Description
This is a part of "Partition Pruning" IEP [1].
Currently we try to extract partitions from map queries and route requests accordingly. Several problems with this approach:
1) Individual map queries may target the same partition, but we never know that, so we may want to setup a merge table when it is not really needed.
2) Sometimes query may reduce to no partitions. In this case we should not execute anything at all and simply return empty result set.
3) If the whole query targets only one partition, we may skip the whole splitter phase!
I propose to do the following:
- Try to extract partition from "original" query.
- If we see that exactly one partition is involved, then original query is a map query, and reduce should be performed in "skip merge table" mode
- If we see that there are no partitions because query is invalid (e.g. id = 1 AND id = 2), then stop and return no results. This decision should be cached in two-step plan.
- If we see that there are some partitions, then we should apply arguments and see the result. If result set is empty - return, but do not cache empty result set decision, as it may change for another set of arguments.
- If none of above hold, then do pushdowns and split, and analyze partitions of individual map queries. For those of them where partition set is empty, we should return empty result set without executing anything.
[1] https://cwiki.apache.org/confluence/display/IGNITE/IEP-24%3A+SQL+Partition+Pruning
Attachments
Issue Links
- blocks
-
IGNITE-10308 Jdbc Thin: implement client side best effort affinity
-
- Resolved
-
- links to
