impala+hdfs+csv格式文件

[创建目录]
hdfs dfs -mkdir -p /user/hdfs/sample_data/csv/device
hdfs dfs -mkdir -p /user/hdfs/sample_data/csv/metrics

[赋予权限]
sudo -u hdfs hadoop fs -chown -R impala:supergroup /user/hdfs/sample_data

[删除目录]
hdfs dfs -rm -r /user/hdfs/sample_data/csv

[上传文件]
hdfs dfs -put -f device.csv /user/hdfs/sample_data/csv/device
hdfs dfs -put -f metrics.csv /user/hdfs/sample_data/csv/metrics

[查看文件]
hdfs dfs -ls /user/hdfs/sample_data/csv/device

[impala建表]
DROP TABLE IF EXISTS device;
CREATE EXTERNAL TABLE device
(
deviceId STRING,
deviceName STRING,
orgId INT
)
ROW FORMAT DELIMITED FIELDS TERMINATED BY ','
LOCATION '/user/hdfs/sample_data/csv/device';

DROP TABLE IF EXISTS metrics;
CREATE EXTERNAL TABLE metrics
(
deviceId STRING,
reading INT,
time STRING
)
ROW FORMAT DELIMITED FIELDS TERMINATED BY ','
LOCATION '/user/hdfs/sample_data/csv/metrics';

[查询数据]
select
T_3C75F1.`deviceId`,
year(T_3C75F1.`time`),
month(T_3C75F1.`time`),
sum(T_3C75F1.`reading`),
count(1)
from (select device.deviceId,reading,metrics.time as time from device,metrics where device.deviceId=metrics.deviceId) as `T_3C75F1`
group by
T_3C75F1.`deviceId`,
year(T_3C75F1.`time`),
month(T_3C75F1.`time`);

耗时：device表1000条，metrics表1亿条（3.7GB）执行上面的查询语句，耗时平均180秒，

[刷新数据]
refresh device;
refresh metrics;

[问题]

1、执行查询时抛异常：
Memory limit exceeded: The memory limit is set too low to initialize spilling operator (id=2). The minimum required memory to spill this operator is 136.00 MB.
Error occurred on backend cdh4:22000 by fragment 5140dddc4be44c9e:5cccbfe200000004
Memory left in process limit: 384.00 KB
Process: Limit=256.00 MB Total=255.62 MB Peak=264.16 MB
RequestPool=root.root: Total=162.28 MB Peak=162.42 MB
Query(5140dddc4be44c9e:5cccbfe200000000): Total=162.28 MB Peak=162.42 MB
Fragment 5140dddc4be44c9e:5cccbfe200000006: Total=2.30 MB Peak=2.74 MB
AGGREGATION_NODE (id=6): Total=2.28 MB Peak=2.28 MB
Exprs: Total=4.00 KB Peak=4.00 KB
EXCHANGE_NODE (id=5): Total=0 Peak=0
DataStreamRecvr: Total=0 Peak=0
DataStreamSender (dst_id=7): Total=3.12 KB Peak=3.12 KB
CodeGen: Total=2.22 KB Peak=451.50 KB
Block Manager: Limit=156.00 MB Total=104.50 MB Peak=104.50 MB
Fragment 5140dddc4be44c9e:5cccbfe200000004: Total=132.81 MB Peak=132.95 MB
Runtime Filter Bank: Total=1.00 MB Peak=1.00 MB
AGGREGATION_NODE (id=3): Total=1.29 MB Peak=1.29 MB
Exprs: Total=8.00 KB Peak=8.00 KB
HASH_JOIN_NODE (id=2): Total=113.12 MB Peak=113.12 MB
Hash Join Builder (join_node_id=2): Total=113.02 MB Peak=113.02 MB
HDFS_SCAN_NODE (id=0): Total=0 Peak=0
EXCHANGE_NODE (id=4): Total=0 Peak=0
DataStreamRecvr: Total=17.37 MB Peak=28.05 MB
DataStreamSender (dst_id=5): Total=6.23 KB Peak=6.23 KB
CodeGen: Total=14.88 KB Peak=2.08 MB
Fragment 5140dddc4be44c9e:5cccbfe200000001: Total=27.18 MB Peak=43.25 MB
HDFS_SCAN_NODE (id=1): Total=27.13 MB Peak=43.20 MB
DataStreamSender (dst_id=4): Total=6.91 KB Peak=6.91 KB
CodeGen: Total=1.38 KB Peak=178.00 KB
RequestPool=root.hue: Total=0 Peak=1.76 MB
RequestPool=root.default: Total=0 Peak=7.40 MB
Untracked Memory: Total=93.34 MB

解决方案：csv文件的体积比较大，同样是1亿条数据，csv文件需要3.7G，而parquet只有261M。如果要是用csv文件，需要调整impala的参数mem_limit，应大于数据文件的大小。