Solr 4 highlights - Mark Miller

Solr 4 Highlights
Mark Miller, LucidImagination.com
markrmiller@gmail.com

Friday, October 21, 11

What I Will Cover
 What are some cool features coming in Solr
4?
 Mark Miller - Lucene/Solr committer, PMC
member, Apache member, blah blah blah
 DirectSolrSpellChecker
 NRT
 Realtime Get
 SolrCloud - search side
 SolrCloud - indexing side (WIP)

4


DirectSolrSpellchecker

 New spellcheck implementation that works with
the new automaton code.
 A practical beneﬁt of this spellchecker is that it
requires no additional data structures (neither in
RAM nor on disk) to do its work.
 Also, no build phase - suggestion are always up
to date.

7


First Came...

LUCENE-1606, LUCENE-2089:

Adds AutomatonQuery, a MultiTermQuery
that matches terms against a finite-state
machine. Implement WildcardQuery and
FuzzyQuery with finite-state methods.
Adds RegexpQuery.

(Robert Muir, Mike McCandless, Uwe Schindler, Mark Miller)

4


Leading too...

LUCENE-2507, SOLR-2571, SOLR-2576:

Added DirectSolrSpellChecker, which
uses Lucene's DirectSpellChecker to
retrieve correction candidates directly
from the term dictionary using levenshtein
automata.

(James Dyer, rmuir)

5


Automaton

6


SolrConfig

<searchComponent name="spellcheck" class="solr.SpellCheckComponent">


<lst name="spellchecker">
<str name="name">default</str>
<str name="field">name</str>
<str name="classname">solr.DirectSolrSpellChecker</str>

7

Realtime-Get

SOLR-2656:

realtime-get, efficiently retrieves the latest
stored fields for specified documents,
even if they are not yet searchable (i.e.
without reopening a searcher)
(yonik)

8


Realtime Get
<updateLog class="solr.FSUpdateLog">
<str name="dir">${solr.data.dir:}</str>
</updateLog>


<requestHandler name="/get" class="solr.RealTimeGetHandler">
<lst name="defaults">
<str name="omitHeader">true</str>
</lst>
</requestHandler>

10

Realtime Get

curl 'http://localhost:8983/solr/update/json?commitWithin=10000000'
-H 'Content-type:application/json' -d '
[{
"id" : "mydoc",
"title" : "realtime-get test!"
}]'

http://localhost:8983/solr/get?ids=mydoc,mydoc

http://localhost:8983/solr/get?id=mydoc&id=mydoc

{"doc":{"id":"mydoc","title":["realtime-get test!"]}}

10


NRT
UpdateHandler Re-Architecture

 Fix commit waits for background merges
 Fix reloaded core does not share
IndexWriter
 Remove unnecessary locking
 Add NRT
 SoftCommit
 SoftAutoCommit

11


Trying to do NRT

 Commit - 1 second.
 Commit - 0.5 seconds.
 Commit - 1.5 minutes. BACKGROUND MERGE

12


Previously....

SolrCore Reload -> SolrCore

IW IW

Usually worked because IW is lazy instantiated...but
overlap is possible. What happens when
continuously feeding Solr and you do a core reload?
IW Lock timeout!

13


Solution

SolrCore Reload -> SolrCore

IW

Reloaded SolrCore now shares a State object

14


NRT AutoCommit

<updateHandler class="solr.DirectUpdateHandler2">

<autoCommit>
<maxTime>60000</maxTime>
</autoCommit>

<autoSoftCommit>
<maxTime>1000</maxTime>
</autoSoftCommit>

15


SolrCloud - Search Side

16


SolrCloud - Search Side

SOLR-1873:

SolrCloud - added shared/central config
and core/shard management via
zookeeper, built-in load balancing, and
infrastructure for future SolrCloud work.
(yonik, Mark Miller)

17


• /solr
ZooKeeper Layout
/solr/configs
/solr/configs/conf1
/solr/configs/conf1/mapping-ISOLatin1Accent.txt
DATA: ...supressed...
/solr/configs/conf1/protwords.txt
/solr/configs/conf1/old_synonyms.txt Centralized Configuration
/solr/configs/conf1/solrconfig.xml
/solr/configs/conf1/stopwords.txt
/solr/configs/conf1/schema.xml
Cluster Composition

/solr/collections/collection1/shards
/solr/collections/collection1/shards/shard4/localhost:50270_solr_
DATA:
node_name=localhost:50270_solr
url=http://localhost:50270/solr/

18


Improvements Coming...
 Modeling the cluster as ZooKeeper nodes is
cool and convenient. But reading the cluster is
costly - many requests.
 This can be mitigated with an incremental
update mode.
 But that makes it difficult to monitor for on the
fly property changes...too many ZooKeeper
Watches needed.
 Solution - store the cluster state as the data for
a single node.
 Downsides: 1MB max data size on zk node.
Possibly less convenient.
19


Ephemeral Nodes

• /solr/live_nodes
/solr/live_nodes/localhost:50272_solr

Nodes know which other nodes are down. There is
some lag - nodes also manage their own list of good
nodes to handle the lag time.

20


Getting Started
http://wiki.apache.org/solr/SolrCloud

21


SolrCloud - Index Side
Goals

 Distributed Indexing
 Fault Tolerant
 Elastic
 Durable

22


Current Vision
 You figure out you need 20 servers at least to
serve your index.
 You specify 20 shards to Solr and launch 60
servers with minimum configuration needed.
 This gives you 20 shards, each with 3 replicas.
 You begin indexing to any of the 60 servers and
documents are pushed around to where they
belong and are visible to search within seconds.
 If servers die, you don’t lose data (unless a full
shard goes down and the data on each server
is unrecoverable).

23


SolrCloud - Index Side
What’s Being worked on this very moment?

 Realtime Get
 Transaction Log
 Versions
 Leader per Shard
 Auto Assign Nodes to Shards
 Distributed IDF

24


Versions
 If two updates to a document are issues around
the same time, how do we know which
document should be the latest? More important,
how can we be sure the same document ‘wins’
on each replica.
 Lamport Clocks, Vector Clocks, a Leader per
shard?

25


Leader Per Shard
 Use ZooKeeper for Leader election.
 Leader can easily version documents.
 Leader is the true record - recovery.
 If the Leader goes down, a replica takes it’s
place.
 Updates briefly block or are buffered with
transaction log.

26


Auto Assign Shards

 Collection lock?
 Optimistic lock?
 Overseer?

27


Cluster

Shard1 Shard2 Shard3

Leaders

Replicas

28


Wrap Up
 These are only a few features - there are many
more already, and more coming.
 Pseudo Join, Pivot Faceting, Pseudo Fields,
Per Segment Faceting, Configure Codec Per
Filed, Use Alternate Scoring Algorithms...

 When? Your guess is as good as mine ;) I hope
sometime early next year.

11


Sources
 Links

• http://wiki.apache.org/solr/
SpellCheckComponent
• http://wiki.apache.org/solr/NearRealtimeSearch
• http://www.lucidimagination.com/blog/
2011/09/07/realtime-get/

13


Contact
 Mark Miller
• markrmiller@gmail.com /
mark.miller@lucidimagination.com
• http://www.lucidimagination.com/blog/author/
markmiller/
• http://twitter.com/heismark

14


Solr 4 highlights - Mark Miller

More Related Content

What's hot (18)

Similar to Solr 4 highlights - Mark Miller (20)

More from lucenerevolution (20)

Recently uploaded (20)

Solr 4 highlights - Mark Miller