Optimizer Hints

The Impala SQL supports query hints, for fine-tuning the inner workings of queries. Specify hints as a temporary workaround for expensive queries, where missing statistics or other factors cause inefficient performance.

Hints are most often used for the resource-intensive Impala queries, such as:

  • Join queries involving large tables, where intermediate result sets are transmitted across the network to evaluate the join conditions.
  • Inserting into partitioned Parquet tables, where many memory buffers could be allocated on each host to hold intermediate results for each partition.

Syntax:

In Impala 2.0 and higher, you can specify the hints inside comments that use either the /* */ or -- notation. Specify a + symbol immediately before the hint name. Recently added hints are only available using the /* */ and -- notation. For clarity, the /* */ and -- styles are used in the syntax and examples throughout this section. With the /* */ or -- notation for hints, specify a + symbol immediately before the first hint name. Multiple hints can be specified separated by commas, for example /* +clustered,shuffle */

  1. SELECT STRAIGHT_JOIN select_list FROM
  2. join_left_hand_table
  3. JOIN /* +BROADCAST|SHUFFLE */
  4. join_right_hand_table
  5. remainder_of_query;
  6. SELECT select_list FROM
  7. join_left_hand_table
  8. JOIN -- +BROADCAST|SHUFFLE
  9. join_right_hand_table
  10. remainder_of_query;
  11. INSERT insert_clauses
  12. /* +SHUFFLE|NOSHUFFLE */
  13. SELECT remainder_of_query;
  14. INSERT insert_clauses
  15. -- +SHUFFLE|NOSHUFFLE
  16. SELECT remainder_of_query;
  17. INSERT /* +SHUFFLE|NOSHUFFLE */
  18. insert_clauses
  19. SELECT remainder_of_query;
  20. INSERT -- +SHUFFLE|NOSHUFFLE
  21. insert_clauses
  22. SELECT remainder_of_query;
  23. UPSERT /* +SHUFFLE|NOSHUFFLE */
  24. upsert_clauses
  25. SELECT remainder_of_query;
  26. UPSERT -- +SHUFFLE|NOSHUFFLE
  27. upsert_clauses
  28. SELECT remainder_of_query;
  29. SELECT select_list FROM
  30. table_ref
  31. /* +{SCHEDULE_CACHE_LOCAL | SCHEDULE_DISK_LOCAL | SCHEDULE_REMOTE}
  32. [,RANDOM_REPLICA] */
  33. remainder_of_query;
  34. INSERT insert_clauses
  35. -- +CLUSTERED
  36. SELECT remainder_of_query;
  37. INSERT insert_clauses
  38. /* +CLUSTERED */
  39. SELECT remainder_of_query;
  40. INSERT -- +CLUSTERED
  41. insert_clauses
  42. SELECT remainder_of_query;
  43. INSERT /* +CLUSTERED */
  44. insert_clauses
  45. SELECT remainder_of_query;
  46. UPSERT -- +CLUSTERED
  47. upsert_clauses
  48. SELECT remainder_of_query;
  49. UPSERT /* +CLUSTERED */
  50. upsert_clauses
  51. SELECT remainder_of_query;
  52. CREATE /* +SHUFFLE|NOSHUFFLE */
  53. table_clauses
  54. AS SELECT remainder_of_query;
  55. CREATE -- +SHUFFLE|NOSHUFFLE
  56. table_clauses
  57. AS SELECT remainder_of_query;
  58. CREATE /* +CLUSTER|NOCLUSTER */
  59. table_clauses
  60. AS SELECT remainder_of_query;
  61. CREATE -- +CLUSTER|NOCLUSTER
  62. table_clauses
  63. AS SELECT remainder_of_query;

The square bracket style hints are supported for backward compatibility, but the syntax is deprecated and will be removed in a future release. For that reason, any newly added hints are not available with the square bracket syntax.

  1. SELECT STRAIGHT_JOIN select_list FROM
  2. join_left_hand_table
  3. JOIN [{ /* +BROADCAST */ | /* +SHUFFLE */ }]
  4. join_right_hand_table
  5. remainder_of_query;
  6. INSERT insert_clauses
  7. [{ /* +SHUFFLE */ | /* +NOSHUFFLE */ }]
  8. [/* +CLUSTERED */]
  9. SELECT remainder_of_query;
  10. UPSERT [{ /* +SHUFFLE */ | /* +NOSHUFFLE */ }]
  11. [/* +CLUSTERED */]
  12. upsert_clauses
  13. SELECT remainder_of_query;

Usage notes:

With both forms of hint syntax, include the STRAIGHT_JOIN keyword immediately after the SELECT and any DISTINCT or ALL keywords to prevent Impala from reordering the tables in a way that makes the join-related hints ineffective.

The STRAIGHT_JOIN hint affects the join order of table references in the query block containing the hint. It does not affect the join order of nested queries, such as views, inline views, or WHERE-clause subqueries. To use this hint for performance tuning of complex queries, apply the hint to all query blocks that need a fixed join order.

To reduce the need to use hints, run the COMPUTE STATS statement against all tables involved in joins, or used as the source tables for INSERT ... SELECT operations where the destination is a partitioned Parquet table. Do this operation after loading data or making substantial changes to the data within each table. Having up-to-date statistics helps Impala choose more efficient query plans without the need for hinting. See Table and Column Statistics for details and examples.

To see which join strategy is used for a particular query, examine the EXPLAIN output for that query. See Using the EXPLAIN Plan for Performance Tuning for details and examples.

Hints for join queries:

The /* +BROADCAST */ and /* +SHUFFLE */ hints control the execution strategy for join queries. Specify one of the following constructs immediately after the JOIN keyword in a query:

  • /* +SHUFFLE */ - Makes that join operation use the “partitioned” technique, which divides up corresponding rows from both tables using a hashing algorithm, sending subsets of the rows to other nodes for processing. (The keyword SHUFFLE is used to indicate a “partitioned join”, because that type of join is not related to “partitioned tables”.) Since the alternative “broadcast” join mechanism is the default when table and index statistics are unavailable, you might use this hint for queries where broadcast joins are unsuitable; typically, partitioned joins are more efficient for joins between large tables of similar size.
  • /* +BROADCAST */ - Makes that join operation use the “broadcast” technique that sends the entire contents of the right-hand table to all nodes involved in processing the join. This is the default mode of operation when table and index statistics are unavailable, so you would typically only need it if stale metadata caused Impala to mistakenly choose a partitioned join operation. Typically, broadcast joins are more efficient in cases where one table is much smaller than the other. (Put the smaller table on the right side of the JOIN operator.)

Hints for INSERT … SELECT and CREATE TABLE AS SELECT (CTAS):

When inserting into partitioned tables, such as using the Parquet file format, you can include a hint in the INSERT or CREATE TABLE AS SELECT(CTAS) statements to fine-tune the overall performance of the operation and its resource usage.

You would only use hints if an INSERT or CTAS into a partitioned table was failing due to capacity limits, or if such an operation was succeeding but with less-than-optimal performance.

  • /* +SHUFFLE */ and /* +NOSHUFFLE */ Hints

    • /* +SHUFFLE */ adds an exchange node, before writing the data, which re-partitions the result of the SELECT based on the partitioning columns of the target table. With this hint, only one node writes to a partition at a time, minimizing the global number of simultaneous writes and the number of memory buffers holding data for individual partitions. This also reduces fragmentation, resulting in fewer files. Thus it reduces overall resource usage of the INSERT or CTAS operation and allows some operations to succeed that otherwise would fail. It does involve some data transfer between the nodes so that the data files for a particular partition are all written on the same node.

      Use /* +SHUFFLE */ in cases where an INSERT or CTAS statement fails or runs inefficiently due to all nodes attempting to write data for all partitions.

      If the table is unpartitioned or every partitioning expression is constant, then /* +SHUFFLE */ will cause every write to happen on the coordinator node.

    • /* +NOSHUFFLE */ does not add exchange node before inserting to partitioned tables and disables re-partitioning. So the selected execution plan might be faster overall, but might also produce a larger number of small data files or exceed capacity limits, causing the INSERT or CTAS operation to fail.

      Impala automatically uses the /* +SHUFFLE */ method if any partition key column in the source table, mentioned in the SELECT clause, does not have column statistics. In this case, use the /* +NOSHUFFLE */ hint if you want to override this default behavior.

    • If column statistics are available for all partition key columns in the source table mentioned in the INSERT ... SELECT or CTAS query, Impala chooses whether to use the /* +SHUFFLE */ or /* +NOSHUFFLE */ technique based on the estimated number of distinct values in those columns and the number of nodes involved in the operation. In this case, you might need the /* +SHUFFLE */ or the /* +NOSHUFFLE */ hint to override the execution plan selected by Impala.

  • /* +CLUSTERED */ and /* +NOCLUSTERED */ Hints

    • /* +CLUSTERED */ sorts data by the partition columns before inserting to ensure that only one partition is written at a time per node. Use this hint to reduce the number of files kept open and the number of buffers kept in memory simultaneously. This technique is primarily useful for inserts into Parquet tables, where the large block size requires substantial memory to buffer data for multiple output files at once. This hint is available in Impala 2.8 or higher.

      Starting in Impala 3.0, /* +CLUSTERED */ is the default behavior for HDFS tables.

    • /* +NOCLUSTERED */ does not sort by primary key before insert. This hint is available in Impala 2.8 or higher.

      Use this hint when inserting to Kudu tables.

      In the versions lower than Impala 3.0, /* +NOCLUSTERED */ is the default in HDFS tables.

Starting from Impala 2.9, INSERT or UPSERT operations into Kudu tables automatically have an exchange and sort node added to the plan that partitions and sorts the rows according to the partitioning/primary key scheme of the target table (unless the number of rows to be inserted is small enough to trigger single node execution). Use the /* +NOCLUSTERED */ and /* +NOSHUFFLE */ hints together to disable partitioning and sorting before the rows are sent to Kudu.

Hints for scheduling of HDFS blocks:

The hints /* +SCHEDULE_CACHE_LOCAL */, /* +SCHEDULE_DISK_LOCAL */, and /* +SCHEDULE_REMOTE */ have the same effect as specifying the REPLICA_PREFERENCE query option with the respective option settings of CACHE_LOCAL, DISK_LOCAL, or REMOTE. The hint /* +RANDOM_REPLICA */ is the same as enabling the SCHEDULE_RANDOM_REPLICA query option.

You can use these hints in combination by separating them with commas, for example, /* +SCHEDULE_CACHE_LOCAL,RANDOM_REPLICA */. See REPLICA_PREFERENCE Query Option (Impala 2.7 or higher only) and SCHEDULE_RANDOM_REPLICA Query Option (Impala 2.5 or higher only) for information about how these settings influence the way Impala processes HDFS data blocks.

Specifying the replica preference as a query hint always overrides the query option setting. Specifying either the SCHEDULE_RANDOM_REPLICA query option or the corresponding RANDOM_REPLICA query hint enables the random tie-breaking behavior when processing data blocks during the query.

Suggestions versus directives:

In early Impala releases, hints were always obeyed and so acted more like directives. Once Impala gained join order optimizations, sometimes join queries were automatically reordered in a way that made a hint irrelevant. Therefore, the hints act more like suggestions in Impala 1.2.2 and higher.

To force Impala to follow the hinted execution mechanism for a join query, include the STRAIGHT_JOIN keyword in the SELECT statement. See Overriding Join Reordering with STRAIGHT_JOIN for details. When you use this technique, Impala does not reorder the joined tables at all, so you must be careful to arrange the join order to put the largest table (or subquery result set) first, then the smallest, second smallest, third smallest, and so on. This ordering lets Impala do the most I/O-intensive parts of the query using local reads on the DataNodes, and then reduce the size of the intermediate result set as much as possible as each subsequent table or subquery result set is joined.

Restrictions:

Queries that include subqueries in the WHERE clause can be rewritten internally as join queries. Currently, you cannot apply hints to the joins produced by these types of queries.

Because hints can prevent queries from taking advantage of new metadata or improvements in query planning, use them only when required to work around performance issues, and be prepared to remove them when they are no longer required, such as after a new Impala release or bug fix.

In particular, the /* +BROADCAST */ and /* +SHUFFLE */ hints are expected to be needed much less frequently in Impala 1.2.2 and higher, because the join order optimization feature in combination with the COMPUTE STATS statement now automatically choose join order and join mechanism without the need to rewrite the query and add hints. See Performance Considerations for Join Queries for details.

Compatibility:

The hints embedded within -- comments are compatible with Hive queries. The hints embedded within /* */ comments or [ ] square brackets are not recognized by or not compatible with Hive. For example, Hive raises an error for Impala hints within /* */ comments because it does not recognize the Impala hint names.

Considerations for views:

If you use a hint in the query that defines a view, the hint is preserved when you query the view. Impala internally rewrites all hints in views to use the -- comment notation, so that Hive can query such views without errors due to unrecognized hint names.

Examples:

For example, this query joins a large customer table with a small lookup table of less than 100 rows. The right-hand table can be broadcast efficiently to all nodes involved in the join. Thus, you would use the /* +broadcast */ hint to force a broadcast join strategy:

  1. select straight_join customer.address, state_lookup.state_name
  2. from customer join /* +broadcast */ state_lookup
  3. on customer.state_id = state_lookup.state_id;

This query joins two large tables of unpredictable size. You might benchmark the query with both kinds of hints and find that it is more efficient to transmit portions of each table to other nodes for processing. Thus, you would use the /* +shuffle */ hint to force a partitioned join strategy:

  1. select straight_join weather.wind_velocity, geospatial.altitude
  2. from weather join /* +shuffle */ geospatial
  3. on weather.lat = geospatial.lat and weather.long = geospatial.long;

For joins involving three or more tables, the hint applies to the tables on either side of that specific JOIN keyword. The STRAIGHT_JOIN keyword ensures that joins are processed in a predictable order from left to right. For example, this query joins t1 and t2 using a partitioned join, then joins that result set to t3 using a broadcast join:

  1. select straight_join t1.name, t2.id, t3.price
  2. from t1 join /* +shuffle */ t2 join /* +broadcast */ t3
  3. on t1.id = t2.id and t2.id = t3.id;

Related information:

For more background information about join queries, see Joins in Impala SELECT Statements. For performance considerations, see Performance Considerations for Join Queries.

Parent topic: Impala SQL Statements