Blocker Description
Stall happens down in SequenceFile in the first call to getPos inside readRecordLength. I tried the johano patch from HADOOP-2172 that restores the positional cache but that didn't seem to be the issue here.
Here is data to support my assertion.
I wrote a little program to make a MapFile of 1M records. I then did 1M random reads from same file. Below are timings from a 0.15.0 and TRUNK as of this afternoon run.
0.15.x branch:
[stack@aa0-000-12 branch-0.15]$ ./bin/hadoop org.apache.hadoop.io.TestMapFile .07/12/15 01:29:02 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable 07/12/15 01:29:02 INFO io.TestMapFile: Writing 1000000 rows to testMapFileRandomAccess 07/12/15 01:32:04 INFO io.TestMapFile: Writing 1000000 records took 182009ms 07/12/15 01:32:04 INFO io.TestMapFile: Reading 1000000 random rows 07/12/15 01:48:02 INFO io.TestMapFile: Reading 1000000 random records took 958243ms Time: 1,140.652 OK (1 test)
For the below test using TRUNK r604352, I amended the test so it output a log message every 100k reads:
[stack@aa0-000-12 hadoop-trunk]$ ./bin/hadoop org.apache.hadoop.io.TestMapFile .07/12/15 01:56:34 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable 07/12/15 01:56:34 INFO io.TestMapFile: Writing 1000000 rows to testMapFileRandomAccess 07/12/15 01:59:38 INFO io.TestMapFile: Writing 1000000 records took 183986ms 07/12/15 01:59:38 INFO io.TestMapFile: Reading 1000000 random rows .........
After 20 minutes it still hadn't printed out the 'read 100k messages' (I had to leave – will fill in final figures later)