Hello,
Same as Sam, I believe the SWAP command is important for important use
cases.
For example, with Solr 3, we do use Current and Temp cores, so that
incremental updates to the index are done live on Current, as well as
searches.
Whenever a full/baseline/from scratch index need to be performed,
incremental index is stopped, a delete *:* on Tempis performed, and all
documents are fed into Temp. Once done and commited, a SWAP between Temp
and Current is performed, and we can continue in nominal incremental
mode. All this time the searches on Current were working and mostly
correct, minus not seeing updates during the full index.
1) currently with SolrCloud, how is it possible to rebuild a new index
from scratch, without too much disruption of live searches?
It is realistic to stop any commit/softcommit, do the delete *:*, then
feed all documents, then only commit ? It seems the transaction log
would be way too large ? Our use case for each collection is around 10
million of documents, a dozen kB each.
2) maybe a workaround is possible using collections, ie. search on
collection1 while indexing collection2, then once done search on
collection2 ?
3) and in what ways could we help to implement a solution on SOLR-3866 ?
André
On 09/25/2012 03:17 AM, sam fang wrote:
Hi Mark,
If can support in future, I think it's great. It's a really useful feature.
For example, user can use to refresh with totally new core. User can build
index on one core. After build done, can swap old core and new core. Then
get totally new core for search.
Also can used in the backup. If one crashed, can easily swap with backup
core and quickly serve the search request.
Best Regards,
Sam
On Sun, Sep 23, 2012 at 2:51 PM, Mark Miller<markrmil...@gmail.com> wrote:
FYI swap is def not supported in SolrCloud right now - even though it may
work, it's not been thought about and there are no tests.
If you would like to see support, I'd add a JIRA issue along with any
pertinent info from this thread about what the behavior needs to be changed
to.
- Mark
On Sep 21, 2012, at 6:49 PM, sam fang<sam.f...@gmail.com> wrote:
Hi Chris,
Thanks for your help. Today I tried again and try to figure out the
reason.
1. set up an external zookeeper server.
2. change /opt/solr/apache-solr-4.0.0-BETA/example/solr/solr.xml
persistent
to true. and run below command to upload config to zk. (renamed multicore
to solr, and need to put zkcli.sh related jar package.)
/opt/solr/apache-solr-4.0.0-BETA/example/cloud-scripts/zkcli.sh -cmd
upconfig -confdir
/opt/solr/apache-solr-4.0.0-BETA/example/solr/core0/conf/
-confname
core0 -z localhost:2181
/opt/solr/apache-solr-4.0.0-BETA/example/cloud-scripts/zkcli.sh -cmd
upconfig -confdir
/opt/solr/apache-solr-4.0.0-BETA/example/solr/core1/conf/
-confname
core1 -z localhost:2181
3. Start jetty server
cd /opt/solr/apache-solr-4.0.0-BETA/example
java -DzkHost=localhost:2181 -jar start.jar
4. publish message to core0
/opt/solr/apache-solr-4.0.0-BETA/example/solr/exampledocs
cp ../../exampledocs/post.jar ./
java -Durl=http://localhost:8983/solr/core0/update -jar post.jar
ipod_video.xml
5. query to core0 and core1 is ok.
6. Click "swap" in the admin page, the query to core0 and core1 is
changing. Previous I saw sometimes returns 0 result. sometimes return 1
result. Today
seems core0 still return 1 result, core1 return 0 result.
7. Then click "reload" in the admin page, the query to core0 and core1.
Sometimes return 1 result, and sometimes return nothing. Also can see
the zk
configuration also changed.
8. Restart jetty server. If do the query, it's same as what I saw in
step 7.
9. Stop jetty server, then log into zkCli.sh, then run command "set
/clusterstate.json {}". then start jetty again. everything back to
normal,
that is what previous swap did in solr 3.6 or solr 4.0 w/o cloud.
From my observation, after swap, seems it put shard information into
actualShards, when user request to search, it will use all shard
information to do the
search. But user can't see zk update until click "reload" button in admin
page. When restart web server, this shard information eventually went to
zk, and
the search go to all shards.
I found there is a option "distrib", and used url like "
http://host1:18000/solr/core0/select?distrib=false&q=*%3A*&wt=xml";, then
only get the data on the
core0. Digged in the code (handleRequestBody method in SearchHandler
class,
seems it make sense)
I tried to stop tomcat server, then use command "set /clusterstate.json
{}"
to clean all cluster state, then use command "cloud-scripts/zkcli.sh -cmd
upconfig" to upload config to zk server, and start tomcat server. It
rebuild the right shard information in zk. then search function back to
normal like what
we saw in 3.6 or 4.0 w/o cloud.
Seems solr always add shard information into zk.
I tested cloud swap on single machine, if each core have one shard in the
zk, after swap, eventually zk has 2 slices(shards) for that core because
now only
do the add. so the search will go to both 2 shards.
and tested cloud swap with 2 machine which each core have 1 shard and 2
slices. Below the configuration in the zk. After swap, eventually zk has
4
for that
core. and search will mess up.
"core0":{"shard1":{
"host1:18000_solr_core0":{
"shard":"shard1",
"roles":null,
"leader":"true",
"state":"active",
"core":"core0",
"collection":"core0",
"node_name":"host1:18000_solr",
"base_url":"http://host1:18000/solr"},
"host2:18000_solr_core0":{
"shard":"shard1",
"roles":null,
"state":"active",
"core":"core0",
"collection":"core0",
"node_name":"host2:18000_solr",
"base_url":"http://host2:18000/solr"}}},
For previous 2 cases, if I stoped tomcat/jetty server, then manullay
upload
configuration to zk, then start tomcat server, zk and search become
normal.
On Fri, Sep 21, 2012 at 3:34 PM, Chris Hostetter
<hossman_luc...@fucit.org>wrote:
: Below is my solr.xml configuration, and already set persistent to
true.
...
: Then publish 1 record to test1, and query. it's ok now.
Ok, first off -- please provide more details on how exactly you are
running Solr. Your initial email said...
In Solr 3.6, core swap function works good. After switch to use Solr
4.0
Beta, and found it doesn't work well.
...but based on your solr.xml file and your logs, it appears you are now
trying to use some of the ZooKeeper/SolrCloud features that didn't even
exist in Solr 3.6, so it's kind of an apples to oranges comparison. i'm
pretty sure that for a simple multicore setup, SWAP still works exactly
as
it did in Solr 3.6.
Wether SWAP works with ZooKeeper/SolrCloud is something i'm not really
clear on -- mainly because i'm not sure what it should mean
conceptually.
Should the two SolrCores swap which collections they are apart of? what
happens if the doc->shard assignment for the two collections means the
same docs woulnd't wind up im those SolrCores? what if the SolrCores are
two different shards of the same collection befor teh SWAP?
FWIW: It wasn't clear from your messsage *how* you had your SolrCloud
system setup, but it appears from your pre-swap log messages that you
are
running a single node, so I tried to reproduce the behavior you were
seeing by seting up a solr home dir like you described, and then
running...
java -Dsolr.solr.home=swap-test/ -DzkRun -Dbootstrap_conf=true
-DnumShards=1 -jar start.jar
...because that was my best guess as to what you were running. But even
then i couldn't get the behavior you described after the swap...
: And found the shardurl is different with the log which search before
swap.
: It’s shard.url=host1:18000/solr/test1-ondeck/|
host1:18000/solr/test1/.
...what i observed after the swap, is that it apperaed as if hte SWAP
had
no effect to client requests, because a client request to
"/solr/test1/select?q=*:*" was distributed under the covers to
"/solr/test1-ondeck/..." which is the new name for the core where the
doc
had been indexed...
Sep 21, 2012 12:03:32 PM org.apache.solr.core.SolrCore execute
INFO: [test1-ondeck] webapp=/solr path=/select
params={distrib=false&wt=javabin&rows=10&version=2&df=text&fl=id,score&shard.url=frisbee:8983/solr/test1-ondeck/&NOW=1348254212677&start=0&q=*:*&isShard=true&fsv=true}
hits=1 status=0 QTime=0
Sep 21, 2012 12:03:32 PM org.apache.solr.core.SolrCore execute
INFO: [test1-ondeck] webapp=/solr path=/select
params={df=text&shard.url=frisbee:8983/solr/test1-ondeck/&NOW=1348254212677&q=*:*&ids=SOLR1000&distrib=false&isShard=true&wt=javabin&rows=10&version=2}
status=0 QTime=0
Sep 21, 2012 12:03:32 PM org.apache.solr.core.SolrCore execute
INFO: [test1] webapp=/solr path=/select params={q=*:*} status=0 QTime=10
...I'm guessing this behavior is because nothing in the SWAP call
bothered
to tell ZK that these two SolrCores swaped names, so when asking to
query
the "test1" collection, ZK says "ok, well the name of the core that's
part
of that collection is test1-ondeck" and that's where the query was
routed.
I only saw behavior similar to what you described (of the "shard.url"
refering to both SolrCores) after i restarted solr -- presumably because
as the solrCores started up, they both notified ZK about their
existence,
and said what collection they (thought they) are a part of, but ZK also
already thinks they are each a part of a differet collection as well
(because nothing bothered to tell ZK otherwise).
So the long and short of it seems to be...
* CoreAdminHandler's SWAP is poorly defined if you are using SolrCloud
(most likely: so is RENAME and ALIAS) - i've opened SOLR-3866.
* This doesn't seem like a regression bug from Solr 3.6, because as
far as i can tell SWAP still works as well as it did in 3.6.
-Hoss
--
André Bois-Crettez
Search technology, Kelkoo
http://www.kelkoo.com/
Kelkoo SAS
Société par Actions Simplifiée
Au capital de € 4.168.964,30
Siège social : 8, rue du Sentier 75002 Paris
425 093 069 RCS Paris
Ce message et les pièces jointes sont confidentiels et établis à l'attention
exclusive de leurs destinataires. Si vous n'êtes pas le destinataire de ce
message, merci de le détruire et d'en avertir l'expéditeur.
