[
https://issues.apache.org/jira/browse/KUDU-1573?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15454402#comment-15454402
]
Todd Lipcon commented on KUDU-1573:
-----------------------------------
I managed to reproduce this here by kernel-panicing a tserver while in the
middle of a heavy workload. Because it was running without log fsync, the last
several log segments existed but had all \0 data. We probably can't recover
from this in a very consistent manner. We should consider a "more durable"
option: enable fsyncing logs but only right before each tablet metadata flush.
I believe this would ensure that the logs would be at least as new as the
latest metadata, and thus would be a "consistent snapshot" to restart from
after a power outage or kernel panic, but still be less expensive than enabling
fsync on every log write.
> In corner cases , tablet could not recovery successfully from node crash.
> ---------------------------------------------------------------------------
>
> Key: KUDU-1573
> URL: https://issues.apache.org/jira/browse/KUDU-1573
> Project: Kudu
> Issue Type: Bug
> Reporter: zhangsong
> Attachments: wal_recovery_dir.zip
>
>
> Last friday ,one of nodes of my kudu cluster crashed and , tablet can not
> recovery successfully after restart kudu-tserver, i observed error messages
> from log:
> (TABLET_DATA_READY): Corruption: Could not open LogReader. Reason: Unable to
> initialize log reader: Segment sequence numbers are not consecutive. Previous
> segment: seqno 0, path
> /export/servers/kudu/tserver_wal_data_7052/wals/ed0d8b3a835e4c27afe695252ad0b8f5.recovery/wal-000000018;
> Current segment: seqno 17, path
> /export/servers/kudu/tserver_wal_data_7052/wals/ed0d8b3a835e4c27afe695252ad0b8f5.recovery/wal-000000017
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
