ExtentRows (i.e. the number of rows per extent) is hardcoded to 8M in:

unsigned ExtentMap::getExtentRows()

{

//	boost::mutex::scoped_lock lk(fConfigCacheMutex);

//	checkReloadConfig( );

	ExtentRows = 0x800000;

	return ExtentRows;

}

The comments say that 1M and 64M are valid values but obviously the config code is commented out so its not configurable without recompiling. I tried changing the value to 0x100000 (1M) and recompiled. Creating a table failed with an extent error (forgot to save it). This suggest that this either never worked in infinidb (the code is this way in the first open source version) or had issues.

Assuming the system can scale to a larger extent map, it would seem that you could get better extent map elimination for many use cases with a smaller extent size.