The row that is being updated could have the non-default column value for the 
default column.  The update may not choose this column. In that case, 
TRAF_REPLACE would blindly replace these columns with the default value. It 
becomes worse if the default column is a timestamp with current_timestamp as 
default.

Update can implicitly select and update the column value. This could be 
beneficial in case of non-wide row in aligned format. In case of wide row, 
hbase format tables are better for update as Ming suggested.

Selva

From: Liu, Ming (Ming) <[email protected]>
Sent: Wednesday, June 6, 2018 5:40 AM
To: [email protected]; [email protected]
Subject: RE: Upsert on part of columns get traf_merge which cause the 
performance not good

hi, Yuan,

You can check a discussion on this at 
https://lists.apache.org/[email protected]:2016-03, search 
for 'upsert semantics' topic.
Default mode should be MERGE as discussed, it was decided that. Correct 
behavior is more important than peformance.
I think you can create Non-Aligned format table, if the table is very 
frequently updated, it will be slower for read queries though.

Others may have better answer here

thanks,
Ming

From: Liu, Yuan (Yuan) <[email protected]<mailto:[email protected]>>
Sent: Wednesday, June 06, 2018 3:33 PM
To: [email protected]<mailto:[email protected]>; 
[email protected]<mailto:[email protected]>
Subject: Upsert on part of columns get traf_merge which cause the performance 
not good

Hi Trafodioneers,

In many cases, we find that "UPDATE" is not that efficient, so sometimes we 
rewrite the "UPDATE" as "UPSERT".

However, if we only write part of columns in upsert, the query plan will get 
"TRAF_MERGE", which is still not efficient.

Only if we write all the columns in upsert, the query plan can get good query 
plan, such as VSBB_UPSERT.

If table has hundrends of columns, this is very complex for users.

It is possible that upsert choose "TRAF_REPLACE" by default, at the same time, 
we keep the values which are not in upsert?


Best regards,
Yuan

Reply via email to