Comprehensive fixes for deployment image triggering #1433

ironcladlou · 2015-03-24T18:43:16Z

The current deployment image triggering logic is not accounting for various edge cases related to image IDs (or lack thereof). Define all the triggering scenarios in the comprehensive unit test and plug the holes in the triggering code so that triggers work consistently.

One known side effect of the current flawed code is mis-firing triggers which update the container to a nil image ID, causing races described in #1443.

Base all comparisons on image ID, and add new state to image change trigger params to store the last image ID used for triggering.

ironcladlou · 2015-03-24T18:43:31Z

@smarterclayton @ncdc @pmorie PTAL.

ncdc · 2015-03-24T18:59:58Z

I'm wondering if we could just call ParseDockerImageReference for both container.Image and latest.DockerImageReference and then compare the 2?

ironcladlou · 2015-03-24T20:24:39Z

I'm wondering if we could just call ParseDockerImageReference for both container.Image and latest.DockerImageReference and then compare the 2?

If ImageRepository.DockerImageReference always eventually contains the latest state (including image ID), I don't see why not. But that's not saying much, since I'm still learning all the IR mechanics on the fly. 😁

ironcladlou · 2015-03-24T21:12:57Z

Talked with @ncdc, there's mork work to do here still. Hashing out unit tests currently.

ncdc · 2015-03-24T21:16:06Z

It's not ImageRepository.DockerImageReference - it's the DockerImageReferences created by parsing container.Image and latest.DockerImageReference. Very confusing, I know :-).

status.tags.items[n].dockerImageReference will always be a pull spec. Sometimes it will just be a pull spec to a tag, sometimes it will be a pull spec to a "pull by id" tag, and once we have v2 registries and Docker 1.6, it will be a pull spec to an image by its digest, for true pull by id semantics.

ironcladlou · 2015-03-24T21:16:15Z

@ncdc Do you think the new test support in ironcladlou@face432 sufficient to express our scenarios?

ncdc · 2015-03-24T21:18:04Z

@ironcladlou I think so

smarterclayton · 2015-03-26T17:00:15Z

Status on this?

ironcladlou · 2015-03-26T17:03:38Z

Paused it while @ncdc was out to make progress on deployment hooks, will be revisiting it this afternoon.

ironcladlou · 2015-03-26T18:49:12Z

@ncdc helped me hash this out. I'm going to make a stateless utility method in the pkg/image/api/helper.go which will be something like:

func TagDiffers(tag string, pullspec string, repo ImageRepository) bool

The table tests in pkg/deploy/controller/imagechange/controller_test.go#TestHandle_matchScenarios are currently conflating tests for the diff logic and the controller's core triggering logic. I'll use them as a guide to make tests for TagDiffers and simplify the controller test.

cc @bparees for builds.

ironcladlou · 2015-03-26T19:54:54Z

Nevermind, the proposed solution still isn't right. We're working on it.

ironcladlou · 2015-03-27T19:13:15Z

Latest branch has big refactorings to the API, image trigger controller, and generator to implement the correct behavior. The e2e tests are already working, but I need to go back and re-do the unit tests.

ironcladlou · 2015-03-30T19:40:23Z

Okay, @ncdc @smarterclayton and @pmorie.

I had to rewrite the deployment image change trigger controller, the generator, and all the related tests. I also had to make an additive/backwards-compatible API change to add a status struct to DeploymentTriggerImageChangeParams so we can correctly track the last triggered image ID.

The changes are significant, but as far as I know, the behavior is actually correct now in terms of ImageRepository interactions. Hopefully you all can do some auditing.

In the meantime, [test].

ironcladlou · 2015-03-30T19:41:31Z

I was also able to remove a metric ton of test cruft during the course of this by consolidating test fixtures and removing dead tests.

smarterclayton · 2015-03-30T19:52:47Z

On Mar 30, 2015, at 3:40 PM, Dan Mace notifications@github.com wrote:

Okay, @ncdc @smarterclayton and @pmorie.

I had to rewrite the deployment image change trigger controller, the generator, and all the related tests. I also had to make an additive/backwards-compatible API change to add a status struct to DeploymentTriggerImageChangeParams so we can correctly track the last triggered image ID.

So I have a question... Why is the pull spec in the pod template not sufficient? Why do we need this other field?
The changes are significant, but as far as I know, the behavior is actually correct now in terms of ImageRepository interactions. Hopefully you all can do some auditing.

In the meantime, [test].

—
Reply to this email directly or view it on GitHub.

ncdc · 2015-03-30T19:54:19Z

Why is the pull spec in the pod template not sufficient? Why do we need this other field?

We need the image id, which won't be in the pull spec if the registry it comes from doesn't support pull by id.

smarterclayton · 2015-03-30T19:57:04Z

Why do we need the image id? If the pull spec doesn't change you can't guarantee you'll get a new image (although you can guarantee you'll get a new pod, you can't force the node to pull a new image).

----- Original Message -----

Why is the pull spec in the pod template not sufficient? Why do we need
this other field?

We need the image id, which won't be in the pull spec if the registry it
comes from doesn't support pull by id.

Reply to this email directly or view it on GitHub:
#1433 (comment)

ncdc · 2015-03-30T20:01:18Z

True, but if the ImagePullPolicy is the default, it will always pull. This is the easiest way to be consistent for all our image change trigger semantics. The code was getting too ugly and too hard to reason through. Going forward in my IS PR, the TagEvents in Status.Tags[tag] must have both DockerImageReference and Image (id) to help enable this. Dan and I spent multiple hours talking through this on IRC and the phone, and this way is easy to understand and follow.

ncdc · 2015-03-30T21:18:26Z

In that case image id isn't helping, you just need a way to trigger the deployment. The important distinction here is that only a change to the image pull spec guarantees a deployment. If we import his registry but don't change the tag, when we trigger a new deployment we're not promising he's going to get a new image.

Agreed, unless the pull policy is set correctly, but we can't easily guarantee that.

He has to rev the tag himself.

From what to what? If the current TagEvent for latest is DockerImageReference=joe/awesome:latest, Image=abcd1234 and he pushes a new image, the only thing that changes in the new TagEvent is Image. The pull spec, DockerImageReference, hasn't changed.

smarterclayton · 2015-03-30T21:27:24Z

He must physically change the tag on the docker end. I'm provisionally ok with someone setting PullAlways (or maybe we need one which is PullAtLeastOnce) after the trigger a deploy but i'm not sure we require this change in order to reflect that.

----- Original Message -----

In that case image id isn't helping, you just need a way to trigger the
deployment. The important distinction here is that only a change to the
image pull spec guarantees a deployment. If we import his registry but
don't change the tag, when we trigger a new deployment we're not promising
he's going to get a new image.

Agreed, unless the pull policy is set correctly, but we can't easily
guarantee that.

He has to rev the tag himself.

From what to what? If the current TagEvent for latest is
DockerImageReference=joe/awesome:latest, Image=abcd1234 and he pushes a
new image, the only thing that changes in the new TagEvent is Image. The
pull spec, DockerImageReference, hasn't changed.

Reply to this email directly or view it on GitHub:
#1433 (comment)

ncdc · 2015-03-30T21:32:14Z

He must physically change the tag on the docker end

Is this what you mean?

docker build -t joe/awesome:revision2 .
docker push joe/awesome:revision2
push button to manually sync status.tags
update ImageStream, set spec.tags[latest] = ImageStreamTag(awesome:revision2)

at this point, with a trigger watching ImageStreamTag(awesome:latest), the pull spec for status.tags[latest].Items[0].DockerImageReference is now joe/awesome:revision2, which is different from before (joe/awesome:latest), thus triggering a deployment.

Isn't it a whole lot easier just to trigger on image id? 😄 We can always educate users to set the pull policy to always in this situation.

smarterclayton · 2015-03-30T21:38:55Z

----- Original Message -----

He must physically change the tag on the docker end

Is this what you mean?

docker build -t joe/awesome:revision2 .

docker push joe/awesome:revision2

push button to manually sync status.tags

update ImageStream, set spec.tags[latest] =
ImageStreamTag(awesome:revision2)

at this point, with a trigger watching ImageStreamTag(awesome:latest), the
pull spec for status.tags[latest].Items[0].DockerImageReference is now
joe/awesome:revision2, which is different from before
(joe/awesome:latest), thus triggering a deployment.

Isn't it a whole lot easier just to trigger on image id? 😄 We can
always educate users to set the pull policy to always in this situation.

You don't need to physically store the image ID to trigger a build - that happens automatically as long as you can make at least one edit to the deployment config (or just to the latest version, but that's a bit tacky right now).

ncdc · 2015-03-30T21:40:54Z

So are you saying then that all that needs to happen is:

docker build -t joe/awesome:latest
docker push joe/awesome:latest
update deployment config
profit?

smarterclayton · 2015-03-30T21:41:03Z

pkg/deploy/controller/imagechange/controller.go

+
+			// If the tag event doesn't have an image ID yet, we can't determine
+			// whether there was really a change
+			if len(latestEvent.Image) == 0 {


Why not? Why can't you just use the pull spec like Builds?

This actually won't be an issue once my ImageStream PR goes in - Image won't ever be empty

What if I want to tag an arbitrary docker repository on the docker hub?

----- Original Message -----

}

if latest.Image != containerImageID {

glog.V(4).Infof("Container %s for config %s: image id changed from %q
to %q; regenerating config", container.Name, labelFor(config),
containerImageID, latest.Image)

configsToGenerate = append(configsToGenerate, config)

firedTriggersForConfig[config.Name] =
append(firedTriggersForConfig[config.Name], params)

}

// Find the latest tag event for the trigger tag

latestEvent, err := imageapi.LatestTaggedImage(imageRepo, params.Tag)

if err != nil {

glog.V(4).Infof("Couldn't find latest tag event for tag %s in
imageRepo %s: %s", params.Tag, labelForRepo(imageRepo), err)

continue

}

// If the tag event doesn't have an image ID yet, we can't determine

// whether there was really a change

if len(latestEvent.Image) == 0 {

This actually won't be an issue once my ImageStream PR goes in - Image won't
ever be empty

Reply to this email directly or view it on GitHub:
https://github.com/openshift/origin/pull/1433/files#r27434928

You have to create an ImageStream for it

Sent from my iPhone

On Mar 30, 2015, at 6:16 PM, Clayton Coleman notifications@github.com wrote:

In pkg/deploy/controller/imagechange/controller.go:

}

if latest.Image != containerImageID {

glog.V(4).Infof("Container %s for config %s: image id changed from %q to %q; regenerating config", container.Name, labelFor(config), containerImageID, latest.Image)

configsToGenerate = append(configsToGenerate, config)

firedTriggersForConfig[config.Name] = append(firedTriggersForConfig[config.Name], params)

}

// Find the latest tag event for the trigger tag

latestEvent, err := imageapi.LatestTaggedImage(imageRepo, params.Tag)

if err != nil {

glog.V(4).Infof("Couldn't find latest tag event for tag %s in imageRepo %s: %s", params.Tag, labelForRepo(imageRepo), err)

continue

}

// If the tag event doesn't have an image ID yet, we can't determine

// whether there was really a change

if len(latestEvent.Image) == 0 {
What if I want to tag an arbitrary docker repository on the docker hub?
…
—
Reply to this email directly or view it on GitHub.

----- Original Message -----

}

if latest.Image != containerImageID {

glog.V(4).Infof("Container %s for config %s: image id changed from %q
to %q; regenerating config", container.Name, labelFor(config),
containerImageID, latest.Image)

configsToGenerate = append(configsToGenerate, config)

firedTriggersForConfig[config.Name] =
append(firedTriggersForConfig[config.Name], params)

}

// Find the latest tag event for the trigger tag

latestEvent, err := imageapi.LatestTaggedImage(imageRepo, params.Tag)

if err != nil {

glog.V(4).Infof("Couldn't find latest tag event for tag %s in
imageRepo %s: %s", params.Tag, labelForRepo(imageRepo), err)

continue

}

// If the tag event doesn't have an image ID yet, we can't determine

// whether there was really a change

if len(latestEvent.Image) == 0 {

You have to create an ImageStream for it

I don't think we are on the same page about that.

@ironcladlou I talked to @smarterclayton and agree that a user shouldn't have to go through the hoops of creating an ImageStream that points to an external Docker image repository just to be able to deploy something. Let's talk tomorrow about how to trigger off of pull spec instead of image (it shouldn't be too much of a change).

Sounds good, will catch up with you and figure out how to refactor.

smarterclayton · 2015-03-31T19:00:10Z

Is this going to fix the integration test race?

ironcladlou · 2015-04-01T15:18:23Z

I re-centered the logic around TagEvent.Status.DockerImageReference rather than .Image. PTAL @smarterclayton @ncdc

ironcladlou · 2015-04-01T17:18:28Z

Not sure yet exactly what failed in that jenkins run, so I want to run this through [test] again.

ironcladlou · 2015-04-01T18:11:10Z

Still not sure what's going on with jenkins. Unit, integration, and e2e tests are passing.

ironcladlou · 2015-04-01T18:16:43Z

#1550 is what is up with Jenkins.

@smarterclayton PTAL, I think your comments are addressed.

smarterclayton · 2015-04-01T18:39:01Z

pkg/deploy/controller/imagechange/controller.go

-					firedTriggersForConfig[config.Name] = append(firedTriggersForConfig[config.Name], params)
-				}
+			// Ensure a change occured
+			if latestEvent.DockerImageReference != params.LastTriggeredImage {


If latestEvent.DockerImageReference is empty you shouldn't trigger an update.

It will never be empty

This is now accounted for.

smarterclayton · 2015-04-01T19:33:02Z

LGTM, squash

Rewrite deployment image triggering and generator logic to correctly interact with ImageRepository objects.

ironcladlou · 2015-04-01T19:39:20Z

Squashed

smarterclayton · 2015-04-01T20:00:34Z

[merge]

openshift-bot · 2015-04-01T20:00:48Z

Evaluated for origin up to ab496df

openshift-bot · 2015-04-02T13:21:35Z

continuous-integration/openshift-jenkins/test SUCCESS (https://ci.openshift.redhat.com/jenkins/job/test_pull_requests_openshift3/1648/)

openshift-bot · 2015-04-02T13:21:51Z

continuous-integration/openshift-jenkins/merge SUCCESS (https://ci.openshift.redhat.com/jenkins/job/test_pull_requests_openshift3/1648/) (Image: devenv-fedora_1188)

Merged by openshift-bot

…service-catalog/' changes from aa27078754..dabde2eb85 dabde2eb85 origin build: add origin tooling b70c076 Reorder class and plan creation; test plan conflict handling (openshift#1459) 4bea012 Use versioned client APIs (openshift#1458) ff4af30 clean up logic for 410 gone deprovision poll (openshift#1452) 3fddf27 clean up logic and fix message for failed poll (openshift#1451) 40926cd Fix typo from openshift#1354 (openshift#1456) ff86ef2 Delete removed serviceplans when they have no instances left (openshift#1444) 8411a16 tweak binding setAndUpdateOrphanMitigation function (openshift#1448) ce28252 Combine apiserver and controller-manager into a single service-catalog image (openshift#1343) 7bbc8ee Check service class / plan before allowing provisioning or plan changes. (openshift#1439) baf28de Create listers before adding event handlers in controller (openshift#1446) 294157d remove setServiceBindingCondition dependency on controller (openshift#1441) 118a0f7 Fix typo in validation (openshift#1447) 117bfbd clean up error logging (openshift#1443) dff470f Move "External" around in some resource names/properties (openshift#1354) 0885edb Adding expectedGot function and using it. (openshift#1440) a7d582e Pretty controller broker (openshift#1442) c5edfaf Set apimachinery build variables with semver info (openshift#1429) 0e90d82 Add a pretty formatter for ClusterService[Class|Plan] (openshift#1408) fb874df Remove deprecated basic auth config support (openshift#1431) f4cd181 Migrate to metav1 methods for manipulating controllerRefs (openshift#1433) 96b286e Make service/plan reference fields on instance spec selectable (openshift#1422) 33f2b04 First example using the pretty context builder. (openshift#1403) 7852917 Stop using corev1.ObjectReference and corev1.LocalObjectReference (openshift#1417) fcf9480 Add tests for plan updates (openshift#1412) 819332e Add root CAs (openshift#1419) b49a76a Clean Makefile a little (openshift#1399) d681da0 Use a separate etcd prefix for each integration test to keep tests isolated (openshift#1415) 314a622 Wire etcd prefix to storage and call complete with options (openshift#1394) REVERT: aa27078754 origin build: add origin tooling git-subtree-dir: cmd/service-catalog/go/src/github.com/kubernetes-incubator/service-catalog git-subtree-split: dabde2eb859b5e31e97c01a704561fc27e1848b2

…service-catalog/' changes from aa27078754..510060232e 510060232e origin build: add origin tooling de45e94 v0.1.0 chart changes (openshift#1468) 0bb9982 Modify Makefile to only specify ldflags once (openshift#1471) 5d6afac Fixes openshift#735: Add repo-sync script for charts (openshift#1453) 630f13f fix lingering unversioned client API (openshift#1466) 6f49128 Fix several logging errors (openshift#1464) 2aece61 Delete removed serviceClasses when they have no instances left (openshift#1450) 179d302 Uncommenting UID field after updating to k8s 1.8 (openshift#1457) b70c076 Reorder class and plan creation; test plan conflict handling (openshift#1459) 4bea012 Use versioned client APIs (openshift#1458) ff4af30 clean up logic for 410 gone deprovision poll (openshift#1452) 3fddf27 clean up logic and fix message for failed poll (openshift#1451) 40926cd Fix typo from openshift#1354 (openshift#1456) ff86ef2 Delete removed serviceplans when they have no instances left (openshift#1444) 8411a16 tweak binding setAndUpdateOrphanMitigation function (openshift#1448) ce28252 Combine apiserver and controller-manager into a single service-catalog image (openshift#1343) 7bbc8ee Check service class / plan before allowing provisioning or plan changes. (openshift#1439) baf28de Create listers before adding event handlers in controller (openshift#1446) 294157d remove setServiceBindingCondition dependency on controller (openshift#1441) 118a0f7 Fix typo in validation (openshift#1447) 117bfbd clean up error logging (openshift#1443) dff470f Move "External" around in some resource names/properties (openshift#1354) 0885edb Adding expectedGot function and using it. (openshift#1440) a7d582e Pretty controller broker (openshift#1442) c5edfaf Set apimachinery build variables with semver info (openshift#1429) 0e90d82 Add a pretty formatter for ClusterService[Class|Plan] (openshift#1408) fb874df Remove deprecated basic auth config support (openshift#1431) f4cd181 Migrate to metav1 methods for manipulating controllerRefs (openshift#1433) 96b286e Make service/plan reference fields on instance spec selectable (openshift#1422) 33f2b04 First example using the pretty context builder. (openshift#1403) 7852917 Stop using corev1.ObjectReference and corev1.LocalObjectReference (openshift#1417) fcf9480 Add tests for plan updates (openshift#1412) 819332e Add root CAs (openshift#1419) b49a76a Clean Makefile a little (openshift#1399) d681da0 Use a separate etcd prefix for each integration test to keep tests isolated (openshift#1415) 314a622 Wire etcd prefix to storage and call complete with options (openshift#1394) REVERT: aa27078754 origin build: add origin tooling git-subtree-dir: cmd/service-catalog/go/src/github.com/kubernetes-incubator/service-catalog git-subtree-split: 510060232e54eb64b294213bb5d7847e169a2fac

…1433)

ironcladlou mentioned this pull request Mar 25, 2015

TestSimpleImageChangeBuildTrigger is racy now #1443

Closed

ironcladlou force-pushed the deploy-trigger-test-fix branch from face432 to d2c42dc Compare March 25, 2015 13:23

ironcladlou changed the title ~~Fall back to pullspecs for IR triggering~~ Comprehensive fixes for deployment image triggering Mar 25, 2015

ironcladlou force-pushed the deploy-trigger-test-fix branch 5 times, most recently from 1db139a to b1d69e8 Compare March 27, 2015 17:55

ironcladlou force-pushed the deploy-trigger-test-fix branch 4 times, most recently from 9df3d81 to 29a2ba5 Compare March 30, 2015 19:36

smarterclayton reviewed Mar 30, 2015
View reviewed changes

ironcladlou force-pushed the deploy-trigger-test-fix branch from 29a2ba5 to 6fc6b3d Compare April 1, 2015 13:06

smarterclayton reviewed Apr 1, 2015
View reviewed changes

Rewrite deployment ImageRepo handling

ab496df

Rewrite deployment image triggering and generator logic to correctly interact with ImageRepository objects.

ironcladlou force-pushed the deploy-trigger-test-fix branch from 9f3e15d to ab496df Compare April 1, 2015 19:35

openshift-bot pushed a commit that referenced this pull request Apr 2, 2015

Merge pull request #1433 from ironcladlou/deploy-trigger-test-fix

71639c1

Merged by openshift-bot

openshift-bot merged commit 71639c1 into openshift:master Apr 2, 2015

ironcladlou deleted the deploy-trigger-test-fix branch April 2, 2015 14:35

smarterclayton added the kind/bug Categorizes issue or PR as related to a bug. label Apr 8, 2015

smarterclayton modified the milestone: 0.5.0 (beta3) Apr 23, 2015

jpeeler pushed a commit to jpeeler/origin that referenced this pull request Feb 1, 2018

Migrate to metav1 methods for manipulating controllerRefs (openshift#…

f4cd181

…1433)

Comprehensive fixes for deployment image triggering #1433

Comprehensive fixes for deployment image triggering #1433

Conversation

ironcladlou commented Mar 24, 2015

ironcladlou commented Mar 24, 2015

ncdc commented Mar 24, 2015

ironcladlou commented Mar 24, 2015

ironcladlou commented Mar 24, 2015

ncdc commented Mar 24, 2015

ironcladlou commented Mar 24, 2015

ncdc commented Mar 24, 2015

smarterclayton commented Mar 26, 2015

ironcladlou commented Mar 26, 2015

ironcladlou commented Mar 26, 2015

ironcladlou commented Mar 26, 2015

ironcladlou commented Mar 27, 2015

ironcladlou commented Mar 30, 2015

ironcladlou commented Mar 30, 2015

smarterclayton commented Mar 30, 2015

ncdc commented Mar 30, 2015

smarterclayton commented Mar 30, 2015

ncdc commented Mar 30, 2015

ncdc commented Mar 30, 2015

smarterclayton commented Mar 30, 2015

ncdc commented Mar 30, 2015

smarterclayton commented Mar 30, 2015

ncdc commented Mar 30, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

smarterclayton commented Mar 31, 2015

ironcladlou commented Apr 1, 2015

ironcladlou commented Apr 1, 2015

ironcladlou commented Apr 1, 2015

ironcladlou commented Apr 1, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

smarterclayton commented Apr 1, 2015

ironcladlou commented Apr 1, 2015

smarterclayton commented Apr 1, 2015

openshift-bot commented Apr 1, 2015

openshift-bot commented Apr 2, 2015

openshift-bot commented Apr 2, 2015