Details
-
Improvement
-
Status: Patch Available
-
Minor
-
Resolution: Unresolved
-
None
-
None
-
None
Description
Join operation on table with skewed data takes most of execution time handling the skewed keys. But mostly we already know about that and even know what is look like the skewed keys.
If we can explicitly assign reducer slots for the skewed keys, total execution time could be greatly shortened.
As for a start, I've extended join grammar something like this.
select * from src a join src b on a.key=b.key skew on (a.key+1 < 50, a.key+1 < 100, a.key < 150);
which means if above query is executed by 20 reducers, one reducer for a.key+1 < 50, one reducer for 50 <= a.key+1 < 100, one reducer for 99 <= a.key < 150, and 17 reducers for others (could be extended to assign more than one reducer later)
This can be only used with common-inner-equi joins. And skew condition should be composed of join keys only.
Work till done now will be updated shortly after code cleanup.
----------------------------
All expressions in the clause "SKEW ON (expr1, expr2, ...)" are called skew condition and consist of skew expression*, which is simple boolean expression for the group and optional CLUSTER/DISTRIBUTED expression. Skew expressions will be evaluated sequentially at runtime, deciding skew group for a row. Each skew group has reserved partition slot(s), to which all rows in a group would be assigned.
The number of partition slot reserved for each skew group is decided by cluster expression. Before submitting the MR job, hive calculates size of each skew groups. If a skew group is "CLUSTER BY 20 PERCENT" and total partition slot (=number of reducer) is, say, 20, the group will reserve 4 partition slots for it, etc.
The optional "DISTRIBUTE BY" decides how the rows in a skew group is dispersed in the range of reserved slots (If there is only one slot for a group, this is meaningless). Currently, three distribution policies are available: RANDOM, KEYS, <expression>.
1. RANDOM : rows from driver alias** are dispersed by random and rows of other aliases are multicasted to all slots (default behavior)
2. KEYS : rows from driver alias are dispersed by hash of keys same for
3. expression : determined by evaluation result of user-provided expression
Only possible with inner, equi, common-joins. Not yet supports join tree merging or vectorization.
- Might be used by other RS users like "SORT BY" or "GROUP BY" (not-yet)
- If there are column statistics for the skewness of the key, it could be possible applied automatically (not-yet)
For example, if 20 reducers are used for the query below,
select count(*) from src a join src b on a.key=b.key skew on ( a.key = '0' CLUSTER BY 10 PERCENT, b.key < '100' CLUSTER BY 20 PERCENT DISTRIBUTE BY upper(b.key), cast(a.key as int) > 300 CLUSTER BY 40 PERCENT DISTRIBUTE BY KEYS);
Skew group-0 would reserve 2 slots (#6~#7), for group-1, 4 slots (#8~#11), for group-2, 8 slots (#12~#19) and others will use remaining 6 slots (#0~#5).
For key='0' from alias a(driver alias), it will be assigned to a slot of group-0 : 6 or 7
For key='0' from alias b(non-driver alias), it will be multicasted to all slots of group-0 : 6 and 7
For key='50' from alias a(non-driver alias), it will be multicasted to all slots of group-1 : 8 and 9 and 10 and 11
For key='50' from alias b(driver alias), it will be assigned to a slot of group-1 : 8 or 9 or 10 or 11
For key='500' from alias a(driver alias), it will be assigned to a slot of group-2 by modulation of hashcode of DISTRIBUTE expression
For key='500' from alias b(non-driver alias), it will be multicasted to all slots of group-2 : #12~#19
For key='200', it's not belong to any skew group and will be processed normally in the range of partition slot 0~5.
*skew expression :
1. all expressions should be made of expression in join condition, which means if join condition is "a.key=b.key", user can make any expression with "a.key" or "b.key". But if join condition is a.key+1=b.key, user cannot make expression with "a.key" solely (or make expression with "a.key+1").
2. all expressions should reference one and only-one side of aliases. For example, simple constant expressions or expressions referencing both side of join condition ("a.key+b.key<100") is not allowed.
3. all functions in expression should be deterministic and stateless.
4. DISTRIBUTED expression should have same alias with skew expression.
**driver alias :
1. driver alias means the sole referenced alias from skew expression, which is important for RANDOM distribution. Rows from driver alias are assigned to one slot, but rows from other aliases will be multicasted to all slots of the group.