Details
-
Task
-
Status: Closed
-
Major
-
Resolution: Won't Fix
-
2.1
-
None
Description
Justification is the same as for IGNITE-6201 (merge joins). If join is equijoin, and participating attributes are not sorted in advance, then instead of executing nested loops, we can do the following:
1) Get the table with smaller number of entries
2) Build a hash table from target attribute to the list of matching entries
3) Iterate over larger table and perform lookup into hash table
Note that this operation might be memory-intensive in case smaller table is large enough still. For this reason we should provide a mechanism to avoid out-of-memory errors. Let's avoid spilling to disk in the first iteration, and try to set a kind of hard limit on how much data can be kept in memory. If we see that this limit cannot be satisfied, then fallback to nested loops.