[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15368750#comment-15368750 ] John Rushford edited comment on TS-4642 at 7/8/16 11:52 PM: @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down when the parent records are shared. If they are shared, services unrelated to the misbehaving origin service would get 502 errors because the shared parents eventually are all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? was (Author: jrushford): @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down when the parent records are shared. If they are shared, services unrelated to the misbehaving origin service would get 502 errors because the shared parents eventually all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15368750#comment-15368750 ] John Rushford edited comment on TS-4642 at 7/8/16 11:51 PM: @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down when the parent records are shared. If they are shared, services unrelated to the misbehaving origin service would get 502 errors because the shared parents eventually all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? was (Author: jrushford): @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down if the parent records are shared. If they are shared, services unrelated to the misbehaving origin service would get 502 errors because the shared parents eventually all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15368750#comment-15368750 ] John Rushford edited comment on TS-4642 at 7/8/16 11:49 PM: @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down if the parent records are shared. If they are shared, services unrelated to the misbehaving origin service would get 502 errors because the shared parents eventually all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? was (Author: jrushford): @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down if the parent records are shared. If they are shared service unrelated to the misbehaving origin service would get 502 errors as the shared parents are all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15368750#comment-15368750 ] John Rushford edited comment on TS-4642 at 7/8/16 11:48 PM: @zwoop in our production environment we have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down if the parent records are shared. If they are shared service unrelated to the misbehaving origin service would get 502 errors as the shared parents are all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? was (Author: jrushford): @zwoop in our production environment have a parent config line for every delivery service and they all use the same proxies. We do this precisely for case #2 so that a a misbehaving origin doesn't cause all parents for all delivery services to be marked down if the parent records are shared. If they are shared service unrelated to the misbehaving origin service would get 502 errors as the shared parents are all marked down. Now for case #1 when a parent proxy becomes unreachable, the non-shared parents are pretty quickly marked down by connection failures across all delivery services but don't forget that when a connection failure is detected to a parent proxy, another parent is selected and that transaction is retried using the new parent. In practice Leif, I have not seen a problem with this. In the diags.log, you see the non-shared parent marked down maybe 20 or more times but that is because there are 20 or more non-shared records. You do not see any 502's because these transactions selected and used another parent proxy. I don't see a problem here myself. What are your thoughts about this? > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15366312#comment-15366312 ] John Rushford edited comment on TS-4642 at 7/7/16 3:59 PM: --- @zwoop - yes i think we should close it but, I'll update the documentation before I do so. I think that most use cases can be accommodated as is. was (Author: jrushford): @zwoop - yes i think we should close it but, I'll update the documentation before I do so. I think that most use cases can accommodated as is. > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Comment Edited] (TS-4642) Parent host healths across multiple configurations is inefficient
[ https://issues.apache.org/jira/browse/TS-4642?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15365377#comment-15365377 ] Leif Hedstrom edited comment on TS-4642 at 7/7/16 12:23 AM: It's unfortunately not as simple as that, since not all requests should have a parent. And some should potentially have different parents. Fwiw, we're working on some configuration changes where we add rules to parent.config to go direct for those (few) domains that should go direct (no parent), and then a dest_domain=. for everything else. Your use case also seems useful, albeit I'd argue that the settings and configurations are not obvious :). was (Author: zwoop): It's unfortunately not as simple as that, since not all requests should have a parent. And some should potentially have different parents. Your use case also seems useful, albeit I'd argue that the settings and configurations are not obvious :). > Parent host healths across multiple configurations is inefficient > - > > Key: TS-4642 > URL: https://issues.apache.org/jira/browse/TS-4642 > Project: Traffic Server > Issue Type: Improvement > Components: Parent Proxy >Reporter: Leif Hedstrom >Assignee: John Rushford > Fix For: sometime > > > It seems that if I have multiple config lines in parent.config, where I use > the same list of parents for multiple rules, each such rules keeps it's own > "health" status? That means that if you repeat a line 100 times, the number > of "failures" until marked down is 100x more than what we configured in > records.config. > It seems that the health and failure counts should be per host, such that > regardless of how many times I use each host in parent.config, it'll > consistently use the settings from records.config. > So, I'm not 100% sure this is how things works, but if they are not, I blame > Phil. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
