-Re: HBase Sink Reliability
Jeff Lord 2013-04-25, 16:58
Mike Percy contributed a most excellent blog post on this topic.
Have you had a chance to read over this?
Tuning the batch size trades throughput vs. latency and duplication under
failure. With a small batch size, throughput decreases, but the risk of
event duplication is reduced if a failure were to occur. With a large batch
size, you get much better throughput, but increased latency, and in the
case of a transaction failure, the number of possible duplicates increases."
On Wed, Apr 24, 2013 at 11:04 PM, David Quigley <[EMAIL PROTECTED]>wrote:
> Thanks all,
> Added a dedicated channel for hdfs and hbase and everything all events are
> making it into their sinks now.
> What is the best tuning strategy for getting events from an exec source ->
> avro sink -> avro source -> hbase sink with the least amount of latency?
> Will batch size and transaction size have any effect on this latency?
> Thanks again
> On Mon, Apr 22, 2013 at 10:58 AM, Israel Ekpo <[EMAIL PROTECTED]> wrote:
>> In addition to what has already been said, if you take a look at your
>> flume log files, you should be able to see exception messages that explain
>> why this is happening.
>> On 22 April 2013 11:11, David Quigley <[EMAIL PROTECTED]> wrote:
>>> I am using flume to write events from webserver to both HDFS and HBase.
>>> All events are being written to HDFS but only about half are making it into
>>> HBase. Is there anything in my configurations which would be causing the
>>> issue? I have both HDFS and HBase sink reading from the same File Channel.
>>> Is it better to have one channel per sink?
>>> # flume config on web server
>>> agent.sources = sourceLog
>>> agent.sources.sourceLog.type = exec
>>> agent.sources.sourceLog.command = tail -F /var/log/clickServer/clicks_out
>>> agent.sources.sourceLog.batchSize = 100
>>> agent.sources.sourceLog.channels = fileChannel
>>> agent.sources.sourceLog.interceptors = itime ihost idatatype idataparent
>>> agent.sources.sourceLog.interceptors.itime.type = timestamp
>>> agent.sources.sourceLog.interceptors.ihost.type = host
>>> agent.sources.sourceLog.interceptors.ihost.useIP = false
>>> agent.sources.sourceLog.interceptors.ihost.hostHeader = host
>>> agent.sources.sourceLog.interceptors.idatatype.type = static
>>> agent.sources.sourceLog.interceptors.idatatype.key = data_type
>>> agent.sources.sourceLog.interceptors.idatatype.value = clicks
>>> agent.sources.sourceLog.interceptors.idataparent.type = static
>>> agent.sources.sourceLog.interceptors.idataparent.key = data_parent
>>> agent.sources.sourceLog.interceptors.idataparent.value = *
>>> agent.channels = fileChannel
>>> agent.channels.fileChannel.type = file
>>> agent.channels.fileChannel.transactionCapacity = 100
>>> agent.channels.fileChannel.checkpointDir >>> /opt/flume/file-channel/checkpoint
>>> agent.channels.fileChannel.dataDirs = /opt/flume/file-channel/data
>>> agent.sinks = AvroSink_main AvroSink_backup_1 AvroSink_backup_2
>>> agent.sinks.AvroSink_main.type = avro
>>> agent.sinks.AvroSink_main.channel = fileChannel
>>> agent.sinks.AvroSink_main.hostname = *
>>> agent.sinks.AvroSink_main.port = 35873
>>> agent.sinks.AvroSink_main.batchSize = 100
>>> agent.sinks.AvroSink_backup_1.type = avro
>>> agent.sinks.AvroSink_backup_1.channel = fileChannel
>>> agent.sinks.AvroSink_backup_1.hostname = *
>>> agent.sinks.AvroSink_backup_1.port = 35873
>>> agent.sinks.AvroSink_backup_1.batchSize = 100
>>> agent.sinks.AvroSink_backup_2.type = avro
>>> agent.sinks.AvroSink_backup_2.channel = fileChannel
>>> agent.sinks.AvroSink_backup_2.hostname = *
>>> agent.sinks.AvroSink_backup_2.port = 35873
>>> agent.sinks.AvroSink_backup_2.batchSize = 100
>>> agent.sinks.AvroSink_backup_3.type = avro
>>> agent.sinks.AvroSink_backup_3.channel = fileChannel
>>> agent.sinks.AvroSink_backup_3.hostname = *