问题
I found this sample interview question and answer posted on toptal reproduced here. But I don't really understand the code. How can a UNION ALL turn into a UNIION (distinct) like that? Also, why is this code faster?
QUESTION
Write a SQL query using UNION ALL (not UNION) that uses the WHERE clause to eliminate duplicates. Why might you want to do this? Hide answer You can avoid duplicates using UNION ALL and still run much faster than UNION DISTINCT (which is actually same as UNION) by running a query like this:
ANSWER
SELECT * FROM mytable WHERE a=X UNION ALL SELECT * FROM mytable WHERE b=Y AND a!=X
The key is the AND a!=X part. This gives you the benefits of the UNION (a.k.a., UNION DISTINCT) command, while avoiding much of its performance hit.
回答1:
But in the example, the first query has a condition on column a, whereas the second query has a condition on column b. This probably came from a query that's hard to optimize:
SELECT * FROM mytable WHERE a=X OR b=Y
This query is hard to optimize with simple B-tree indexing. Does the engine search an index on column a? Or on column b? Either way, searching the other term requires a table-scan.
Hence the trick of using UNION to separate into two queries for one term each. Each subquery can use the best index for each search term. Then combine the results using UNION.
But the two subsets may overlap, because some rows where b=Y may also have a=X in which case such rows occur in both subsets. Therefore you have to do duplicate elimination, or else see some rows twice in the final result.
SELECT * FROM mytable WHERE a=X
UNION DISTINCT
SELECT * FROM mytable WHERE b=Y
UNION DISTINCT is expensive because typical implementations sort the rows to find duplicates. Just like if you use SELECT DISTINCT ....
We also have a perception that it's even more "wasted" work if the two subset of rows you are unioning have a lot of rows occurring in both subsets. It's a lot of rows to eliminate.
But there's no need to eliminate duplicates if you can guarantee that the two sets of rows are already distinct. That is, if you guarantee there is no overlap. If you can rely on that, then it would always be a no-op to eliminate duplicates, and therefore the query can skip that step, and therefore skip the costly sorting.
If you change the queries so that they are guaranteed to select non-overlapping subsets of rows, that's a win.
SELECT * FROM mytable WHERE a=X
UNION ALL
SELECT * FROM mytable WHERE b=Y AND a!=X
These two sets are guaranteed to have no overlap. If the first set has rows where a=X and the second set has rows where a!=X then there can be no row that is in both sets.
The second query therefore only catches some of the rows where b=Y, but any row where a=X AND b=Y is already included in the first set.
So the query achieves an optimized search for two OR terms, without producing duplicates, and requiring no UNION DISTINCT operation.
回答2:
The question will be correct if the table has unique identifier - primary key. Otherwise every select can return many the same rows.
To understand why it can faster let's look at how database executes UNION ALL and UNION.
The first is simple joining results from two independent queries. These queries can be processed in parallel and taken to client one by one.
The second is joining + distinction. To distinct records from 2 queries db needs to have all them in memory or if memory is not enough db needs to store them to temporary table and next select unique ones. This is where performance degradation can be. DB's are pretty smart and distinction algorithms are developed good but for large result sets it could be a problem anyway.
UNION ALL + additional WHERE condition can be faster if an index will be used while filtering. So, here the performance magic.
回答3:
I guess it will work
select col1 From (
select row_number() over (partition by col1 order by col1) as b, col1
from (
select col1 From u1
union all
select col1 From u2 ) a
) x
where x.b =1
回答4:
This will also do the same trick:
select * from (
select * from table1
union all
select * from table2
) a group by
columns
having count(*) >= 1
or
select * from table1
union all
select * from table2 b
where not exists (select 1 from table1 a where a.col1 = b.col1)
来源：https://stackoverflow.com/questions/41729082/sql-union-all-to-eliminate-duplicates