pod lifecycles

Tags: kubernetes (k8s)

Three states

• waiting
• running
• terminated
  • use kubectl describe

Terminating Gracefully

• Cleanups?
• Resources?
• Forced termination?
• SIGTERM/SIGKILL
• Lifecycle
  1. Set the grace period -> terminating state -> traffic
  2. preStop hook -> grace of 2 seconds
  3. send SIGTERM to pid 1
  4. grace period sends SIGKILL
  5. API server deletes the pod’s API object

Health Probes

• Kubelet brings up the containers and keeps them running, also responsible for restarts
• uses some Process Health Checks to check the status
• Stalls?
• Types of probes
  • Exec
    • Runs every so often, checks for the exit code
  • TCP socket
  • HTTP get
• Liveness Probes
  • Check if the container is alive or dead
  • Always define a liveness probe for prod apps
  • Have the application expose a health-check API that doesn’t require auth
  • Keep probes light as CPU resources
• Readniess Probes
• Startup probes
  • Indicates whether the container has started
  • Disables other probes until this is done

Hooks

postStart hook

• Runs after container is created, async
• No guarantees of running
• postStart is a blocking call
• container remains in Waiting until the handler completes

preStop

• Call sent to container before terminated
• use when reacting to SIGTERM
• always ends before SIGKILL is sent
• used for graceful termination

initContainers

• runs various preconditions before the container startup
• exaples
  • DB, create user accounts, DB migrations

Scheduling and resources

• Effective request/limit is the max(sum request/limit for all the application containers, effective request limit)

Init vs postStart hook vs Startup Probe

• Probes and hooks run in the same container, init containers run on the whole pod
• hooks have no guarantees
• usages:
  • postStart
    • signal to external listeners
  • init
    • initalization
  • probes
    • slow starting containers