Speed and Optimization¶
This page contains information and guidelines for improving the speed, stability, and general optimization of your JupyterHub deployment.
Picking a Scheduler Strategy¶
Kubernetes offers very flexible ways to determine how it distributes pods on your nodes. The JupyterHub helm chart supports two common configurations, see below for a brief description of each.
Spread¶
- Behavior: This spreads user pods across as many nodes as possible.
- Benefits: A single node going down will not affect too many users. If you do not have explicit memory & cpu limits, this strategy also allows your users the most efficient use of RAM & CPU.
- Drawbacks: This strategy is less efficient when used with autoscaling.
This is the default strategy. To explicitly specify it, use the following in
your config.yaml:
singleuser:
schedulerStrategy: spread
Pack¶
- Behavior: This packs user pods into as few nodes as possible.
- Benefits: This reduces your resource utilization, which is useful in conjunction with autoscalers.
- Drawbacks: A single node going down might affect more user pods than using a “spread” strategy (depending on the node).
When you use this strategy, you should specify limits and guarantees for memory and cpu. This will make your users’ experience more predictable.
To explicitly specify this strategy, use the following in your config.yaml:
singleuser:
schedulerStrategy: pack
Pre-pulling¶
Pulling a user’s images to a node forces a user to wait before the user’s server is started. Sometimes, the wait can be 5 to 10 minutes. Pre-pulling the images on all the nodes can cut this wait time to a few seconds. Let’s look at how pre-pulling works.
Hook - image pulling before upgrades¶
With the pre-pulling hook, which is enabled by default, the user’s container
image is pulled on all nodes whenever a helm install or helm upgrade is
performed. While this causes helm install and helm upgrade to take several
minutes as the update is scheduled after the pulling has completed, the users
waiting time will decrease and become more reliable.
With the default helm upgrade settings, a helm install or helm upgrade will
allow 5 minutes of image pulling before timing out. This wait time is
configurable by passing the --wait <seconds> flag to the helm commands.
We recommend using pre-pulling. For the rare cases where you have a good reason
to disable it, pre-pulling can be disabled. To disable the pre-pulling during
helm install and helm upgrade, you can use the following snippet in your
config.yaml:
prePuller:
hook:
enabled: false
Continuous - image pulling for added nodes¶
Cluster size can change through manual addition of nodes or autoscaling. When a new node is added to the cluster, the new node does not yet have the user image. A user using this new node would be forced to wait while the image is pulled from scratch. Ideally, it would be helpful to pre-pull images when the new node is added to the cluster.
With the continuous pre-puller enabled (disabled by default), the user’s container image will be pre-pulled when a new node is added. New nodes can for example be added manually or by a cluster autoscaler. The continuous pre-puller uses a daemonset to force kubernetes to pull the user image on all nodes as soon as a node is present. The continuous pre-puller uses minimal resources on all nodes and greatly speeds up the user pod start time.
The continuous pre-puller is disabled by default. To enable it, use the
following snippet in your config.yaml:
prePuller:
continuous:
enabled: true
Pre-pulling additional images¶
By default, the pre-puller only pulls the singleuser image & the networktools image (if access to cloud metadata is disabled). If you have customizations that need additional images present on all nodes, you can ask the pre-puller to also pull an arbitrary number of additional images.
prePuller:
extraImages:
ubuntu-xenial:
name: ubuntu
tag: 16.04
policy: IfNotPresent
This snippet will pre-pull the ubuntu:16.04 image on all nodes, for example.
You can pre-pull any number of images.
