A ready pod which can be specialized was fetched for every function earlier, this has been changed to a queue and cache implementation in client-go to improve performance.
Poolmanager when tested at high load had some issues and this PR fixes one set of them which were found so far.
Co-authored-by: Vishal <vishal-biyani@users.noreply.github.com>
Concurrency in the pool manager allows specializing pods concurrently based on a specified limit.
Co-authored-by: Vishal <vishal-biyani@users.noreply.github.com>
Attribute enableServiceLinks can be now configured in PodSpec so that environment variables are not injected in function pods and DNS is used for discovery
Executor wrongly deletes role bindings if the user
creates an environment in the reserved namespaces. This
PR is a quick fix to solve the problem by checking if
an environment is under reserved namespaces.
The root cause of the problem is that eagerPoolCreator
tries to create the deployment when the poolmanager
is trying to delete it. To avoid this, start eager pool
creator after executor starts serving requests.
The pod template is embedded inside the deployment. So if
the pod annotation contains instance-id, the deployment
will get updated and thus triggers a rolling update whenever
a new executor starts which is unwanted.
After this PR, poolmanager will patches instance-id when a
pod is chosen for a function.
For newdeploy, unlike poolmanager manages the lifecycle
of function pod directly, newdeploy is only responsible
to create the deployment so we append instance-id to top-
level controller (deployment) only.
We used to update timestamp in the deployment environment field
in order to trigger a rolling update when the function referenced
resources get updated. However, use timestamp means we are not
able to avoid triggering a rolling update when executor tries to adopt
orphaned deployment due to timestamp changed which is unwanted.
In order to let executor adopt deployment without triggering a rolling
update, we need an identical way to get a value that can reflect res-
ources changed without affecting by time.
To achieve this goal, the sum of the resource version of all referenced
resources is a good fit for our scenario since the sum of the resource
version is always the same as long as no resources changed.
When a new executor starts up, it adopts the orphan kubernetes resources created
by the old executor instance. However, the adopted resource won't reflect the changes
come with the new executor, for example, the fetcher image inside won't be changed.
To solve this, executor updates the resource spec (HPA/Deployment/Service) with the
latest resources spec. By doing this, we can prevent the inconsistency between resources
created by different executor instance, also minimizes the impact on users.
Previously, once the executor is deleted for reasons (like upgrade or cluster scale-in),
the new executor deletes all existing resources created by the old executor and creates
new one. This mechanism becomes a problem when there are requests connecting to the
existing pods. Also in the worst case, the cluster may not have enough resources to create
new pods and cause service downtime.
This PR let each executor type adopts existing resources before starting the executor
API services, and so the alive connections won't experience failure. However, the requests
send to the function that doesn't have alive function pods will still fail due to the
executor is in bootstrapping.
The pool manager keeps terminating function pod periodically even there are
traffic to the function. The root cause is that executor, poolmgr, newdeploy
manage their own functionServiceCache separately. And when router taps a
function, executor updates the access time of the function service entry in its
own cache without notifying executor types to do the update as well. Hence,
the access time of function service entry in poolmanager cache never gets updated.
Due to the access time never gets updated, the idle pod reaper in poolmanager
then thinks the function pod is in idle state and recycle it.
This PR removes the cache in executor itself, and when router tries to tap a function,
executor will call executor type to tap the function and update access time.
The router taps function service one by one which is inefficient and
increases the burden of executor. This PR aggregates all requests into
one to solve the problem mentioned above.
When a function is created before the creation of the environment it's used, the newdeploy will not be able to create kube objs. Hence no function service record is inserted into the cache.
When getFuncSvc is called, the newdeploy tries to find the record in service cache in order to create kube objs with the same name used in previous kubeobjs creation. However, due to no record in the cache, a NotFound error is returned directly and causes the problem. To solve this, we use fn meta UID to ensure we always get the same obj name instead of getting it from the cache.
When a function with executor type newdeploy got created, Newdeploy
is expected to create deployment/HPA/service for it and insert serviceEntry
to the cache for later use. Once clients call the function, newdeploy returns
the serviceEntry to the router.
However, the log shows that the newdeploy was unable to find the entry and
prints "Resource not found - key 'xxx' not found". The root cause is that the
informer controller instead of processing items in parallel, it dispatches XXFunc
to process items one by one. So if there is any problem during the creation of the
kubernetes resource, it takes a longer time to process the next item and hence
the serviceEntry was not inserted before clients call the function.
This PR lets the newdeploy to process items in extra goroutines instead of blocking
the process loop. It's a workaround to solve the problem above, we should consider
using workqueue to solve it in the future.
If a user deploys fission in the namespace which is different from the one in the single YAML file generated by helm, fission components won't be able to talk to each other due to the wrong namespace appends after the service address.
This PR adds --namespace when generating the YAML file to prevent the mismatch problem.
The merge function executor used wasn't merge container correctly, and it
didn't merge all fields in spec except volumeMount & Env which confused people.
To apply the user-configured container correctly, this PR changes the way of merge
and follows rules:
1. Slices are merged and return an error if the elements in the slice have name conflicts.
2. Maps are merged, the value of map of dst container are overridden if the key is the same.
3. The rest of the fields of dst container are overridden directly.
In 1.9, kubernetes prompted Deployment to apps/v1 and finally deprecated
extensions/v1beta1 support in 1.16. Users will see the following error message
in log when fission components try to submit Deployment to the k8s API server
and specify extensions/v1beta1 as API group.
error: unable to recognize "deployment": no matches for kind "Deployment" in version "extensions/v1beta1"
This PR updates call to k8s API server and apiKind in chart files to avoid fission failure in 1.16.
