[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13030224#comment-13030224 ] Hudson commented on HBASE-3721: --- Integrated in HBase-TRUNK #1909 (See [https://builds.apache.org/hudson/job/HBase-TRUNK/1909/]) Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Fix For: 0.92.0 Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721-v6.patch, 3721.txt, LoadIncrementalHFiles.java From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. New configuration parameter hbase.loadincremental.threads.max is introduced which sets the maximum number of threads for parallel bulk load. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13029694#comment-13029694 ] Ted Yu commented on HBASE-3721: --- From Adam: I did a number of runs of loading a single set of HFiles with and without the patch, and it does seem the patch improves the load speed. I'll need to run more extensively to get accurate numbers, but with the patch I'm seeing ranges from 3-7 minutes vs 5-11 without the patch. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721-v6.patch, 3721.txt, LoadIncrementalHFiles.java From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. New configuration parameter hbase.loadincremental.threads.max is introduced which sets the maximum number of threads for parallel bulk load. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13029728#comment-13029728 ] stack commented on HBASE-3721: -- @Ted v6 patch is what he tested? Is that what I should commit? Thanks. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721-v6.patch, 3721.txt, LoadIncrementalHFiles.java From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. New configuration parameter hbase.loadincremental.threads.max is introduced which sets the maximum number of threads for parallel bulk load. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13029732#comment-13029732 ] Ted Yu commented on HBASE-3721: --- Version 6 is what Adam used. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721-v6.patch, 3721.txt, LoadIncrementalHFiles.java From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. New configuration parameter hbase.loadincremental.threads.max is introduced which sets the maximum number of threads for parallel bulk load. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028458#comment-13028458 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review639 --- Does it work? If it does, I'm good w/ applying it. There are some questions in the below. See what you think Ted. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1284 Nothing is done w/ the result here. Should it be logged or something? /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1285 There are a bunch of these in this patch... white space. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1286 Will multiple threads be trying to get a unique name at the same time? Is this a good enough 'unique' name -- table name and incrementing number? Is this per unique table-based name to isolate thread writes to the fs? - Michael On 2011-04-29 20:48:41, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-04-29 20:48:41) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1097897 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028475#comment-13028475 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. Does it work? If it does, I'm good w/ applying it. There are some questions in the below. See what you think Ted. I ran unit tests (TestHFileOutputFormat and TestLoadIncrementalHFiles) on my patch. bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java, line 212 bq. https://reviews.apache.org/r/572/diff/2/?file=17706#file17706line212 bq. bq. Nothing is done w/ the result here. Should it be logged or something? The return type is Void. I do log errors. bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java, line 233 bq. https://reviews.apache.org/r/572/diff/2/?file=17706#file17706line233 bq. bq. There are a bunch of these in this patch... white space. Will remove white spaces in next patch. bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java, line 235 bq. https://reviews.apache.org/r/572/diff/2/?file=17706#file17706line235 bq. bq. Will multiple threads be trying to get a unique name at the same time? Is this a good enough 'unique' name -- table name and incrementing number? Is this per unique table-based name to isolate thread writes to the fs? I changed regionCount to AtomicLong. The unique name is to isolate writes to fs from different threads. - Ted --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review639 --- On 2011-04-29 20:48:41, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-04-29 20:48:41) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1097897 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028479#comment-13028479 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/ --- (Updated 2011-05-03 22:28:11.346203) Review request for hbase and Todd Lipcon. Changes --- Cleaned up white spaces. Summary --- I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. I added the following method to HConnection/HConnectionManager: public T void getRegionServerWithRetries(ExecutorService pool, ListServerCallableT callables, Object[] results) This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. This addresses bug HBASE-3721. https://issues.apache.org/jira/browse/HBASE-3721 Diffs (updated) - /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1099118 Diff: https://reviews.apache.org/r/572/diff Testing --- TestLoadIncrementalHFiles and TestHFileOutputFormat pass. Thanks, Ted Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028586#comment-13028586 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java, line 212 bq. https://reviews.apache.org/r/572/diff/2/?file=17706#file17706line212 bq. bq. Nothing is done w/ the result here. Should it be logged or something? bq. bq. Ted Yu wrote: bq. The return type is Void. bq. I do log errors. OK. Makes sense. - Michael --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review639 --- On 2011-05-03 22:28:11, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-05-03 22:28:11) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1099118 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028587#comment-13028587 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. Does it work? If it does, I'm good w/ applying it. There are some questions in the below. See what you think Ted. bq. bq. Ted Yu wrote: bq. I ran unit tests (TestHFileOutputFormat and TestLoadIncrementalHFiles) on my patch. I was more asking if you'd loaded up any files with. Do you think we should get someone like Adam Portley to try it? Will I ask them? They are the ones who wanted this originally complaining about slow load speed, is that right? Good on you Ted. - Michael --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review639 --- On 2011-05-03 22:28:11, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-05-03 22:28:11) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1099118 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028591#comment-13028591 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- bq. On 2011-05-03 21:51:39, Michael Stack wrote: bq. Does it work? If it does, I'm good w/ applying it. There are some questions in the below. See what you think Ted. bq. bq. Ted Yu wrote: bq. I ran unit tests (TestHFileOutputFormat and TestLoadIncrementalHFiles) on my patch. bq. bq. Michael Stack wrote: bq. I was more asking if you'd loaded up any files with. Do you think we should get someone like Adam Portley to try it? Will I ask them? They are the ones who wanted this originally complaining about slow load speed, is that right? Good on you Ted. I hope Adam would try out the latest patch. - Ted --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review639 --- On 2011-05-03 22:28:11, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-05-03 22:28:11) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1099118 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13028608#comment-13028608 ] stack commented on HBASE-3721: -- I asked him in private mail. Thanks Ted. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13027178#comment-13027178 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/ --- (Updated 2011-04-29 20:48:41.082584) Review request for hbase and Todd Lipcon. Changes --- Simplified the changes for this JIRA according to Todd's review. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. Summary --- I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. I added the following method to HConnection/HConnectionManager: public T void getRegionServerWithRetries(ExecutorService pool, ListServerCallableT callables, Object[] results) This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. This addresses bug HBASE-3721. https://issues.apache.org/jira/browse/HBASE-3721 Diffs (updated) - /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1097897 Diff: https://reviews.apache.org/r/572/diff Testing --- TestLoadIncrementalHFiles and TestHFileOutputFormat pass. Thanks, Ted Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13026146#comment-13026146 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review597 --- /src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java https://reviews.apache.org/r/572/#comment1242 I don't necessarily see the point of this being in HConnectionManager, since it's just a general submission of a bunch of callables. If HConnectionManager did something smarter so that threads sleeping for retry didn't lock up a thread to sleep, it would make more sense here. /src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java https://reviews.apache.org/r/572/#comment1243 since these arrays only get filled in in the case that there are exceptions, we can expect they'll be very small in general, and the extra code to track actionCount isn't likely to make any difference. It's just a micro-optimization for something that isn't a bottleneck. /src/main/java/org/apache/hadoop/hbase/client/HTable.java https://reviews.apache.org/r/572/#comment1244 why is this public? /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1246 rather than working in batches, why not just have one thread which is submitting tasks to the executor, and another thread which is pulling off completed ones? ie I don't see the point of batch size at all instead of something more fluid - a normal producer/consumer kind of design. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1245 i don't like this dependency. Check out Guava's ThreadFactoryBuilder. - Todd On 2011-04-09 14:00:23, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-04-09 14:00:23) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq./src/main/java/org/apache/hadoop/hbase/client/HConnection.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HTable.java 1090500 bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1090500 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13026849#comment-13026849 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review611 --- /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1255 Your suggestion is good. Looking at existing usage of ExecutorService in SplitTransaction and HConnectionManager, we submit all Callable's and then wait for them to complete. I will try to use some data structure, such as Mapbyte[], Future, so that producer and consumer work concurrently. - Ted On 2011-04-09 14:00:23, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-04-09 14:00:23) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq./src/main/java/org/apache/hadoop/hbase/client/HConnection.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HTable.java 1090500 bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1090500 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13026854#comment-13026854 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/#review612 --- /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java https://reviews.apache.org/r/572/#comment1256 Another way is to keep submitting Callable's in doBulkLoad() and save its Future. When there is no more Callable to submit, we call Future.get() for each saved Future. - Ted On 2011-04-09 14:00:23, Ted Yu wrote: bq. bq. --- bq. This is an automatically generated e-mail. To reply, visit: bq. https://reviews.apache.org/r/572/ bq. --- bq. bq. (Updated 2011-04-09 14:00:23) bq. bq. bq. Review request for hbase and Todd Lipcon. bq. bq. bq. Summary bq. --- bq. bq. I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. bq. I added the following method to HConnection/HConnectionManager: bq. public T void getRegionServerWithRetries(ExecutorService pool, bq. ListServerCallableT callables, Object[] results) bq. This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). bq. bq. I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size bq. hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. bq. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. bq. bq. bq. This addresses bug HBASE-3721. bq. https://issues.apache.org/jira/browse/HBASE-3721 bq. bq. bq. Diffs bq. - bq. bq./src/main/java/org/apache/hadoop/hbase/client/HConnection.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java 1090500 bq./src/main/java/org/apache/hadoop/hbase/client/HTable.java 1090500 bq. /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1090500 bq. bq. Diff: https://reviews.apache.org/r/572/diff bq. bq. bq. Testing bq. --- bq. bq. TestLoadIncrementalHFiles and TestHFileOutputFormat pass. bq. bq. bq. Thanks, bq. bq. Ted bq. bq. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism on multi-core computer. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13017867#comment-13017867 ] jirapos...@reviews.apache.org commented on HBASE-3721: -- --- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/572/ --- Review request for hbase and Todd Lipcon. Summary --- I refactored LoadIncrementalHFiles so that tryLoad() queues work items in ListServerCallableVoid. doBulkLoad() periodically sends batch of ServerCallable's to HBase cluster. I added the following method to HConnection/HConnectionManager: public T void getRegionServerWithRetries(ExecutorService pool, ListServerCallableT callables, Object[] results) This method uses thread pool to send multiple ServerCallable's through getRegionServerWithRetries(ServerCallableT callable). I introduced two new config parameters: hbase.loadincremental.threads.max and hbase.loadincremental.batch.size hbase.loadincremental.batch.size is for configuring the batch size above which HConnection.getRegionServerWithRetries() would be called. In Adam's case, there're many small HFiles. LoadIncrementalHFiles shouldn't wait until all HFiles have been scanned. hbase.loadincremental.threads.max controls the maximum number of threads in thread pool. This addresses bug HBASE-3721. https://issues.apache.org/jira/briwse/HBASE-3721 Diffs - /src/main/java/org/apache/hadoop/hbase/client/HConnection.java 1090500 /src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java 1090500 /src/main/java/org/apache/hadoop/hbase/client/HTable.java 1090500 /src/main/java/org/apache/hadoop/hbase/mapreduce/LoadIncrementalHFiles.java 1090500 Diff: https://reviews.apache.org/r/572/diff Testing --- TestLoadIncrementalHFiles and TestHFileOutputFormat pass. Thanks, Ted Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu Assignee: Ted Yu Attachments: 3721-v2.txt, 3721-v3.txt, 3721-v4.txt, 3721.txt From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (HBASE-3721) Speedup LoadIncrementalHFiles
[ https://issues.apache.org/jira/browse/HBASE-3721?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13015041#comment-13015041 ] Ted Yu commented on HBASE-3721: --- LoadIncrementalHFiles may split StoreFile. The above proposal only works if there is no such splitting. Speedup LoadIncrementalHFiles - Key: HBASE-3721 URL: https://issues.apache.org/jira/browse/HBASE-3721 Project: HBase Issue Type: Improvement Components: util Reporter: Ted Yu From Adam Phelps: from the logs it looks like 1% of the hfiles we're loading have to be split. Looking at the code for LoadIncrementHFiles (hbase v0.90.1), I'm actually thinking our problem is that this code loads the hfiles sequentially. Our largest table has over 2500 regions and the data being loaded is fairly well distributed across them, so there end up being around 2500 HFiles for each load period. At 1-2 seconds per HFile that means the loading process is very time consuming. Currently server.bulkLoadHFile() is a blocking call. We can utilize ExecutorService to achieve better parallelism. -- This message is automatically generated by JIRA. For more information on JIRA, see: http://www.atlassian.com/software/jira