(s).  If your reaction to hearing this is that it
>>>>>> *should* notice, you're probably right, but in a typical use case,
>>each
>>>>>> server should only be running one Solr instance, so this would
>>never happen.
>>>>>>
>>>>>> There is only one instance where I can think of where I would
>>recommend
>>>>>> running multiple instances per server, and that is when the
>>required
>>>>>> heap size for a single instance would be VERY large.  Running two
>>>>>> instances with smaller heaps can yield better performance.
>>>>>>
>>>>>> See this issue:
>>>>>>
>>>>>> https://issues.apache.org/jira/browse/SOLR-6027
>>>>>>
>>>>>> Thanks,
>>>>>> Shawn
>>>>>>
>>>
>
> --
> Sorry for being brief. Alternate email is rickleir at yahoo dot com

Mime
View raw message