[jira] [Commented] (HBASE-12217) System load average based client pushback
[ https://issues.apache.org/jira/browse/HBASE-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16111993#comment-16111993 ] stack commented on HBASE-12217: --- Unscheduling good idea that has not seen progress in a while. > System load average based client pushback > - > > Key: HBASE-12217 > URL: https://issues.apache.org/jira/browse/HBASE-12217 > Project: HBase > Issue Type: New Feature >Reporter: Andrew Purtell >Assignee: Andrew Purtell > > If a RegionServer host is already heavily loaded* then it might not be best > to accept more work in the form of coprocessor invocations. This could > generalize to all RPC work, perhaps as part of a broader admission control > initiative, but I think it makes sense to start small in an obvious place. > *: We could use % CPU utilization or the UNIX 1min or 5min load average to > determine this, and provide an option for choosing between those > alternatives. -- This message was sent by Atlassian JIRA (v6.4.14#64029)
[jira] [Commented] (HBASE-12217) System load average based client pushback
[ https://issues.apache.org/jira/browse/HBASE-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15337287#comment-15337287 ] Mikhail Antonov commented on HBASE-12217: - Adaptive load control is totally something we should start doing more of; but now kicking out of 1.3 > System load average based client pushback > - > > Key: HBASE-12217 > URL: https://issues.apache.org/jira/browse/HBASE-12217 > Project: HBase > Issue Type: New Feature >Reporter: Andrew Purtell >Assignee: Andrew Purtell > Fix For: 2.0.0, 1.3.0 > > > If a RegionServer host is already heavily loaded* then it might not be best > to accept more work in the form of coprocessor invocations. This could > generalize to all RPC work, perhaps as part of a broader admission control > initiative, but I think it makes sense to start small in an obvious place. > *: We could use % CPU utilization or the UNIX 1min or 5min load average to > determine this, and provide an option for choosing between those > alternatives. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HBASE-12217) System load average based client pushback
[ https://issues.apache.org/jira/browse/HBASE-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14596548#comment-14596548 ] Sean Busbey commented on HBASE-12217: - Understood. I only meant to provide an example of how we'd make use of an expanded set of host metrics elsewhere in the codebase if the chosen library included them. Apologies for not making that clearer. System load average based client pushback - Key: HBASE-12217 URL: https://issues.apache.org/jira/browse/HBASE-12217 Project: HBase Issue Type: New Feature Reporter: Andrew Purtell Assignee: Andrew Purtell Fix For: 2.0.0, 1.3.0 If a RegionServer host is already heavily loaded* then it might not be best to accept more work in the form of coprocessor invocations. This could generalize to all RPC work, perhaps as part of a broader admission control initiative, but I think it makes sense to start small in an obvious place. *: We could use % CPU utilization or the UNIX 1min or 5min load average to determine this, and provide an option for choosing between those alternatives. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HBASE-12217) System load average based client pushback
[ https://issues.apache.org/jira/browse/HBASE-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14596414#comment-14596414 ] Andrew Purtell commented on HBASE-12217: bq. Over in HBASE-13103 Lars Hofhansl calls out several other metrics we'd want to know about when auto-tuning number of regions. Pardon, but this issue isn't about auto-tuning the number of regions. This issue is about providing backpressure to the client based on server side conditions. bq. I think we'd be interested in more than just load average if it was available. There are other proposed strategies on other JIRAs, linked as related items on HBASE-5162. When/if we get around to this work the idea is the load estimation strategy and client response will both be pluggable. It may be more challenging to make them composable. That would be a stretch goal IMHO. System load average based client pushback - Key: HBASE-12217 URL: https://issues.apache.org/jira/browse/HBASE-12217 Project: HBase Issue Type: New Feature Reporter: Andrew Purtell Assignee: Andrew Purtell Fix For: 2.0.0, 1.3.0 If a RegionServer host is already heavily loaded* then it might not be best to accept more work in the form of coprocessor invocations. This could generalize to all RPC work, perhaps as part of a broader admission control initiative, but I think it makes sense to start small in an obvious place. *: We could use % CPU utilization or the UNIX 1min or 5min load average to determine this, and provide an option for choosing between those alternatives. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HBASE-12217) System load average based client pushback
[ https://issues.apache.org/jira/browse/HBASE-12217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14310910#comment-14310910 ] Andrew Purtell commented on HBASE-12217: I recently came across [metrics-sigar|https://github.com/cb372/metrics-sigar], which integrates Coda Hale's Metrics with the [SIGAR|https://support.hyperic.com/display/SIGAR/Home] cross-platform native-based library for gathering system information. It's interesting but of course requires native binaries be deployed and a system property set to point to them. Since we are just interested in system load I think we are better off doing it like YARN does: a generic interface (ResourceCalculatorPlugin) and OS-specific implementations (like LinuxResourceCalculatorPlugin, which knows how to parse procfs), but let's mention both options. System load average based client pushback - Key: HBASE-12217 URL: https://issues.apache.org/jira/browse/HBASE-12217 Project: HBase Issue Type: New Feature Reporter: Andrew Purtell If a RegionServer host is already heavily loaded* then it might not be best to accept more work in the form of coprocessor invocations. This could generalize to all RPC work, perhaps as part of a broader admission control initiative, but I think it makes sense to start small in an obvious place. *: We could use % CPU utilization or the UNIX 1min or 5min load average to determine this, and provide an option for choosing between those alternatives. -- This message was sent by Atlassian JIRA (v6.3.4#6332)