Bryan Beaudreault created HBASE-28000:
-----------------------------------------
Summary: Empty Increments in batch calls are too expensive
Key: HBASE-28000
URL: https://issues.apache.org/jira/browse/HBASE-28000
Project: HBase
Issue Type: Bug
Reporter: Bryan Beaudreault
When an Increment is submitted via table.increment, we have a check which
throws an exception if the Increment is empty. There is no such check when
Increment is submitted through table.batch().
When this Increment makes its way to the server, it makes it all the way down
past taking a row lock before the first check as to whether the Increment has
any columns in it. Finally when we go to fetch the current cells from the table
for atomic update, the fact that getFamilyMap() is empty starts to matter. Even
still, beyond that I think we might still end up writing a somewhat useless
WALEdit.
The big issue here is the row lock, as this causes contention under certain
hotspot access patterns which greatly exacerbate the hotspot.
On the server side, we should check for whether the Increment is empty before
even attempting to serve it. On the client side, we should probably also
iterate the batch and check for emptiness.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
