[
https://issues.apache.org/jira/browse/HBASE-8031?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13600639#comment-13600639
]
Keith Turner commented on HBASE-8031:
-------------------------------------
I though of two example bugs rewriting could possibly masquerade. I am sure
there are more.
Client code reports data as flushed when its not really flushed, its actually
still flushing data in background threads out of order. If client dies and
you verify, you may see lost data. If you always rewrite data after client
dies you may never see this flush bug. Maybe close really does flush
everything properly, so under normal runs you never lose data. Data loss does
not always happen on RS, could happen in client code. We run continuous ingest
for a long period and then just kill all of the clients. This example makes me
think we should kill ingest clients more frequently as part of testing (in
addition to killing various server processes).
3 out 4 racks are accidentally powercycled by admins during test. Tiny bit of
data is lost, alot of data is rewritten, may cover up data loss.
> Adopt goraci as an Integration test
> -----------------------------------
>
> Key: HBASE-8031
> URL: https://issues.apache.org/jira/browse/HBASE-8031
> Project: HBase
> Issue Type: Improvement
> Components: test
> Reporter: Enis Soztutar
> Assignee: Enis Soztutar
> Fix For: 0.95.0, 0.98.0, 0.94.7
>
> Attachments: hbase-8031_v1.patch
>
>
> As you might know, I am a big fan of the goraci test that Keith Turner has
> developed, which in turn is inspired by the Accumulo test called Continuous
> Ingest.
> As much as I hate to say it, having to rely on gora and and external github
> library makes using this lib cumbersome. And lately we had to use this for
> testing against secure clusters and with Hadoop2, which gora does not support
> for now.
> So, I am proposing we add this test as an IT in the HBase code base so that
> all HBase devs can benefit from it.
> The original source code can be found here:
> * https://github.com/keith-turner/goraci
> * https://github.com/enis/goraci/
> From the javadoc:
> {code}
> Apache Accumulo [0] has a simple test suite that verifies that data is not
> * lost at scale. This test suite is called continuous ingest. This test runs
> * many ingest clients that continually create linked lists containing 25
> * million nodes. At some point the clients are stopped and a map reduce job
> is
> * run to ensure no linked list has a hole. A hole indicates data was lost.··
> *
> * The nodes in the linked list are random. This causes each linked list to
> * spread across the table. Therefore if one part of a table loses data, then
> it
> * will be detected by references in another part of the table.
> *
> Below is rough sketch of how data is written. For specific details look at
> * the Generator code.
> *
> * 1 Write out 1 million nodes· 2 Flush the client· 3 Write out 1 million that
> * reference previous million· 4 If this is the 25th set of 1 million nodes,
> * then update 1st set of million to point to last· 5 goto 1
> *
> * The key is that nodes only reference flushed nodes. Therefore a node should
> * never reference a missing node, even if the ingest client is killed at any
> * point in time.
> *
> * Some ASCII art time:
> * [ . . . ] represents one batch of random longs of length WIDTH
> *
> * _________________________
> * | ______ |
> * | | ||
> * __+_________________+_____ ||
> * v v v |||
> * first = [ . . . . . . . . . . . ] |||
> * ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ |||
> * | | | | | | | | | | | |||
> * prev = [ . . . . . . . . . . . ] |||
> * ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ |||
> * | | | | | | | | | | | |||
> * current = [ . . . . . . . . . . . ] |||
> * |||
> * ... |||
> * |||
> * last = [ . . . . . . . . . . . ] |||
> * | | | | | | | | | | |-----|||
> * | |--------||
> * |___________________________|
> {code}
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
