[jira] [Comment Edited] (HBASE-20603) Histogram metrics should reset min and max
[ https://issues.apache.org/jira/browse/HBASE-20603?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16489597#comment-16489597 ] Andrew Purtell edited comment on HBASE-20603 at 5/24/18 7:15 PM: - bq. Do you think creating a new FastLongHistogram on every metrics snapshot (and all of the stuff inside of it) a concern There is no other way to do it without modifying FLH itself. Sure, I could hack that, but then it changes everything, even legacy users of FLH. If you think that is fine, then I could do whatever I like with FLH. Some unit tests for FLH that expect new bins to inherit state from old bins will have to be removed. Note internally at every snapshot in FLH a new Bins object is allocated, which is the bulk of allocation for FLH state. It's true that this change does in effect a double allocation of Bins for every reset. That is a concession to legacy as mentioned above. I'm not sure how much of a concern it will be in practice. I doubt the extra object will ever leave eden or TLAB, and won't have a measurable impact unless we construct an artificial test case that resets a million histograms at once or something like that. was (Author: apurtell): > Do you think creating a new FastLongHistogram on every metrics snapshot (and > all of the stuff inside of it) a concern There is no other way to do it without modifying FLH itself. Sure, I could hack that, but then it changes everything, even legacy users of FLH. If you think that is fine, then I could do whatever I like with FLH. Some unit tests for FLH that expect new bins to inherit state from old bins will have to be removed. Note internally at every snapshot in FLH a new Bins object is allocated, which is the bulk of allocation for FLH state. It's true that this change does in effect a double allocation of Bins for every reset. That is a concession to legacy as mentioned above. I'm not sure how much of a concern it will be in practice. I doubt the extra object will ever leave eden or TLAB, and won't have a measurable impact unless we construct an artificial test case that resets a million histograms at once or something like that. > Histogram metrics should reset min and max > -- > > Key: HBASE-20603 > URL: https://issues.apache.org/jira/browse/HBASE-20603 > Project: HBase > Issue Type: Bug > Components: metrics >Reporter: Andrew Purtell >Assignee: Andrew Purtell >Priority: Minor > Attachments: HBASE-20603-branch-1-WIP.patch > > > It's weird that the bins are reset at every monitoring interval but min and > max are tracked over the lifetime of the process. Makes it impossible to set > alarms on max value as they'll never shut off unless the process is > restarted. Histogram metrics should reset min and max at snapshot time too. > For discussion. -- This message was sent by Atlassian JIRA (v7.6.3#76005)
[jira] [Comment Edited] (HBASE-20603) Histogram metrics should reset min and max
[ https://issues.apache.org/jira/browse/HBASE-20603?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16481166#comment-16481166 ] Andrew Purtell edited comment on HBASE-20603 at 5/24/18 1:52 AM: - Parking a WIP patch that does what I want, but only touches hbase-metrics-api's Histogram. I don't want to touch FastLongHistogram because it has some legacy use. Every histogram metric derived from types in hbase-metrics-api, the majority, especially the call time metrics we are especially interested in, pick up the behavioral change. was (Author: apurtell): Parking a WIP patch that does what I want, but only touches hbase-metrics-api's Histogram. I don't want to touch FastLongHistogram because it has a lot of legacy use all over the place. Every histogram metric derived from types in hbase-metrics-api, the majority, especially the call time metrics we are especially interested in, pick up the behavioral change. > Histogram metrics should reset min and max > -- > > Key: HBASE-20603 > URL: https://issues.apache.org/jira/browse/HBASE-20603 > Project: HBase > Issue Type: Bug > Components: metrics >Reporter: Andrew Purtell >Assignee: Andrew Purtell >Priority: Minor > Attachments: HBASE-20603-branch-1-WIP.patch > > > It's weird that the bins are reset at every monitoring interval but min and > max are tracked over the lifetime of the process. Makes it impossible to set > alarms on max value as they'll never shut off unless the process is > restarted. Histogram metrics should reset min and max at snapshot time too. > For discussion. -- This message was sent by Atlassian JIRA (v7.6.3#76005)
[jira] [Comment Edited] (HBASE-20603) Histogram metrics should reset min and max
[ https://issues.apache.org/jira/browse/HBASE-20603?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=16481166#comment-16481166 ] Andrew Purtell edited comment on HBASE-20603 at 5/24/18 1:36 AM: - Parking a WIP patch that does what I want, but only touches hbase-metrics-api's Histogram. I don't want to touch FastLongHistogram because it has a lot of legacy use all over the place. Every histogram metric derived from types in hbase-metrics-api, the majority, especially the call time metrics we are especially interested in, pick up the behavioral change. was (Author: apurtell): Parking a WIP patch that does what I want, but only touches hbase-metrics-api's Histogram. I don't want to touch FastLongHistogram because it has a lot of legacy use all over the place. We are most interested in applying the proposed new semantics to call metrics, so will propose that change separately as an update of call metrics to use hbase-metrics-api's Histogram, at some future time if this is accepted. > Histogram metrics should reset min and max > -- > > Key: HBASE-20603 > URL: https://issues.apache.org/jira/browse/HBASE-20603 > Project: HBase > Issue Type: Bug > Components: metrics >Reporter: Andrew Purtell >Assignee: Andrew Purtell >Priority: Minor > Attachments: HBASE-20603-branch-1-WIP.patch > > > It's weird that the bins are reset at every monitoring interval but min and > max are tracked over the lifetime of the process. Makes it impossible to set > alarms on max value as they'll never shut off unless the process is > restarted. Histogram metrics should reset min and max at snapshot time too. > For discussion. -- This message was sent by Atlassian JIRA (v7.6.3#76005)