Optimize performance for UnionScanExec and MemBuffer

[lcwangchao]

Enhancement

The UnionScan is not very efficient, for example:

• UnionScanExec will filter all rows/indexes in Open instead of Next. It will take more time when the SQL has a LIMIT because Open will not considerate it.
• Maybe we can implementMemBuffer more efficient.

The union scan's performance will affect the performance of temporary table, cached table and queries in a txn with a lot of uncommitted rows.

[tiancaiamao]

cd executor;
 go test -run XXX -bench BenchmarkUnionScan -cpuprofile cpu.out -benchtime 45s

The inefficency comes from too parts:

1. Open() always drain all the data even it might be useless in the Limit scenario. It's better to make it a streaming API
2. The decode / encode and row format translation cost. When loading, we decode kv -> row in []Datum representation. and when merging, []Datum -> []Datum, and finally output is translated from row representation to chunk, i.e. []Datum -> chunk

[ekexium]

Hey @lcwangchao do you have more elaboration in "Maybe we can implementMemBuffer more efficient."? The flamegraph shows that time spent in memdb is quite little. Did you mean codec-related work?

[tiancaiamao]

Hey @lcwangchao do you have more elaboration in "Maybe we can implementMemBuffer more efficient."? The flamegraph shows that time spent in memdb is quite little. Did you mean codec-related work?

After my previous optimization, much of that had been improved.