Revised design for Spillable Mem Channel
Review Request #12060 - Created June 24, 2013 and updated
Revised design for Spillable Mem Channel. We no longer have Spillable channel config pointing to another channel (by name) as in the previous design. Spillable Channel instead derives from FileChannel (as per https://issues.apache.org/jira/browse/FLUME-1227?focusedCommentId=13628201&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13628201) Essence of this design: - SC derives from File channel and maintains an in memory queue. If memory queue is full, events are sent to disk overflow (i.e. File channel). - SC maintains a 'Drain-Order' queue (DOQ) for remembering the order in which the incoming events were interleaved between main memory and - Put transaction: All the elements in the putList are committed to mem queue if it has space, else written to disk (ie file channel). Head of DOQ is updated to indicate where the elements put. - Take transaction : Tail of DOQ is consulted to determine whether the next set of events are to be taken from the memory queue or from disk overflow. DOQ's tail is updated after events are taken out. SC Configuration: - Accepts all the File Channel settings - Introduces one additional setting: 'memoryCapacity' which indicates the number of items it can hold in memory Sample config ... a1.channels = c1 a1.sinks = logger a1.sources = src a1.sources.src.type = exec a1.sources.src.command = seq 1 100000 a1.sources.src.batchSize = 10 a1.sources.src.channels = c1 a1.sinks.logger.type = logger a1.sinks.logger.channel = c1 a1.channels.c1.type = spillablememory a1.channels.c1.checkpointDir = /tmp/flume/checkpoint a1.channels.c1.dataDirs = /tmp/flume/data a1.channels.c1.memoryCapacity = 10 a1.channels.c1.keep-alive = 2
Wrote a set of Unit tests. A few are failing and need to be finished up.
Diff revision 3