INNER JOIN vs LEFT JOIN performance in SQL Server

Solution 1:

A LEFT JOIN is absolutely not faster than an INNER JOIN. In fact, it's slower; by definition, an outer join (LEFT JOIN or RIGHT JOIN) has to do all the work of an INNER JOIN plus the extra work of null-extending the results. It would also be expected to return more rows, further increasing the total execution time simply due to the larger size of the result set.

(And even if a LEFT JOIN were faster in specific situations due to some difficult-to-imagine confluence of factors, it is not functionally equivalent to an INNER JOIN, so you cannot simply go replacing all instances of one with the other!)

Most likely your performance problems lie elsewhere, such as not having a candidate key or foreign key indexed properly. 9 tables is quite a lot to be joining so the slowdown could literally be almost anywhere. If you post your schema, we might be able to provide more details.


Edit:

Reflecting further on this, I could think of one circumstance under which a LEFT JOIN might be faster than an INNER JOIN, and that is when:

  • Some of the tables are very small (say, under 10 rows);
  • The tables do not have sufficient indexes to cover the query.

Consider this example:

CREATE TABLE #Test1
(
    ID int NOT NULL PRIMARY KEY,
    Name varchar(50) NOT NULL
)
INSERT #Test1 (ID, Name) VALUES (1, 'One')
INSERT #Test1 (ID, Name) VALUES (2, 'Two')
INSERT #Test1 (ID, Name) VALUES (3, 'Three')
INSERT #Test1 (ID, Name) VALUES (4, 'Four')
INSERT #Test1 (ID, Name) VALUES (5, 'Five')

CREATE TABLE #Test2
(
    ID int NOT NULL PRIMARY KEY,
    Name varchar(50) NOT NULL
)
INSERT #Test2 (ID, Name) VALUES (1, 'One')
INSERT #Test2 (ID, Name) VALUES (2, 'Two')
INSERT #Test2 (ID, Name) VALUES (3, 'Three')
INSERT #Test2 (ID, Name) VALUES (4, 'Four')
INSERT #Test2 (ID, Name) VALUES (5, 'Five')

SELECT *
FROM #Test1 t1
INNER JOIN #Test2 t2
ON t2.Name = t1.Name

SELECT *
FROM #Test1 t1
LEFT JOIN #Test2 t2
ON t2.Name = t1.Name

DROP TABLE #Test1
DROP TABLE #Test2

If you run this and view the execution plan, you'll see that the INNER JOIN query does indeed cost more than the LEFT JOIN, because it satisfies the two criteria above. It's because SQL Server wants to do a hash match for the INNER JOIN, but does nested loops for the LEFT JOIN; the former is normally much faster, but since the number of rows is so tiny and there's no index to use, the hashing operation turns out to be the most expensive part of the query.

You can see the same effect by writing a program in your favourite programming language to perform a large number of lookups on a list with 5 elements, vs. a hash table with 5 elements. Because of the size, the hash table version is actually slower. But increase it to 50 elements, or 5000 elements, and the list version slows to a crawl, because it's O(N) vs. O(1) for the hashtable.

But change this query to be on the ID column instead of Name and you'll see a very different story. In that case, it does nested loops for both queries, but the INNER JOIN version is able to replace one of the clustered index scans with a seek - meaning that this will literally be an order of magnitude faster with a large number of rows.

So the conclusion is more or less what I mentioned several paragraphs above; this is almost certainly an indexing or index coverage problem, possibly combined with one or more very small tables. Those are the only circumstances under which SQL Server might sometimes choose a worse execution plan for an INNER JOIN than a LEFT JOIN.

Solution 2:

There is one important scenario that can lead to an outer join being faster than an inner join that has not been discussed yet.

When using an outer join, the optimizer is always free to drop the outer joined table from the execution plan if the join columns are the PK of the outer table, and none of the outer table columns are referenced outside of the outer join itself. For example SELECT A.* FROM A LEFT OUTER JOIN B ON A.KEY=B.KEY and B.KEY is the PK for B. Both Oracle (I believe I was using release 10) and Sql Server (I used 2008 R2) prune table B from the execution plan.

The same is not necessarily true for an inner join: SELECT A.* FROM A INNER JOIN B ON A.KEY=B.KEY may or may not require B in the execution plan depending on what constraints exist.

If A.KEY is a nullable foreign key referencing B.KEY, then the optimizer cannot drop B from the plan because it must confirm that a B row exists for every A row.

If A.KEY is a mandatory foreign key referencing B.KEY, then the optimizer is free to drop B from the plan because the constraints guarantee the existence of the row. But just because the optimizer can drop the table from the plan, doesn't mean it will. SQL Server 2008 R2 does NOT drop B from the plan. Oracle 10 DOES drop B from the plan. It is easy to see how the outer join will out-perform the inner join on SQL Server in this case.

This is a trivial example, and not practical for a stand-alone query. Why join to a table if you don't need to?

But this could be a very important design consideration when designing views. Frequently a "do-everything" view is built that joins everything a user might need related to a central table. (Especially if there are naive users doing ad-hoc queries that do not understand the relational model) The view may include all the relevent columns from many tables. But the end users might only access columns from a subset of the tables within the view. If the tables are joined with outer joins, then the optimizer can (and does) drop the un-needed tables from the plan.

It is critical to make sure that the view using outer joins gives the correct results. As Aaronaught has said - you cannot blindly substitute OUTER JOIN for INNER JOIN and expect the same results. But there are times when it can be useful for performance reasons when using views.

One last note - I haven't tested the impact on performance in light of the above, but in theory it seems you should be able to safely replace an INNER JOIN with an OUTER JOIN if you also add the condition <FOREIGN_KEY> IS NOT NULL to the where clause.

Solution 3:

If everything works as it should it shouldn't, BUT we all know everything doesn't work the way it should especially when it comes to the query optimizer, query plan caching and statistics.

First I would suggest rebuilding index and statistics, then clearing the query plan cache just to make sure that's not screwing things up. However I've experienced problems even when that's done.

I've experienced some cases where a left join has been faster than a inner join.

The underlying reason is this: If you have two tables and you join on a column with an index (on both tables). The inner join will produce the same result no matter if you loop over the entries in the index on table one and match with index on table two as if you would do the reverse: Loop over entries in the index on table two and match with index in table one. The problem is when you have misleading statistics, the query optimizer will use the statistics of the index to find the table with least matching entries (based on your other criteria). If you have two tables with 1 million in each, in table one you have 10 rows matching and in table two you have 100000 rows matching. The best way would be to do an index scan on table one and matching 10 times in table two. The reverse would be an index scan that loops over 100000 rows and tries to match 100000 times and only 10 succeed. So if the statistics isn't correct the optimizer might choose the wrong table and index to loop over.

If the optimizer chooses to optimize the left join in the order it is written it will perform better than the inner join.

BUT, the optimizer may also optimize a left join sub-optimally as a left semi join. To make it choose the one you want you can use the force order hint.

Solution 4:

Try both queries (the one with inner and left join) with OPTION (FORCE ORDER) at the end and post the results. OPTION (FORCE ORDER) is a query hint that forces the optimizer to build the execution plan with the join order you provided in the query.

If INNER JOIN starts performing as fast as LEFT JOIN, it's because:

  • In a query composed entirely by INNER JOINs, the join order doesn't matter. This gives freedom for the query optimizer to order the joins as it sees fit, so the problem might rely on the optimizer.
  • With LEFT JOIN, that's not the case because changing the join order will alter the results of the query. This means the engine must follow the join order you provided on the query, which might be better than the optimized one.

Don't know if this answers your question but I was once in a project that featured highly complex queries making calculations, which completely messed up the optimizer. We had cases where a FORCE ORDER would reduce the execution time of a query from 5 minutes to 10 seconds.

Solution 5:

Have done a number of comparisons between left outer and inner joins and have not been able to find a consisten difference. There are many variables. Am working on a reporting database with thousands of tables many with a large number of fields, many changes over time (vendor versions and local workflow) . It is not possible to create all of the combinations of covering indexes to meet the needs of such a wide variety of queries and handle historical data. Have seen inner queries kill server performance because two large (millions to tens of millions of rows) tables are inner joined both pulling a large number of fields and no covering index exists.

The biggest issue though, doesn't seem to appeaer in the discussions above. Maybe your database is well designed with triggers and well designed transaction processing to ensure good data. Mine frequently has NULL values where they aren't expected. Yes the table definitions could enforce no-Nulls but that isn't an option in my environment.

So the question is... do you design your query only for speed, a higher priority for transaction processing that runs the same code thousands of times a minute. Or do you go for accuracy that a left outer join will provide. Remember that inner joins must find matches on both sides, so an unexpected NULL will not only remove data from the two tables but possibly entire rows of information. And it happens so nicely, no error messages.

You can be very fast as getting 90% of the needed data and not discover the inner joins have silently removed information. Sometimes inner joins can be faster, but I don't believe anyone making that assumption unless they have reviewed the execution plan. Speed is important, but accuracy is more important.