While looking at the HLog code I saw some funny methods that are *only* called from tests.
public long append(HRegionInfo regionInfo, HLogKey logKey, WALEdit logEdit,
HTableDescriptor htd, boolean doSync)
which seems to be (almost) a carbon copy of:
private long append(HRegionInfo info, byte  tableName, WALEdit edits, UUID clusterId,
final long now, HTableDescriptor htd, boolean doSync)
I plan to remove the former.
Should we think about the notion of "group commits". Many relational databases delay the commit of a transaction a bit in order to group multiple commits together.
Only after all such grouped transactions are committed are the respective clients notified.
HBase does a bit of this. If Puts hit the same RegionServer at the same time, the edits get added to WAL and all outstanding edits are sync'ed at the end of the Put.
We could widen the window by delaying this by (say) a second or so, and thus reduce the number of hflushes (or more importantly hsyncs, once I finally get to adding them to HBase).
Or is the current logic good enough?
谢良 2013-06-15, 01:02
Jean-Daniel Cryans 2013-06-15, 02:34