Change-queue job failures this weekend
Hi, We seen a great deal of noise coming from the change queue this weekend. While a part of it is due to actual code regressions, some of that was actually due to two separate infra issues. One issue we had was with building FC26 packages - it turns out that a yum-incompatible update if the 'cmake' package was introduced to the FC26 updates repo. Since for the time being we use 'yum' to setup the mock environments, the build jobs for FC26 started failing. This issue was actually reported to us [1]. To resolve this - we rolled back the FC26 mirror to a time before the breaking change was introduced, and then re-triggered the merge events for all the patches that failed building to introduce passing build to the change queue. The second issue had to do with the introduction of FC27 slaves - it seems that slaves were misconfigured [2] and did not include vary basic packages like 'git' - this caused the CQ master job to simply crash and stop queue processing. To resolve this issue we disabled the FC27 slaves, resumed CQ operation and then re-sent all changes that failed to be added into the queue. We are in the final phases of integrating a new oVirt release, so proper CQ operation is crucial at this time. Additionally, due to a substantial amount of regressions introduced last week, the CQ currently has a huge backlog of ~180 changes to work through, this means that every bisection takes 8 OST runs, so we have no CQ minutes to spare. The FC27 slaves issue cost us 11 hours in which the CQ was not running. It also manifested itself in failures of the 'standard-enqueue' job. These kinds of failures need to be handled promptly or be avoided altogether. Build failures can make the CQ waste time too, as it runs bisections to detect and remove changes that fail to build. At this time, a single failed build can waste up to 8 hours! Lets try to be more careful about introducing infrastructure changes to the system at sensitive times, and be more vigilant about failure reports from jenkins. [1]: https://ovirt-jira.atlassian.net/browse/OVIRT-1854 [2]: https://ovirt-jira.atlassian.net/browse/OVIRT-1855 -- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that. I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one. In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST. For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ): 1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually ) Thoughts? On Sun, Jan 21, 2018 at 10:42 AM, Barak Korren <bkorren@redhat.com> wrote:
Hi,
We seen a great deal of noise coming from the change queue this weekend. While a part of it is due to actual code regressions, some of that was actually due to two separate infra issues.
One issue we had was with building FC26 packages - it turns out that a yum-incompatible update if the 'cmake' package was introduced to the FC26 updates repo. Since for the time being we use 'yum' to setup the mock environments, the build jobs for FC26 started failing.
This issue was actually reported to us [1].
To resolve this - we rolled back the FC26 mirror to a time before the breaking change was introduced, and then re-triggered the merge events for all the patches that failed building to introduce passing build to the change queue.
The second issue had to do with the introduction of FC27 slaves - it seems that slaves were misconfigured [2] and did not include vary basic packages like 'git' - this caused the CQ master job to simply crash and stop queue processing.
To resolve this issue we disabled the FC27 slaves, resumed CQ operation and then re-sent all changes that failed to be added into the queue.
We are in the final phases of integrating a new oVirt release, so proper CQ operation is crucial at this time. Additionally, due to a substantial amount of regressions introduced last week, the CQ currently has a huge backlog of ~180 changes to work through, this means that every bisection takes 8 OST runs, so we have no CQ minutes to spare.
The FC27 slaves issue cost us 11 hours in which the CQ was not running. It also manifested itself in failures of the 'standard-enqueue' job. These kinds of failures need to be handled promptly or be avoided altogether.
Build failures can make the CQ waste time too, as it runs bisections to detect and remove changes that fail to build. At this time, a single failed build can waste up to 8 hours!
Lets try to be more careful about introducing infrastructure changes to the system at sensitive times, and be more vigilant about failure reports from jenkins.
[1]: https://ovirt-jira.atlassian.net/browse/OVIRT-1854 [2]: https://ovirt-jira.atlassian.net/browse/OVIRT-1855
-- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted _______________________________________________ Infra mailing list Infra@ovirt.org http://lists.ovirt.org/mailman/listinfo/infra
-- Eyal edri MANAGER RHV DevOps EMEA VIRTUALIZATION R&D Red Hat EMEA <https://www.redhat.com/> <https://red.ht/sig> TRIED. TESTED. TRUSTED. <https://redhat.com/trusted> phone: +972-9-7692018 irc: eedri (on #tlv #rhev-dev #rhev-integ)
On 21 January 2018 at 12:39, Eyal Edri <eedri@redhat.com> wrote:
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that.
I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one.
In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST.
For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ):
1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually )
Thoughts?
We have a staging system - we should be using it for staging.... -- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
On Sun, Jan 21, 2018 at 12:47 PM, Barak Korren <bkorren@redhat.com> wrote:
On 21 January 2018 at 12:39, Eyal Edri <eedri@redhat.com> wrote:
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that.
I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one.
In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST.
For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ):
1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually )
Thoughts?
We have a staging system - we should be using it for staging....
Do we have OST tests or manual job avaialble there? In any case, this doesn't contradict what I suggested, even if you test on staging, there could be differences from the production system, so we should take care when we upgrade regardless. Another point when scheduling an upgrade, is to talk to infra owner or the CI team and understand if we currently have a large Q in CQ or known failures, so it might be best to wait a bit until its cleared.
-- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
-- Eyal edri MANAGER RHV DevOps EMEA VIRTUALIZATION R&D Red Hat EMEA <https://www.redhat.com/> <https://red.ht/sig> TRIED. TESTED. TRUSTED. <https://redhat.com/trusted> phone: +972-9-7692018 irc: eedri (on #tlv #rhev-dev #rhev-integ)
Due to a bug with CPU model type in OST, right now, we have only 2 available physical slaves to run tests (srv04 and srv05). All the other slaves are temporary offline until the issue will be solved. Please, avoid any changes to srv04 and srv05. On Sun, Jan 21, 2018 at 12:50 PM, Eyal Edri <eedri@redhat.com> wrote:
On Sun, Jan 21, 2018 at 12:47 PM, Barak Korren <bkorren@redhat.com> wrote:
On 21 January 2018 at 12:39, Eyal Edri <eedri@redhat.com> wrote:
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that.
I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one.
In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST.
For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ):
1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually )
Thoughts?
We have a staging system - we should be using it for staging....
Do we have OST tests or manual job avaialble there? In any case, this doesn't contradict what I suggested, even if you test on staging, there could be differences from the production system, so we should take care when we upgrade regardless.
Another point when scheduling an upgrade, is to talk to infra owner or the CI team and understand if we currently have a large Q in CQ or known failures, so it might be best to wait a bit until its cleared.
-- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
--
Eyal edri
MANAGER
RHV DevOps
EMEA VIRTUALIZATION R&D
Red Hat EMEA <https://www.redhat.com/> <https://red.ht/sig> TRIED. TESTED. TRUSTED. <https://redhat.com/trusted> phone: +972-9-7692018 <+972%209-769-2018> irc: eedri (on #tlv #rhev-dev #rhev-integ)
_______________________________________________ Infra mailing list Infra@ovirt.org http://lists.ovirt.org/mailman/listinfo/infra
-- DANIEL BELENKY RHV DEVOPS
On 21 January 2018 at 12:50, Eyal Edri <eedri@redhat.com> wrote:
On Sun, Jan 21, 2018 at 12:47 PM, Barak Korren <bkorren@redhat.com> wrote:
On 21 January 2018 at 12:39, Eyal Edri <eedri@redhat.com> wrote:
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that.
I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one.
In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST.
For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ):
1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually )
Thoughts?
We have a staging system - we should be using it for staging....
Do we have OST tests or manual job avaialble there?
We can add them easily, or simply run Lago manually when needed.
In any case, this doesn't contradict what I suggested, even if you test on staging, there could be differences from the production system, so we should take care when we upgrade regardless.
Yes, but at least we'd know we green lighted the new configuration - I'm sure in this case we could have found at least some of the issues on staging (Like the fc27 issues for example) and could have avoided expansive production failures. Another point when scheduling an upgrade, is to talk to infra owner or the
CI team and understand if we currently have a large Q in CQ or known failures, so it might be best to wait a bit until its cleared.
-- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
On Sun, Jan 21, 2018 at 1:01 PM, Barak Korren <bkorren@redhat.com> wrote:
On 21 January 2018 at 12:50, Eyal Edri <eedri@redhat.com> wrote:
On Sun, Jan 21, 2018 at 12:47 PM, Barak Korren <bkorren@redhat.com> wrote:
On 21 January 2018 at 12:39, Eyal Edri <eedri@redhat.com> wrote:
There is another issue, which is currently failing all CQ, and its related to the new IBRS CPU model. It looks like all of the lago slaves were upgraded to new Libvirt and kernel on Friday, while we still don't have a fix on lago-ost-plugin for that.
I think there was a misunderstanding about what to upgrade, and it might have been understood that only the bios upgrade breaks it and not the kernel one.
In any case, we're currently fixing the issue, either by downgrading the relevant pkgs on lago slaves or adding the mapping to new CPU types from OST.
For future, I suggest a few updates to maintenance work on Jenkins slaves ( VMs or BM ):
1. Let's avoid doing an upgrade close to a weekend ( i.e not on Thu-Sun ), so all the team can be around to help if needed or if something unexpected happens. 2. When we have a system-wide upgrade scheduled, like all BM slaves or VMs for a specific OS, let's adopt a gradual upgrade with a few days window in between, e.g, if we need to upgrade all Lago slaves, let's upgrade 1-2 and wait to see if nothing breaks and continue after we verify OST runs ( either seeing on CQ or running manually )
Thoughts?
We have a staging system - we should be using it for staging....
Do we have OST tests or manual job avaialble there?
We can add them easily, or simply run Lago manually when needed.
In any case, this doesn't contradict what I suggested, even if you test on staging, there could be differences from the production system, so we should take care when we upgrade regardless.
Yes, but at least we'd know we green lighted the new configuration - I'm sure in this case we could have found at least some of the issues on staging (Like the fc27 issues for example) and could have avoided expansive production failures.
Another point when scheduling an upgrade, is to talk to infra owner or the
CI team and understand if we currently have a large Q in CQ or known failures, so it might be best to wait a bit until its cleared.
Adding infra-support so we can gather this info and prepare a maintanaince / upgrade checklist to add to the oVirt infra docs. Let's continue the discussion, suggestion on that ticket.
-- Barak Korren RHV DevOps team , RHCE, RHCi Red Hat EMEA redhat.com | TRIED. TESTED. TRUSTED. | redhat.com/trusted
-- Eyal edri MANAGER RHV DevOps EMEA VIRTUALIZATION R&D Red Hat EMEA <https://www.redhat.com/> <https://red.ht/sig> TRIED. TESTED. TRUSTED. <https://redhat.com/trusted> phone: +972-9-7692018 irc: eedri (on #tlv #rhev-dev #rhev-integ)
participants (3)
-
Barak Korren -
Daniel Belenky -
Eyal Edri