Sujith created CARBONDATA-527:
---------------------------------
Summary: Greater than/less-than/Like filters optimization for
dictionary columns
Key: CARBONDATA-527
URL: https://issues.apache.org/jira/browse/CARBONDATA-527
Project: CarbonData
Issue Type: Improvement
Reporter: Sujith
Current design
In greater than/less-than/Like filters, system first iterates each row present
in the dictionary cache for identifying valid filter actual members by
applying the filter expression , once evaluation done system will hold the list
of identified valid filter actual member values(String), now in next step again
system will look up the dictionary cache in order to identify the dictionary
surrogate values of the identified members. this look up is an additional cost
to our system even though the look up methodology is an binary search in
dictionary cache.
Proposed design/solution:
Identify the dictionary surrogate values in filter expression evaluation step
itself when actual dictionary values will be scanned for identifying valid
filter members .
Keep a dictionary counter variable which will be increased when system
iterates through the dictionary cache in order to retrieve each actual member
stored in dictionary cache , after this system will evaluate each row against
the filter expression to identify whether its a valid filter member or not,
while doing this process itself counter value can be taken as valid selected
dictionary value since the actual member values and its dictionary values will
be kept in same order in dictionary cache as the iteration order.
thus it will eliminate the further dictionary look up step which is required
to retrieve the dictionary surrogate value against identified actual valid
filter member. this can also increase significantly the filter query
performance of such filter queries which require expression evaluation to
identify it the filter members by looking up dictionary cache, like greater
than/less-than/Like filters .
Note : this optimization is applicable for dictionary columns.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)