lucene-solr-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Erick Erickson <>
Subject Re: SolrCloud replicas consistently out of sync
Date Mon, 16 May 2016 16:02:26 GMT
8 nodes, 4 shards apiece? All in the same JVM? People have gotten by
the GC pain by running in separate JVMs with less Java memory each on
big beefy machines.... That's not a recommendation as much as an

That aside, unless you have some very strange stuff going on this is
totally weird. Are you hitting OOM errors at any time you have this
problem? Once you hit an OOM error, all bets are off about how Java
behaves. If you are hitting those, you can't hope for stability until
you fix that issue. In your writeup there's some evidence for this
when you say that if you index multiple docs at a time you get

Do your Solr logs show any anomalies? My guess is that you'll see
exceptions in your Solr logs that will shed light on the issue.


On Mon, May 16, 2016 at 8:03 AM, Stephen Weiss <> wrote:
> Hi everyone,
> I'm running into a problem with SolrCloud replicas and thought I would ask the list to
see if anyone else has seen this / gotten past it.
> Right now, we are running with only one replica per shard.  This is obviously a problem
because if one node goes down anywhere, the whole collection goes offline, and due to garbage
collection issues, this happens about once or twice a week, causing a great deal of instability.
 If we try to increase to 2 replicas per shard, once we index new documents and the shards
autocommit, the shards all get out of sync with each other, with different numbers of documents,
different numbers of documents deleted, different facet counts - pretty much totally divergent
indexes.  Shards always show green and available, and never go into recovery or any other
state as to indicate there's a mismatch.  There are also no errors in the logs to indicate
anything is going wrong.  Even long after indexing has finished, the replicas never come back
into sync.  The only way to get consistency again is to delete one set of replicas and then
add them back in.  Unfortunately, when we do this, we invariably discover that many documents
(2-3%) are missing from the index.
> We have tried setting the min_rf parameter, and have found that when setting min_rf=2,
we almost never get back rf=2.  We almost always get rf=1, resend the request, and it basically
just goes into an infinite loop.  The only way to get rf=2 to come back is to only index one
document at a time.  Unfortunately, we have to update millions of documents a day and it isn't
really feasible to index this way, and even when indexing one document at a time, we still
occasionally find ourselves in an infinite loop.  This doesn't appear to be related to the
documents we are indexing - if we stop the index process and bounce solr, the exact same document
will go through fine the next time until indexing stops up on another random document.
> We have 8 nodes, with 4 shards a piece, all running one collection with about 900M documents.
 An important note is that we have a block join system with 3 tiers of documents (products
-> skus -> sku_history).  During indexing, we are forced to delete all documents for
a product prior to adding the product back into the index, in order to avoid orphaned children
/ grandchildren.  All documents are consistently indexed with the top-level product ID so
that we can delete all child/grandchild documents prior to updating the document.  So, for
each updated document, we are sending through a delete call followed by an add call.  We have
tried putting both the delete and add in the same update request with the same results.
> All we see out there on Google is that none of what we're seeing should be happening.
> We are currently running Solr 6.0 with Zookeeper 3.4.6.  We experienced the same behavior
on 5.4 as well.
> --
> Steve
> ________________________________
> WGSN is a global foresight business. Our experts provide deep insight and analysis of
consumer, fashion and design trends. We inspire our clients to plan and trade their range
with unparalleled confidence and accuracy. Together, we Create Tomorrow.
> WGSN<> is part of WGSN Limited, comprising of market-leading
products including<>, WGSN Lifestyle & Interiors<>,
WGSN INstock<>, WGSN StyleTrial<>
and WGSN Mindset<>, our bespoke consultancy
> The information in or attached to this email is confidential and may be legally privileged.
If you are not the intended recipient of this message, any use, disclosure, copying, distribution
or any action taken in reliance on it is prohibited and may be unlawful. If you have received
this message in error, please notify the sender immediately by return email and delete this
message and any copies from your computer and network. WGSN does not warrant that this email
and any attachments are free from viruses and accepts no liability for any loss resulting
from infected email transmissions.
> WGSN reserves the right to monitor all email through its networks. Any views expressed
may be those of the originator and not necessarily of WGSN. WGSN is powered by Ascential plc<>,
which transforms knowledge businesses to deliver exceptional performance.
> Please be advised all phone calls may be recorded for training and quality purposes and
by accepting and/or making calls from and/or to us you acknowledge and agree to calls being
> WGSN Limited, Company number 4858491
> registered address:
> Ascential plc, The Prow, 1 Wilder Walk, London W1B 5AP
> WGSN Inc., tax ID 04-3851246, registered office c/o National Registered Agents, Inc.,
160 Greentree Drive, Suite 101, Dover DE 19904, United States
> 4C Serviços de Informação Ltda., CNPJ/MF (Taxpayer's Register): 15.536.968/0001-04,
Address: Avenida Cidade Jardim, 377, 7˚ andar CEP 01453-000, Itaim Bibi, São Paulo
> 4C Business Information Consulting (Shanghai) Co., Ltd, 富新商务信息咨询(上海)有限公司,
registered address Unit 4810/4811, 48/F Tower 1, Grand Gateway, 1 Hong Qiao Road, Xuhui District,

View raw message