

(See Job Lifecycle for an explanation of when drain scripts run.)

Release job can have a drain script that will run when the job is restarted or stopped. This script allows the job to clean up and get into a state where it can be safely stopped. For example: - when writing a release for a load balancer, each node can safely stop accepting new connections and drain existing connections before fully stopping. - when writing a release for a database or other distributed datastore, you can ensure that data is correctly replicated/distributed on all nodes before stopping the node.

Job Configuration¶

To add a drain script to a release job:

Create a script with any name in the templates directory of a release job.
In the templates section of the release job spec file, add the script name and the bin/drain directory as a key value pair.

Example:

---
name: nginx
templates:
  drain-web-requests.erb: bin/drain

Note

Drain script from each release job will run if they are deployed on 3093+ stemcells. Before only the first release job's drain script ran.

Script Implementation¶

Drain script is usually just a regular shell script. Since drain script is executed in a similar way as other release job scripts (start, stop, pre-start scripts) you can use job's package dependencies.

Drain script should be idempotent. It may be called multiple times before or after process is stopped.

You must ensure that your drain script exits in one of following ways:

exit with a non-0 exit code to indicate drain script failed
exit with 0 exit code and also print an integer followed by a newline to stdout (nothing else must be printed to stdout):

static draining: If the drain script prints a zero or a positive integer, BOSH sleeps for that many seconds before continuing.

dynamic draining: If the drain script prints a negative integer, BOSH sleeps for that many seconds, then calls the drain script again.

Note

BOSH re-runs a script indefinitely as long as the script exits with a exit code 0 and outputs a negative integer.

Tip

It's recommended to only use static draining as dynamic draining will be eventually deprecated. If you can't provide an upper bound on how long BOSH should wait before continuing (as required in case of static draining) you can sleep/retry inside the drain script since BOSH guarantees that drain scripts will not timeout.

Note that if drain script causes monitored job processes to exit, monit will not call stop script for that job.

Environment Variables¶

Drain script can access the following environment variables:

BOSH_JOB_STATE: JSON description of the current job state
BOSH_JOB_NEXT_STATE: JSON description of the new job state that is being applied

Currently, only persistent disk size is provided in those two variables. For example:

{"persistent_disk":2048}

With this, the drain script can determine when the size of the persistent disk changes and take action.

But more importantly, it can detect when the current node is being deleted. This is important when the node is part of a cluster and needs to gracefully say goodbye to its pairs before leaving forever.

When the drain script is run before the node is deleted, then the new persistent disk size is zero. For example, you would be able to see these values when echoing them.

cat /var/vcap/jobs/my-job/bon/drain
(
  echo BOSH_JOB_STATE=$BOSH_JOB_STATE
  echo BOSH_JOB_NEXT_STATE=$BOSH_JOB_NEXT_STATE
) > /var/vcap/sys/log/my-job/drain.stdout.log

cat /var/vcap/sys/log/my-job/drain.stdout.log
# BOSH_JOB_STATE={"persistent_disk":2048}
# BOSH_JOB_NEXT_STATE={"persistent_disk":0}

You'll find here an example script for an etcd member to leave its etcd cluster gracefully.

Command-line arguments¶

The first argument passed to the drain script indicates the intended job lifecycle action. It can have the following values (note: "job" here actually means "instance"): * job_changed indicating that the instance will be restarted * job_shutdown indicating that the instance will be stopped and subsequently the VM will terminated

The second argument passed to the drain script indicates whether the job hash has changed. It can have the following values: * hash_changed indicating that the job's packages or rendered templates have changed. * hash_unchanged indicating that the job's packages and rendered templates have not changed.

Logs¶

Currently logs from the drain script are not saved on disk by default, though release author may choose to do so explicitly. We are planning to eventually make it more consistent with pre-start script logging.

Examples¶

Load-balancer¶

#!/bin/bash

# check if the process is running
pid_path=/var/vcap/sys/run/worker/worker.pid
if [ ! -f "$pid_path" ]; then echo 0; exit 0; fi
pid=$(<"$pid_path")
if ! ps -p "$pid" >/dev/null; then echo 0; exit 0; fi

# process is running; send signal to instruct the process to drain active 
# connections and to steer traffic to the other instances of this job
kill -USR2 $pid

# we know that the operation above takes 15 seconds to complete; inform BOSH 
# that it can continue after waiting 15 seconds
echo 15; exit 0

Stateful distributed job¶

#!/bin/bash

# check if the process is running
pid_path=/var/vcap/sys/run/worker/worker.pid
if [ ! -f "$pid_path" ]; then echo 0; exit 0; fi
pid=$(<"$pid_path")
if ! ps -p "$pid" >/dev/null; then echo 0; exit 0; fi

function cluster_in_sync() {
  # run command to check if the data in the cluster is correctly replicated
}

function heal_cluster() {
  # run command to ask the cluster to heal its internal state
}

# check if the cluster is in sync. if it is not in sync trigger a healing
# operation and wait until healing is complete and the cluster is in sync
if ! cluster_in_sync; then
  heal_cluster
  while ! cluster_in_sync; do
    sleep 10
  done
fi

# ask the process to terminate, and wait until the process exits
kill "$pid"
while ps -p "$pid" >/dev/null; then sleep 1; fi

# the cluster was in sync and the local node has been shut down; tell BOSH to
# proceed with the operation
echo 0; exit 0