First, I want to thank Billie, Todd, and Eric for their help so far. It is much 
appreciated.

Now, on to the results. As I stated before I'm attempting to run the Accumulo 
auto tests from test/system/auto. When run standalone, most tests complete 
successfully on MapR but there have been some issues. That said, I'm really 
glad that Accumulo has all of these tests. Very useful for verification.

First, we found one item that required a change to MapR:

simple.bulkSplitOptimization.BulkSplitOptimizationTest failed with a seek error 
which has now been addressed thanks to Billie's help. This is MapR change 6539 
and we'll put that into a release soon.

Secondly, the Accumulo tests seem to have various minor glitches. I've 
documented them here for reference. I hope this is useful to the Accumulo 
community. I could open JIRA defects if you want but I don't have an account on 
the tracking system. Just let me know.

1) simple.examples.Examples was failing with a class loader issue. By examining 
simple/examples.py I was able to determine that the package names were not 
correct when referring to certain subtests, such as RandomBatchScanner. 
Specifically it referred to 
org.apache.accumulo.examples.client.RandomBatchScanner rather than the 
appropriate org.apache.accumulo.examples.simple.client.RandomBatchScanner. The 
same is true for RandomBatchWriter.

2) simple.mapreduce.MapReduceTest fails because it assumes ZOOKEEPER_HOME is 
set as an environment variable. Most of the other scripts appear to get this 
information from TestUtils.py. The workaround is simply to set that variable 
before using run.py.

3) simple.dynamic.DynamicClassloader fails with class loader issues referencing 
both Accumulo and Hadoop classes. I examined simple/dynamic.py and determined 
that the class path references weren't sufficient where javac was invoked. I 
added the following to the class path command line options for javac and now it 
works:  
/opt/accumulo-1.4.0/lib/*:/opt/mapr/lib/*:/opt/mapr/hadoop/hadoop-0.20.2/lib/*

Obviously that's a bit of a hack. I'm sure someone more familiar with the code 
can do this better.

4) stress.weird.LateLastContact fails with a ZOOKEEPER_HOME reference just like 
MapReduceTest.

5) simple.deleterows.DeleteRowsSplitTest fails with a timeout. The test is 
coded to wait 120 seconds which wasn't' quite long enough on my system. I 
changed it to 180 seconds and it finishes cleanly after 138 seconds. Is this 
typical or expected? 

6) simple.zooCacheTest.ZooCacheTest has one minor issue. The issue is that the 
test times out on my system after 60 seconds. I changed the timeout to 120 
seconds and the test completes successfully in 95 seconds. Is this typical or 
expected?

7) MapR does not normally install a full zookeeper install on each node. 
Instead we install a zookeeper "client" on each node which is basically the 
zookeeper JAR file - /opt/mapr/lib/zookeeper-3.3.2.jar. If I set ZOOKEEPER_HOME 
to /opt/mapr/lib almost everything works fine since Accumulo seems to only need 
the JAR file at that location. The one exception is the 
stress.weird.LateLastContact which directly references zkClient.sh (in 
stress/weird.py) which isn't part of the MapR install. I wanted your opinion on 
this. Is the test's behavior crucial? Does Accumulo need more of zookeeper than 
the JAR? I want Accumulo to work,  so if there really is a strong need for more 
of Zookeeper I'll recommend we make adjustments to the product once we 
understand the requirement.

Third, there are still two tests that are failing when run in standalone mode 
and I haven't determined the issue: simple.largeRow.LargeRowTest and 
simple.batchScanSplit.BatchScanSplitTest. I will post details on those in a 
moment as separate messages as this email is long enough.

Thanks again for all of your help so far,
Keys
________________________________
Keys Botzum
Senior Principal Technologist
WW Systems Engineering
[email protected]
443-718-0098
MapR Technologies
http://www.mapr.com


Reply via email to