Re: [Hive] Slow Loading Data Process with Parquet over 30k Partitions
Hi Slava, We would be interested in reviewing your patch. Can you please provide more details? Is there any other way to disable the partition creation step? Thanks, Chris On 4/13/15 10:59 PM, Slava Markeyev wrote: This is something I've encountered when doing ETL with hive and having it create 10's of thousands partitions. The issue is each partition needs to be added to the metastore and this is an expensive operation to perform. My work around was adding a flag to hive that optionally disables the metastore partition creation step. This may not be a solution for everyone as that table then has no partitions and you would have to run msck repair but depending on your use case, you may just want the data in hdfs. If there is interest in having this be an option I'll make a ticket and submit the patch. -Slava On Mon, Apr 13, 2015 at 10:40 PM, Xu, Cheng A cheng.a...@intel.com mailto:cheng.a...@intel.com wrote: Hi Tianqi, Can you attach hive.log as more detailed information? +Sergio __ __ Yours, Ferdinand Xu __ __ *From:*Tianqi Tong [mailto:tt...@brightedge.com mailto:tt...@brightedge.com] *Sent:* Friday, April 10, 2015 1:34 AM *To:* user@hive.apache.org mailto:user@hive.apache.org *Subject:* [Hive] Slow Loading Data Process with Parquet over 30k Partitions __ __ Hello Hive, I'm a developer using Hive to process TB level data, and I'm having some difficulty loading the data to table. I have 2 tables now: __ __ -- table_1: CREATE EXTERNAL TABLE `table_1`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition1 STRING) STORED AS RCfile __ __ -- table_2: CREATE EXTERNAL TABLE `table_2`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition2 STRING) STORED AS Parquet __ __ I'm doing an INSERT OVERWRITE to table_2 from SELECT FROM table_1 with dynamic partitioning, and the number of partitions grows dramatically from 1500 to 40k (because I want to use something else as partitioning). The mapreduce job was fine. Somehow the process stucked at Loading data to table default.table_2 (yearmonth=null, domain_prefix=null) , and I've been waiting for hours. __ __ Is this expected when we have 40k partitions? __ __ -- Refs - Here are the parameters that I used: export HADOOP_HEAPSIZE=16384 set PARQUET_FILE_SIZE=268435456; set parquet.block.size=268435456; set dfs.blocksize=268435456; set parquet.compression=SNAPPY; SET hive.exec.dynamic.partition.mode=nonstrict; SET hive.exec.max.dynamic.partitions=50; SET hive.exec.max.dynamic.partitions.pernode=5; SET hive.exec.max.created.files=100; __ __ __ __ Thank you very much! Tianqi Tong -- Slava Markeyev | Engineering | Upsight
Re: [Hive] Slow Loading Data Process with Parquet over 30k Partitions
I've created HIVE-10385 and attached a patch. Unit tests to come. -Slava On Fri, Apr 17, 2015 at 1:34 PM, Chris Roblee chr...@unity3d.com wrote: Hi Slava, We would be interested in reviewing your patch. Can you please provide more details? Is there any other way to disable the partition creation step? Thanks, Chris On 4/13/15 10:59 PM, Slava Markeyev wrote: This is something I've encountered when doing ETL with hive and having it create 10's of thousands partitions. The issue is each partition needs to be added to the metastore and this is an expensive operation to perform. My work around was adding a flag to hive that optionally disables the metastore partition creation step. This may not be a solution for everyone as that table then has no partitions and you would have to run msck repair but depending on your use case, you may just want the data in hdfs. If there is interest in having this be an option I'll make a ticket and submit the patch. -Slava On Mon, Apr 13, 2015 at 10:40 PM, Xu, Cheng A cheng.a...@intel.com mailto:cheng.a...@intel.com wrote: Hi Tianqi, Can you attach hive.log as more detailed information? +Sergio __ __ Yours, Ferdinand Xu __ __ *From:*Tianqi Tong [mailto:tt...@brightedge.com mailto: tt...@brightedge.com] *Sent:* Friday, April 10, 2015 1:34 AM *To:* user@hive.apache.org mailto:user@hive.apache.org *Subject:* [Hive] Slow Loading Data Process with Parquet over 30k Partitions __ __ Hello Hive, I'm a developer using Hive to process TB level data, and I'm having some difficulty loading the data to table. I have 2 tables now: __ __ -- table_1: CREATE EXTERNAL TABLE `table_1`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition1 STRING) STORED AS RCfile __ __ -- table_2: CREATE EXTERNAL TABLE `table_2`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition2 STRING) STORED AS Parquet __ __ I'm doing an INSERT OVERWRITE to table_2 from SELECT FROM table_1 with dynamic partitioning, and the number of partitions grows dramatically from 1500 to 40k (because I want to use something else as partitioning). The mapreduce job was fine. Somehow the process stucked at Loading data to table default.table_2 (yearmonth=null, domain_prefix=null) , and I've been waiting for hours. __ __ Is this expected when we have 40k partitions? __ __ -- Refs - Here are the parameters that I used: export HADOOP_HEAPSIZE=16384 set PARQUET_FILE_SIZE=268435456; set parquet.block.size=268435456; set dfs.blocksize=268435456; set parquet.compression=SNAPPY; SET hive.exec.dynamic.partition.mode=nonstrict; SET hive.exec.max.dynamic.partitions=50; SET hive.exec.max.dynamic.partitions.pernode=5; SET hive.exec.max.created.files=100; __ __ __ __ Thank you very much! Tianqi Tong -- Slava Markeyev | Engineering | Upsight -- Slava Markeyev | Engineering | Upsight http://www.linkedin.com/in/slavamarkeyev http://www.linkedin.com/in/slavamarkeyev
Re: [Hive] Slow Loading Data Process with Parquet over 30k Partitions
This is something I've encountered when doing ETL with hive and having it create 10's of thousands partitions. The issue is each partition needs to be added to the metastore and this is an expensive operation to perform. My work around was adding a flag to hive that optionally disables the metastore partition creation step. This may not be a solution for everyone as that table then has no partitions and you would have to run msck repair but depending on your use case, you may just want the data in hdfs. If there is interest in having this be an option I'll make a ticket and submit the patch. -Slava On Mon, Apr 13, 2015 at 10:40 PM, Xu, Cheng A cheng.a...@intel.com wrote: Hi Tianqi, Can you attach hive.log as more detailed information? +Sergio Yours, Ferdinand Xu *From:* Tianqi Tong [mailto:tt...@brightedge.com] *Sent:* Friday, April 10, 2015 1:34 AM *To:* user@hive.apache.org *Subject:* [Hive] Slow Loading Data Process with Parquet over 30k Partitions Hello Hive, I'm a developer using Hive to process TB level data, and I'm having some difficulty loading the data to table. I have 2 tables now: -- table_1: CREATE EXTERNAL TABLE `table_1`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition1 STRING) STORED AS RCfile -- table_2: CREATE EXTERNAL TABLE `table_2`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition2 STRING) STORED AS Parquet I'm doing an INSERT OVERWRITE to table_2 from SELECT FROM table_1 with dynamic partitioning, and the number of partitions grows dramatically from 1500 to 40k (because I want to use something else as partitioning). The mapreduce job was fine. Somehow the process stucked at Loading data to table default.table_2 (yearmonth=null, domain_prefix=null) , and I've been waiting for hours. Is this expected when we have 40k partitions? -- Refs - Here are the parameters that I used: export HADOOP_HEAPSIZE=16384 set PARQUET_FILE_SIZE=268435456; set parquet.block.size=268435456; set dfs.blocksize=268435456; set parquet.compression=SNAPPY; SET hive.exec.dynamic.partition.mode=nonstrict; SET hive.exec.max.dynamic.partitions=50; SET hive.exec.max.dynamic.partitions.pernode=5; SET hive.exec.max.created.files=100; Thank you very much! Tianqi Tong -- Slava Markeyev | Engineering | Upsight
Re: [Hive] Slow Loading Data Process with Parquet over 30k Partitions
That is too many partitions. Way to much overhead in anything that has that many partitions. On Tue, Apr 14, 2015 at 12:53 PM, Tianqi Tong tt...@brightedge.com wrote: Hi Slava and Ferdinand, Thanks for the reply! Later when I was looking at the hive.log, I found Hive was indeed calculating the partition stats, and the log looks like: …. 2015-04-14 09:38:21,146 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(296)) - Updating partition stats fast for: parquet_table 2015-04-14 09:38:21,147 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(299)) - Updated size to 5533480 2015-04-14 09:38:44,511 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(296)) - Updating partition stats fast for: parquet_table 2015-04-14 09:38:44,512 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(299)) - Updated size to 66246 2015-04-14 09:39:07,554 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(296)) - Updating partition stats fast for: parquet_table 2015-04-14 09:39:07,555 WARN [main]: hive.log (MetaStoreUtils.java:updatePartitionStatsFast(299)) - Updated size to 418925 …. One interesting thing is, it's getting slower and slower. Right after I launched the job, it took less than 1s to calculate for one partition. Now it's taking 20+s for each one. I tried hive.stats.autogather=false, but somehow it didn't seem to work. I also ended up hard coding a little bit to the Hive source code. In my case, I have around 4 partitions with one file (varies from 1M to 1G) in each of them. Now it's been 4 days and the first job I launched is still not done yet, with partition stats. Thanks Tianqi Tong *From:* Slava Markeyev [mailto:slava.marke...@upsight.com] *Sent:* Monday, April 13, 2015 11:00 PM *To:* user@hive.apache.org *Cc:* Sergio Pena *Subject:* Re: [Hive] Slow Loading Data Process with Parquet over 30k Partitions This is something I've encountered when doing ETL with hive and having it create 10's of thousands partitions. The issue is each partition needs to be added to the metastore and this is an expensive operation to perform. My work around was adding a flag to hive that optionally disables the metastore partition creation step. This may not be a solution for everyone as that table then has no partitions and you would have to run msck repair but depending on your use case, you may just want the data in hdfs. If there is interest in having this be an option I'll make a ticket and submit the patch. -Slava On Mon, Apr 13, 2015 at 10:40 PM, Xu, Cheng A cheng.a...@intel.com wrote: Hi Tianqi, Can you attach hive.log as more detailed information? +Sergio Yours, Ferdinand Xu *From:* Tianqi Tong [mailto:tt...@brightedge.com] *Sent:* Friday, April 10, 2015 1:34 AM *To:* user@hive.apache.org *Subject:* [Hive] Slow Loading Data Process with Parquet over 30k Partitions Hello Hive, I'm a developer using Hive to process TB level data, and I'm having some difficulty loading the data to table. I have 2 tables now: -- table_1: CREATE EXTERNAL TABLE `table_1`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition1 STRING) STORED AS RCfile -- table_2: CREATE EXTERNAL TABLE `table_2`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition2 STRING) STORED AS Parquet I'm doing an INSERT OVERWRITE to table_2 from SELECT FROM table_1 with dynamic partitioning, and the number of partitions grows dramatically from 1500 to 40k (because I want to use something else as partitioning). The mapreduce job was fine. Somehow the process stucked at Loading data to table default.table_2 (yearmonth=null, domain_prefix=null) , and I've been waiting for hours. Is this expected when we have 40k partitions? -- Refs - Here are the parameters that I used: export HADOOP_HEAPSIZE=16384 set PARQUET_FILE_SIZE=268435456; set parquet.block.size=268435456; set dfs.blocksize=268435456; set parquet.compression=SNAPPY; SET hive.exec.dynamic.partition.mode=nonstrict; SET hive.exec.max.dynamic.partitions=50; SET hive.exec.max.dynamic.partitions.pernode=5; SET hive.exec.max.created.files=100; Thank you very much! Tianqi Tong -- Slava Markeyev | Engineering | Upsight
RE: [Hive] Slow Loading Data Process with Parquet over 30k Partitions
Hi Tianqi, Can you attach hive.log as more detailed information? +Sergio Yours, Ferdinand Xu From: Tianqi Tong [mailto:tt...@brightedge.com] Sent: Friday, April 10, 2015 1:34 AM To: user@hive.apache.org Subject: [Hive] Slow Loading Data Process with Parquet over 30k Partitions Hello Hive, I'm a developer using Hive to process TB level data, and I'm having some difficulty loading the data to table. I have 2 tables now: -- table_1: CREATE EXTERNAL TABLE `table_1`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition1 STRING) STORED AS RCfile -- table_2: CREATE EXTERNAL TABLE `table_2`( `keyword` string, `domain` string, `url` string ) PARTITIONED BY (yearmonth INT, partition2 STRING) STORED AS Parquet I'm doing an INSERT OVERWRITE to table_2 from SELECT FROM table_1 with dynamic partitioning, and the number of partitions grows dramatically from 1500 to 40k (because I want to use something else as partitioning). The mapreduce job was fine. Somehow the process stucked at Loading data to table default.table_2 (yearmonth=null, domain_prefix=null) , and I've been waiting for hours. Is this expected when we have 40k partitions? -- Refs - Here are the parameters that I used: export HADOOP_HEAPSIZE=16384 set PARQUET_FILE_SIZE=268435456; set parquet.block.size=268435456; set dfs.blocksize=268435456; set parquet.compression=SNAPPY; SET hive.exec.dynamic.partition.mode=nonstrict; SET hive.exec.max.dynamic.partitions=50; SET hive.exec.max.dynamic.partitions.pernode=5; SET hive.exec.max.created.files=100; Thank you very much! Tianqi Tong