Find below the general outline of hdfs scan for text formatted hive tables.
Compiler returns a list of scan ranges and the begin range and number of ranges to be done by each instance of TCB in TDB. This list of scan ranges is also re-computed at run time possibly based on a CQD
The scan range for a TCB can come from the same or different hdfs files. TCB creates two threads to read these ranges.Two ranges (for the TCB) are initially assigned to these threads. As and when a range is completed, the next range (assigned for the TCB) is picked up by the thread. Ranges are read in multiples of hdfs scan buffer size at the TCB level. Default hdfs scan buffer size is 64 MB. Rows from hdfs scan buffer is processed and moved into up queue. If the range contains a record split, then the range is extended to read up to range tail IO size to get the full row. The range that had the latter part of the row ignores it because the former range processes it. Record split at the file level is not possible and/or not supported.
For compression, the compiler returns the range info such that the hdfs scan buffer can hold the full uncompressed buffer.
Reader threads feature too complex to maintain in C++
Error handling at the layer below the TCB is missing or errors are not propagated to work method causing incorrect results
Possible multiple copying of data
Libhdfs calls are not optimized. It was observed that the method Ids are being obtained many times. Need to check if this problem still exists.
Now that we clearly know what is expected, it could be optimized better
- Reduced scan buffer size for smoother data flow
- Better thread utilization
- Avoid multiple copying of data.
Unable to comprehend the need for two threads for pre-fetch especially when one range is completed fully before the data from next range is processed.
Following are the hdfsCalls used by programs at exp and executor directory.
Make changes to use direct Java APIs for these calls. However, come up with better mechanism to move the data from Java and JNI, avoid unnecessary copying of data, better thread management via Executor concepts in Java. Hence it won’t be direct mapping of these calls to hdfs Java API. Instead, use the abstraction like what is being done for HBase access.
I believe newer implementation will be optimized better and hence improved performance. (but not many folds)