Oracle always uses HASH JOIN even when both tables are huge?

Question

my understanding is that HASH JOIN only makes sense when one of the 2 tables is small enough to fit into memory as a hash table.

but when I gave a query to oracle, with both tables having several hundred million rows, oracle still came up with a hash join explain plan. even when I tricked it with OPT_ESTIMATE(rows = ....) hints, it always decides to use HASH JOIN instead of merge sort join.

so I wonder how is HASH JOIN possible in the case of both tables being very large?

thanks Yang

Answer 1

So I wonder how is HASH JOIN possible in the case of both tables being very large?

It would be done in multiple passes: the driven table is read and hashed in chunks, the leading table is scanned several times.

This means that with limited memory hash join scales at O(N^2) while merge joins scales at O(N) (with no sorting needed of course), and on really large tables merge outperforms hash joins. However, the tables should be really large so that benefits of single read would outweight drawbacks of non-sequential access, and you would need all data from them (usually aggregated).

Given the RAM sized on modern servers, we are talking about really large reports on really large databases which take hours to build, not something you would really see in everyday live.

MERGE JOIN may also be useful when the output recordset is limited with rownum < N. But this means that the joined inputs should be already sorted which means they both be indexed which means NESTED LOOPS is available too, and that's what is usually chosen by the optimizer, since this is more efficient when the join condition is selective.

With their current implementations, MERGE JOIN always scans and NESTED LOOPS always seeks, while a more smart combination of both methods (backed up by statistics) would be preferred.

You may want to read this article in my blog:

• Things SQL needs: MERGE JOIN that would seek

Answer 2

Hash joins obviously work best when everything can fit in memory. But that does not mean they are not still the best join method when the table can't fit in memory. I think the only other realistic join method is a merge sort join.

If the hash table can't fit in memory, than sorting the table for the merge sort join can't fit in memory either. And the merge join needs to sort both tables. In my experience, hashing is always faster than sorting, for joining and for grouping.

But there are some exceptions. From the Oracle® Database Performance Tuning Guide, The Query Optimizer:

Hash joins generally perform better than sort merge joins. However, sort merge joins can perform better than hash joins if both of the following conditions exist:

  The row sources are sorted already.
  A sort operation does not have to be done.

---

Test

Instead of creating hundreds of millions of rows, it's easier to force Oracle to only use a very small amount of memory.

This chart shows that hash joins outperform merge joins, even when the tables are too large to fit in (artificially limited) memory:

---

Notes

For performance tuning it's usually better to use bytes than number of rows. But the "real" size of the table is a difficult thing to measure, which is why the chart displays rows. The sizes go approximately from 0.375 MB up to 14 MB. To double-check that these queries are really writing to disk you can run them with /*+ gather_plan_statistics */ and then query v$sql_plan_statistics_all.

I only tested hash joins vs merge sort joins. I didn't fully test nested loops because that join method is always incredibly slow with large amounts of data. As a sanity check, I did compare it once with the last data size, and it took at least several minutes before I killed it.

I also tested with different _area_sizes, ordered and unordered data, and different distinctness of the join column (more matches is more CPU-bound, less matches is more IO bound), and got relatively similar results.

However, the results were different when the amount of memory was ridiculously small. With only 32K sort|hash_area_size, merge sort join was significantly faster. But if you have so little memory you probably have more significant problems to worry about.

There are still many other variables to consider, such as parallelism, hardware, bloom filters, etc. People have probably written books on this subject, I haven't tested even a small fraction of the possibilities. But hopefully this is enough to confirm the general consensus that hash joins are best for large data.

---

Code

Below are the scripts I used:

--Drop objects if they already exist
drop table test_10k_rows purge;
drop table test1 purge;
drop table test2 purge;

--Create a small table to hold rows to be added.
--("connect by" would run out of memory later when _area_sizes are small.)
--VARIABLE: More or less distinct values can change results.  Changing
--"level" to something like "mod(level,100)" will result in more joins, which
--seems to favor hash joins even more.
create table test_10k_rows(a number, b number, c number, d number, e number);
insert /*+ append */ into test_10k_rows
    select level a, 12345 b, 12345 c, 12345 d, 12345 e
    from dual connect by level <= 10000;
commit;

--Restrict memory size to simulate running out of memory.
alter session set workarea_size_policy=manual;

--1 MB for hashing and sorting
--VARIABLE: Changing this may change the results.  Setting it very low,
--such as 32K, will make merge sort joins faster.
alter session set hash_area_size = 1048576;
alter session set sort_area_size = 1048576;

--Tables to be joined
create table test1(a number, b number, c number, d number, e number);
create table test2(a number, b number, c number, d number, e number);

--Type to hold results
create or replace type number_table is table of number;

set serveroutput on;

--
--Compare hash and merge joins for different data sizes.
--
declare
    v_hash_seconds number_table := number_table();
    v_average_hash_seconds number;
    v_merge_seconds number_table := number_table();
    v_average_merge_seconds number;

    v_size_in_mb number;
    v_rows number;
    v_begin_time number;
    v_throwaway number;

    --Increase the size of the table this many times
    c_number_of_steps number := 40;
    --Join the tables this many times
    c_number_of_tests number := 5;

begin
    --Clear existing data
    execute immediate 'truncate table test1';
    execute immediate 'truncate table test2';

    --Print headings.  Use tabs for easy import into spreadsheet.
    dbms_output.put_line('Rows'||chr(9)||'Size in MB'
        ||chr(9)||'Hash'||chr(9)||'Merge');

    --Run the test for many different steps
    for i in 1 .. c_number_of_steps loop
        v_hash_seconds.delete;
        v_merge_seconds.delete;
        --Add about 0.375 MB of data (roughly - depends on lots of factors)
        --The order by will store the data randomly.
        insert /*+ append */ into test1
        select * from test_10k_rows order by dbms_random.value;

        insert /*+ append */ into test2
        select * from test_10k_rows order by dbms_random.value;

        commit;

        --Get the new size
        --(Sizes may not increment uniformly)
        select bytes/1024/1024 into v_size_in_mb
        from user_segments where segment_name = 'TEST1';

        --Get the rows.  (select from both tables so they are equally cached)
        select count(*) into v_rows from test1;
        select count(*) into v_rows from test2; 

        --Perform the joins several times
        for i in 1 .. c_number_of_tests loop
            --Hash join
            v_begin_time := dbms_utility.get_time;
            select /*+ use_hash(test1 test2) */ count(*) into v_throwaway
            from test1 join test2 on test1.a = test2.a;
            v_hash_seconds.extend;
            v_hash_seconds(i) := (dbms_utility.get_time - v_begin_time) / 100;

            --Merge join
            v_begin_time := dbms_utility.get_time;
            select /*+ use_merge(test1 test2) */ count(*) into v_throwaway
            from test1 join test2 on test1.a = test2.a;
            v_merge_seconds.extend;
            v_merge_seconds(i) := (dbms_utility.get_time - v_begin_time) / 100;
        end loop;

        --Get average times.  Throw out first and last result.
        select ( sum(column_value) - max(column_value) - min(column_value) ) 
            / (count(*) - 2)
        into v_average_hash_seconds
        from table(v_hash_seconds);

        select ( sum(column_value) - max(column_value) - min(column_value) ) 
            / (count(*) - 2)
        into v_average_merge_seconds
        from table(v_merge_seconds);

        --Display size and times
        dbms_output.put_line(v_rows||chr(9)||v_size_in_mb||chr(9)
            ||v_average_hash_seconds||chr(9)||v_average_merge_seconds);

    end loop;
end;
/