Impala队列内存参数分析

问题

对Impala队列内存的几个参数分析了下，欢迎指正

队列资源池的几个内存配置

Maximum Query Memory Limit

某个队列资源池，一个查询在一个Impala节点上下执行需要的最小内存
Minimum Query Memory Limit

某个队列资源池，一个查询在一个Impala节点上下执行需要的最大内存
最大内存

可用于此池中执行的所有查询的最大内存

给一个Impala队列提交查询时，Impala如何判断是否接受查询请求

实验sql

 set request_pool = hqueue;

 select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

查询sql分析

[ip:21000] testdb> explain select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: explain select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

+------------------------------------------------------------------------------------+

| Explain String                                                                     |

+------------------------------------------------------------------------------------+

| Max Per-Host Resource Reservation: Memory=8.00MB Threads=3                         |

| Per-Host Resource Estimates: Memory=256MB                                          |

| WARNING: The following tables are missing relevant table and/or column statistics. |

| testdb.testtable                                              |

|                                                                                    |

| PLAN-ROOT SINK                                                                     |

| |                                                                                  |

| 02:MERGING-EXCHANGE [UNPARTITIONED]                                                |

| |  offset: 5                                                                       |

| |  order by: acctset_code ASC                                                      |

| |  limit: 5                                                                        |

| |                                                                                  |

| 01:TOP-N [LIMIT=10]                                                                |

| |  order by: acctset_code ASC                                                      |

| |  row-size=22B cardinality=10                                                     |

| |                                                                                  |

| 00:SCAN HDFS [testdb.testtable]                               |

|    partitions=138/138 files=140 size=808.62MB                                      |

|    predicates: acctset_code = '00001'                                              |

|    row-size=22B cardinality=16                                                     |

+------------------------------------------------------------------------------------+

注意的地方：这里面有个单节点需要内存值256M，不过Impala估算的不一定准确。

实验1

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maximum Query Memory Limit
root.hqueue	500M	260M	270M

提交结果：

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 10:54:55 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=f54d764cf100d474:a89eec5c00000000

ERROR: Rejected query from pool root.hqueue: request memory needed 780.00 MB is greater than pool max mem resources 500.00 MB.

猜测是因为：260M（查询最小内存） * 3 =780M > 500M

实验2

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maxmum Query Memory Limit
root.hqueue	500M	250M	270M

提交结果：

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 10:58:28 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=39423b17b20dc603:66c4de7400000000

ERROR: Rejected query from pool root.hqueue: request memory needed 768.23 MB is greater than pool max mem resources 500.00 MB.

猜测是因为：256M（查询计划里面估计的单节点内存） * 3 = 768M > 500M，综合实验1和实验2，估计Impala在判断查询是否会超内存时，对估计值和Minimum Query Memory Limit参数，会有个 Max(估计值,Minimum Query Memory Limit)操作，在实验1中，即Max(256M,260M)，实验2中，即Max(256,250)。

实验3

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maxmum Query Memory Limit
root.hqueue	500M	250M	252M

提交结果：

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 11:09:42 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=e24e74d387c201b5:9e72143600000000

ERROR: Rejected query from pool root.hqueue: request memory needed 756.00 MB is greater than pool max mem resources 500.00 MB

猜测是因为：252M * 3 = 756M > 500M，结合实验2，估计Impala在判断查询是否会超内存时，对于Maxmum Query Memory Limit参数，会有个Min操作，即Min(Max(估计值,Minimum Query Memory Limit),Maxmum Query Memory Limit)，在本例中，即Min(Max(256M,250M),252M)

实验4

mem_limit：指定查询每个节点需要的内存

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maxmum Query Memory Limit
root.hqueue	500M	100M	200M

[ip:21000] testdb> set mem_limit=170M;

MEM_LIMIT set to 170M

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 13:53:31 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=ba4fa4a44d2dac9d:b24a60d600000000

ERROR: Rejected query from pool root.hqueue: request memory needed 510.00 MB is greater than pool max mem resources 500.00 MB.

[ip:21000] testdb> set mem_limit=210M;

MEM_LIMIT set to 210M

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 13:54:07 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=ca49acba3c002727:2d69557a00000000

ERROR: Rejected query from pool root.hqueue: request memory needed 600.00 MB is greater than pool max mem resources 500.00 MB

分析：mem_limit=170M时，Min(Max(170,100),200) * 3 = 510M > 500M；mem_limit=210M时，Min(Max(210,100),200) * 3 = 600M > 500；猜测，指定mem_limit时，Impala会使用mem_limit值来代替自己估计的内存使用值，并结合Minimum Query Memory Limit和Maxmum Query Memory Limit来判断内存是否会超过最大内存，从而决定是否拒绝查询请求。

实验5

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maxmum Query Memory Limit
root.hqueue	500M	39M	39M

[ip:21000] testdb> select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5;

Query: select tally_id, acctset_code  from testtable where acctset_code='00001'order by acctset_code limit 5 offset 5

Query submitted at: 2020-06-23 15:26:42 (Coordinator: http://ip:25000)

Query progress can be monitored at: http://ip:25000/query_plan?query_id=234ca270d3731d06:9980e6fd00000000

ERROR: Rejected query from pool root.hqueue: minimum memory reservation is greater than memory available to the query for buffer reservations. Memory reservation needed given the current plan: 8.00 MB. Adjust either the mem_limit or the pool config (max-query-mem-limit, min-query-mem-limit) for the query to allow the query memory limit to be at least 40.00 MB. Note that changing the mem_limit may also change the plan. See the query profile for more information about the per-node memory requirements.

以下配置，查询成功提交并执行

Left-Aligned	Left-Aligned	Left-Aligned	Left-Aligned
队列名称	最大内存	Minimum Query Memory Limit	Maxmum Query Memory Limit
root.hqueue	500M	40M	40M

分析：max-query-mem-limit, min-query-mem-limit，不能设置的太小，测试环境中，单个节点最少需要40M

结论

当查询指定mem_limit，以下条件拒绝提交查询，报内存不够

__Min(Max(mem_limit,Minimum Query Memory Limit),Maxmum Query Memory Limit) * 节点数 __> 最大内存
未指定mem_limit，以下条件拒绝提交查询，报内存不够，估计值可以通过explain获得，不过Impala估计的不准

__Min(Max(估计值,Minimum Query Memory Limit),Maxmum Query Memory Limit) * 节点数 __> 最大内存
max-query-mem-limit, min-query-mem-limit，不能设置的太小，测试环境中，单个节点最少需要40M

建议

配置Maxmum Query Memory Limit * 节点数 <= 最大内存，查询应该不会被reject
若队列资源池中没有配置Minimum Query Memory Limit和Maxmum Query Memory Limit参数，那么从之前的结论也可以看出，Impala会根据__估计值 * 节点数__ 是否大于最大内存来判断是否拒绝该查询，但因为Impala估计出的单节点内存上限值很不准确，所以这种情况，可以通过 set mem_limit = XXM，人为设置一个合理的大小，后续Impala会根据__mem_limit__ * 节点数来判断是否会超过最大内存