[
https://issues.apache.org/jira/browse/SOLR-15585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17401365#comment-17401365
]
Ishan Chattopadhyaya commented on SOLR-15585:
---------------------------------------------
{quote}In this scenario, all TLOG replicas are offline correct? Otherwise
another TLOG replica will claim leadership, and the PULL replicas will fetch
the up-to-date index from them.
{quote}
Correct. Imagine a situation were there's 1 TLOG replica + 1 PULL replica.
A "graceful" shutdown of a server should never result in data loss (writes were
acknowledged, but never reflect in the system in steady state), and hence the
need for such checks before the shutdown is deemed successful.
* Add shutdown behavior to check that you are the last TLOG replica, and have
to replicate the data before shutting down
+1, exactly my intention.
{quote}Write advice in the ref guide to always have at least one TLOG replica
available when doing cluster operations (and definitely have multiple TLOG
replicas existing for each shard)
{quote}
This is good general advice, but won't work for those who only have a total of
one replica that can assume leadership (1 TLOG or 1 NRT) along with 1+ PULL
replicas.
> Graceful shutdown can cause data loss with PULL replicas
> --------------------------------------------------------
>
> Key: SOLR-15585
> URL: https://issues.apache.org/jira/browse/SOLR-15585
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Reporter: Ishan Chattopadhyaya
> Priority: Major
>
> When using TLOG (leader)+PULL replicas, a clean shutdown of the node
> containing leaders (for which PULL replicas exist on other nodes) can be
> complete even before PULL replicas get a chance to sync all recently
> committed segments. One solution can be to have a check and enforce a
> replication on the node containing the leaders of such shards before shutting
> down.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
