This exemplar BOSH release is intended as a collection of recommended practices for authoring a BOSH release, with a particular focus on jobs. Not all advice here should be taken verbatim for all releases, so we have made an attempt to explain the reasoning behind all recommendations so that the conscientious release author can adapt the recommendations to their unique circumstances.
- Order Independence
- BOSH Job Lifecycle
- Advice for authoring
specfiles - Template Advice
- Monit Advice
- Exporting Logs
- Bosh DNS
- Backup and Restore features
The primary goals of a BOSH job are to maximize availability, remove the need for manual operation, and contribute to quick deployments. To accomplish this, it is important to understand the BOSH deployment lifecycle and what your job should be doing to take advantage of it.
A process should be able to start and stay running without dependencies being ready. Only once its required dependencies (those needed to respond to requests) are reporting as healthy should it report as healthy. This allows jobs to be started in any order and then propagate health upwards through their dependents when they're ready. We think removing this burden of job ordering from the operator reduces complexity and increases system reliability.
On the other hand, stopping can follow a similar but reversed pattern. If your service is being load-balanced then you
can use a signal in your drain script to mark yourself as unhealthy. After a period of time or after no more requests
are detected you can stop the process completely. If you aren't behind a load balancer then you will have to find
another way to tell upstream components about your imminent demise.
BPM stands for BOSH Process Manager. You can get more information about it in the BPM chapter of the BOSH documentation. The BPM release repository is also a great place where to find more information about BPM.
We provide here a sample-app job that shows how simple things can be when delegating daemons
management to BPM. That said, BPM doesn't provide an answer to all use cases yet. Classical BOSH pre-start and drain
scripts are still necessary for the use cases documented below.
One note about BPM pre_start hooks: contrarily to BOSH pre-start, BPM pre_start scripts are run at every
monit start, i.e. even when the BOSH instance gets back up after being evicted by the underlying Cloud
infrastructure. Actually, you should put in your BPM pre_start what you would have put in your start script before
exec-ing your process.
Every BOSH job goes through a specific lifecycle for starting and stopping. The BOSH website has a great overview of the lifecycle. You may wish to review this before reading on to our recommendations for the individual parts of the lifecycle.
Pre-Start (docs)
Pre-start scripts are run before BOSH hands off control to Monit, so will not necessarily run every time a job starts
(e.g. if a VM is rebooted outside of BOSH's control). However, pre-start scripts will run at least once for every new
release version that is deployed, and have no timeout (in contrast to a start script). Therefore, pre-start
scripts can best be used for performing lengthy set-up of persistent state that must be done for a new version of a
release, such as database migrations. However, because the pre-start script does not necessarily run in every VM your
job will run in, do not perform any work in temporary directories such as /var/vcap/sys/run (see VM Configuration
Locations for the list of temporary directories).
In general, a pre-start script should not be necessary and the start script should be sufficient. If you find a
pre-start script is necessary, keep the above caveats in mind.
The start script has two main responsibilities: writing the process ID (PID) to a .pid file and starting the main
process. All setup required for the process to run (that has not been done in the pre-start script) must be done at
this time. Monit places a short timeout on the .pid file being written, so the work done in the start script should
be focused on executing your process and getting it healthy quickly.
Note: This script will likely run many times, so ensure that it is idempotent (that it can be run repeatedly without causing the system to enter a bad state)
The general workflow of a start script:
-
Create a log directory for log persistence in
/var/vcap/sys/log/<job name>- Idempotency recommendation: Do not fail if the log directory already exists
-
Create a run directory to contain the pidfile in
/var/vcap/sys/run/<job name>- Idempotency recommendation: Do not fail if the run directory already exists
-
Change ownership permissions on the
log/andrun/directories to uservcapand groupvcap -
Run your process with
start-stop-daemon, which will manage your process's pidfile, ensuring multiple instances of the process are not run simultaneously. A recommended usage ofstart-stop-daemonlooks like this:/sbin/start-stop-daemon \ # Write the pidfile, and error if it exists # and refers to an already-running process --pidfile "$PIDFILE" \ --make-pidfile \ # Run the process as the less-privileged vcap user --chuid vcap:vcap \ # Start the given process and redirect its logs --start \ --exec /var/vcap/packages/paragon/bin/web \ >> "$LOG_DIR/web.stdout.log" \ 2>> "$LOG_DIR/web.stderr.log"Refer to the start script of the pararagon job for a more-complete example.
Note: The start script is executed as root. Do not assume you can only break your own process.
Note: If your process cannot start quickly, consider moving long-running tasks to pre-start or post-start.
Note for Windows Releases: The Windows BOSH Agent does not use Monit and manages starting the script directly. If
your pre-start and post-start scripts have been written following the guidance in this document, the Windows BOSH
Agent will be able to start your process correctly.
Job logs should be put in the /var/vcap/sys/log/<job> directory. You can redirect the logs from your server (if
they're output on stdout and stderr) by using standard redirection:
exec ... \
>> "/var/vcap/sys/log/<job>/<process>.stdout.log" \
2>> "/var/vcap/sys/log/<job>/<process>.stderr.log"If your process takes log locations as parameters, use that instead:
exec ... \
--out-log-file "/var/vcap/sys/log/<job>/<process>.stdout.log" \
--err-log-file "/var/vcap/sys/log/<job>/<process>.stderr.log"What is important here, is that all log files must end with the .log extention in order to be properly rotated by
BOSH. If any other extension is used for files where daemon append logs, you might end up with full disks and experience
failing nodes. For those who are interested to know more, BOSH enforces a Logrotate config
in /etc/logrotate.d/vcap.
For detail on forwarding logs to external locations, please see Exporting Logs.
Logging the activity of start, stop is not critical when leveraging start-stop-daemon because they become really
simple. Some pre-start or post-start script are not trivial though, and need logging for debug purpose. In such
case, we recommend you prepend some time indication to the logged lines. This is useful for debugging nodes that have
been deployed for a long time, in order to distinguish old errors from recent errors.
function start_logging() {
exec \
> >(prepend_datetime >> /var/vcap/sys/log/<job>/<process>.stdout.log) \
2> >(prepend_datetime >> /var/vcap/sys/log/<job>/<process>.stderr.log)
}
function prepend_datetime() {
awk -W interactive '{ system("echo -n [$(date +%FT%T%z)]"); print " " $0 }'
}
function main() {
start_logging
...
}
main "$@"As a good remark, you'll note that the above code starts multiple processes for every line of logs. Yes indeed, this way
of doing is not to be applied to all cases. Here, starting multiple processes for every line of logs is acceptable
though, because scripts like start, stop, pre- start, post-start or drain are not supposed to be executed
often or to produce massive amounts of logs.
Post-Start (docs)
post-start is useful for custom health-checks to ensure you job has started correctly. For example, if your process
starts quickly, but takes time to discover services or connect to its backend, you may wish to use post-start to query
the readiness of your job to start handling requests. If a post-start script is provided, BOSH will not consider a job
to be ready until it has exited successfully.
When ensuring in a post-start that your job's is healthy and its dependencies are met, you are responsible for
implementing the proper timeouts. Most of the time, these timeouts will need to be configurable in order for your job to
adapt to slow or loaded IaaS infrastructures. A good way to do this is to expose a timeout scale factor in your job's
configuration properties.
Post-Deploy (docs)
post-deploy is useful for tasks that must occur once all instances of a deployment are started, such as checking the
health of an entire deployment, enabling a feature after a successful upgrade, or finishing database migrations for an
entire cluster.
Job is unmonitored before any stop scripts can run, so you can safely exit in either pre-stop, drain or monit stop
without the job becoming listed as unhealthy.
Pre-Stop (docs)
The feature for pre-stop hook scripts has been released in April 2019. The pre-stop script runs right
before drain, with no timeout, and is exposed more information to help release authors take proper decision about
future state of the bosh instance, whether it is going away temporarily of permanently.
When pre-stop scripts are executed, daemon processes are no more monitored by Monit, so pre-stop is the proper place
for draining client connections, and should be favored over using the drain script, that abides to legacy complex
rules (related to “static” or “dynamic” draining modes, that should be avoided).
The tricks in drain scripts based on BOSH_JOB_NEXT_STATE (see below) should be abandonned in favor of the newer
BOSH_VM_NEXT_STATE, BOSH_INSTANCE_NEXT_STATE and BOSH_DEPLOYMENT_NEXT_STATE environment variables, that are now
exposed to pre-stop scripts. See the documentation for more details
Drain (docs)
Drain scripts are the legacy hooks into the BOSH job lifecycle that are run before stopping the job via Monit. They've
typically been used for services which must perform some work before being shut down, for example flushing a request
queue or evacuating containers from a Diego cell. Now such things should be implemented in the newer pre-stop script.
From this point on, and until the end of this section, everything that is said about drain should be ported to
pre-stop instead.
As a rule of thumb, if monit stop-ping your job could cause dropped connections or a lack of availability, a drain
script should be used to prevent this. Most commonly, your drain script will send a request to a drain endpoint on
your process and wait for it to return rather than implementing the drain behavior itself.
A concrete example is the gorouter, which has a configurable drain_wait parameter. When
non-zero, gorouter's drain script will instruct gorouter to report itself as unhealthy to its load-balancer with the
intent of being removed from the balanced instance group before shutting down and rejecting requests. When monit stop
is called, the router will already be receiving no connections, so will not drop connections when it is shut down
quickly. This is the lame duck pattern.
Drain scripts have no timeout, so should take whatever time necessary to block on any draining work. One may, however,
run the risk of writing a drain script that never finishes and blocks a deployment. A well-written drain script is
guaranteed to finish, such as by adding your own sensible timeout around draining work (e.g. 10 minutes).
When your draining process has completed, your drain script should output a "0" to STDOUT to inform BOSH that draining is complete.
Open question: If your job needs to wait for another job to drain first, what is the best way to block on that? If you have any ideas, please let us know.
In a drain script, you might need to tell the difference between your process being temporarily stopped, or
permanently decommissioned. This is especially relevant for cluster nodes that share a global state. A node that
permanently leaves the cluster might have to re-distribute its data to the other nodes that stay in the cluster (this is
for instance the case with Cassandra nodes) or just tell the others that it's no use waiting for it to come back
(MongoDB calls this “stepping down”).
In such case, BOSH proposed a hack-ish check on the permanent disk size being set to zero in the future. This
information is provided by the JSON document exposed by the BOSH_JOB_NEXT_STATE environment variable, as detailed in
the Environment Variables section of the drain script
documentation.
You can see two examples of such scripts that follow this best practice:
- For etcd, in cloudfoundry-incubator/cfcr-etcd-release/jobs/etcd/templates/bin/drain.erb
- For Cassandra, in orange-cloudfoundry/cassandra-boshrelease/jobs/cassandra/templates/bin/drain
The stop program in your job's monit file is executed after the drain script (if present) has finished running. It
can also be run directly by an operator if they execute monit stop <job> on the machine (though operators should
always consider running bosh stop <instance-group>/<index> instead of using monit stop directly). By default, there
is a 30 second timeout on this script completing. Monit will assume scripts taking longer than this have failed. We do
not recommend changing this value if you need more time. Instead, you should do all the work necessary in your drain
script such that your service can shutdown quickly (where “quickly” generally means in under 10 seconds).
If you're not using drain, then we recommend that you send SIGTERM to your process, which should cause it to start
shutting down. If your process shuts down at this point then you're good to go. If for some reason your process locks up
or is unable to exit for some other reason then you should send SIGKILL before the timeout. If your language runtime
supports dumping the stacks of all running threads on SIGQUIT (Go and Java do) then you can send that signal just
before the SIGKILL to aid debugging why the process is stuck. If you are not using a SIGKILL respecting runtime then
adding this functionality to your own program normally isn't difficult.
start-stop-daemon can be used for this:
/sbin/start-stop-daemon \
# Remove the pidfile after killing the process
--pidfile "$PIDFILE" \
--remove-pidfile \
# Send SIGTERM, wait for the process to die for 20 seconds
# If the process has not died, send SIGQUIT and wait for 1 second
# If the process has still not died, send SIGKILL
--retry TERM/20/QUIT/1/KILL \
# If the process is already gone, do not error
--oknodo \
--stop
Refer to the stop script of the pararagon job for a more-complete example.
If you're using drain to kill the process then your process may already be shut down by the time that monit stop is
called. In this case we do not need to do anything further in the stop executable.
Note for Windows Releases: The Windows BOSH Agent does not use Monit and manages starting the script directly. If
your pre-start and post-start scripts have been written following the guidance in this document, the Windows BOSH
Agent will be able to stop your process correctly.
Post-Stop (docs)
The feature for post-stop hook scripts has been released in March 2018. When post-stop is run, the
Monit process is gone. post-stop may possibly cleanup some data before the VM is shut down. post-stop is not exposed
the same information as pre-stop for taking decisionsuppon permanent deletion of the VM.
We haven't see so far any use-case that requires running code in the post-stop script specifically, but we'd lov to
hear your feedback and get suggestions/contributions in PRs.
The properties you include in your spec file create the product surface for your job from the perspective of the
operator component. Job properties can make it much easier or much harder to operate a deployment, so one should be
mindful of the operator when making decisions about properties.
- Properties should not have a “namespace” for the job itself (e.g.
my-job.port,my-job.hostname), but should only use namespace to create property groups that the job cares about (e.g.database.*andblobstore.*). - If a property does not need to be configured specially for every deployment and a reasonable default exists, it should be provided. This lets an operator have a terser deployment manifest, which is easier to generate, read, and modify.
- Include a description for every property whenever it is slightly likely that it would help with understanding. It can be easy for an operator to misinterpret a property and use it incorrectly.
Here is an example of properly written description, from the ATC job of Concourse
external_url:
description: |
Externally reachable URL of the ATCs. Required for OAuth. This will be
auto-generated using the IP of each ATC VM if not specified, however
this is only a reasonable default if you have a single instance.
Typically this is the URL that you as a user would use to reach your CI.
For multiple ATCs it would go to some sort of load balancer.
example: https://ci.concourse.ciYou can find other good examples for Cassandra or MongoDB. Generally,
using the pipe | notation for the description: YAML property, and providing a text with lines wrapped around the
80th column, is best for readability.
Links (docs)
Links allow jobs to provide and consume configuration that needs to be shared between jobs, which can greatly reduce the amount of configuration required in a deployment manifest. Instead, jobs can declare which information they need and BOSH will provide that information automatically. When links are used correctly, there are two main benefits:
- IP address information (e.g. locating servers from other jobs) does not need to be provided at all by the operator, so default manifests are intrinsically much more network-agnostic and portable.
- Configuration can be provided in a manifest to only one job, which will then export it as a link consumed by downstream jobs, making it much easier to modify that configuration and reduce human error. For example, if the port a server listens on is provided as a link, a change to the server's configuration will automatically propagate to all jobs that need to communicate with it.
If links are used whenever a job depends on configuration from another job, manifests can become much simpler and
deployments can become more reliable. When authoring your release, all properties you include in your spec file should
change the runtime behavior of your process and should not include IP addresses, domain names, credentials, or
certificates of other jobs. If you are including these properties, they are candidates for links instead. If the job
you depend on does not provide the information as a link, please consider submitting an issue or PR to the maintainer.
Sometimes it is necessary to allow operators to override individual properties within a link. For example, if your job uses a link's IP information to find a dependent service, but a particular deployment may be using custom DNS for service discovery, your job template could prefer the DNS property over the link. That template could look like this:
<%
require "json"
server = nil
if_p("database_location") do |prop|
server = prop
end.else do
server = link("database").instances[0].address
end
-%>
server: <%= server.to_json %>Any template in your job can use ERB (even the monit files!). While this is extremely powerful it can be very
difficult to understand and maintain complex templates. Therefore, we advise avoiding any ERB in your control scripts
wherever possible. The control flow of starting and stopping your program should be deterministic and simple. All ERB
should be relegated to static configuration files so that properties can be interpolated.
When you inject configuration properties with the ERB syntax in shell scripts or configuration files, you should take great care for properly encoding special caracters that are meaningful to the target syntax. Failing at doing so could cause shell scripts to erase critical data, or mis-configuration in YAML files to cause data loss.
With Bash shell scripts, we recommend encoding properties using shellwords:
#!/usr/bin/env bash
<%
require "shellwords"
def esc(x)
Shellwords.shellescape(x)
end
-%>
some_property=<%= esc(p('some_property')) %> # <- Please observe that you must not quote the value hereWith YAML config files, we recommend encoding properties with json:
---
<% require "json" -%>
# The name of the cluster. This is mainly used to prevent machines in
# one logical cluster from joining another.
cluster_name: <%= p("cluster_name").to_json %> # <- Please observe that you must not quote the value hereFor complex logic in YAML templates, you can adopt the full-Ruby option, where you
first build a Ruby Hash for your YAML configuration, and then YAML.dump() it, which properly performs the expected
encoding.
Don't hesitate to contact us and submit advices or best-practice encoding for other syntax schemes.
Your templates should be simple enough to not need testing, such as by doing a simple passthrough to property parsing
in your own code, but you may find yourself with more-complex templating needs, such as when writing a release wrapping
third-party code. Testing BOSH templates is classically a hard topic, but it has been greatly simplified with a new
version of the bosh-template Gem, and a proper documentation of unit testing with it.
As it might still be difficult to test BOSH template rendering, you could also imagine having a simple template (e.g.
<% p('my-properties').to_json %>) that you would transform with a custom executable, which you can test in your
standard unit-testing workflow.
We advise you to use the #!/usr/bin/env bash sheebang variant.
The ShellCheck tool can be used to find common errors in your bash scripts. This yet another reason to
keep ERB out of your scripts!
You can also possibly use BATS unit tests, or similar testing frameworks.
Try to keep your Monit configuration as simple as possible. The BOSH team is planning on removing Monit soon - do not couple yourself to it! This includes such changes as monitoring memory and changing away from the default timeout.
You must specify group vcap in your monit file because the agent uses this tag to find
processes it should be managing on the machine.
BOSH itself does not handle forwarding logs off-system. If you have written your logs appropriately as described in the Logging section, operators can choose the correct log-forwarding mechanism for their deployment using job co-location in the deployment manifest or BOSH addons. Operators may choose to use something like google-fluentd or syslog-release. If you are also responsible for providing a deployment manifest generation tool, you may wish to provide the option to add syslog-release forwarding to all components.
Backwards Compatibility Warning: If your release is currently responsible for forwarding logs off-system, by following this guide and removing that functionality, you are putting the onus for logging on deployment authors (e.g. cf-deployment or your local friendly closed-source proprietary offering) to configure logging. If your release has traditionally done this, deployment authors may not know how to update the deployment and logs could be lost. If you are removing this functionality, please coordinate appropriately with deployment authors and operators.
metron_agent has functionality to forward syslogs, but this has been superseded by syslog-release. metron_agent
should only be used to forward logs to loggregator.
Many releases currently include complex setup to forward logs to both /var/vcap/sys/log as well as syslog. This
is not necessary if operators use one of the log-forwarding options mentioned above. Drawbacks of the tee and logger
approach include starting multiple processes for every line of logs, fairly subtle behavior that can easily break or
security vulnerabilities. If your release has any code like the following, please remove it and follow the
recommendations above:
DO NOT USE THIS CODE
exec > \
>(
tee -a >(logger -p user.info -t vcap.$(basename $0).stdout) | \
awk -W interactive '{ gsub(/\\n/, ""); system("echo -n [$(date +\"%Y-%m-%d %H:%M:%S%z\")]"); print " " $0 }' \
>> /var/vcap/sys/log/<job>/<process>.stdout.log
)
exec 2> \
>(
tee -a >(logger -p user.error -t vcap.$(basename $0).stderr) | \
awk -W interactive '{ gsub(/\\n/, ""); system("echo -n [$(date +\"%Y-%m-%d %H:%M:%S%z\")]"); print " " $0 }' \
>> /var/vcap/sys/log/<job>/<process>.stderr.log
)
DO NOT USE THIS CODEIf you must use this, you should understand it. The exec calls redirect STDOUT and STDERR respectively,
sending them to a sub-shell that calls tee. tee splits the output to 1) syslog via logger and 2) the BOSH log
directory (but not before appending timestamps with awk).
Bosh DNS is an important feature in Bosh. It basically brings consistent cloud-agnostic internal load-balancer features for free.
Bosh DNS is also particularily powerful for implementing cloud-agnostic split-DNS features. Internally, Bosh DNS can mask public DNS names, and return private IP addresses instead. With this, the cluster can use public DNS names and still get the traffic routed internally. Without it, the traffic would exit the cluster and immediately re-enter it, resulting in unnecessary network hops.
Release jobs can define their own DNS aliases in a dns/aliases.json file. Here we
demonstrate how to write such file. Indeed, when referring to
default Bopsh DNS names, you'll need to get the network name from Bosh Cloud Config, and there is a trick for this with
the .methods(false) helper on the .networks OpenStruct.
The Bosh dns/health script is actually misnamed, as it is actually a “readiness” check script. Indeed, when a
bin/dns/healthy script exits with non-zero (error) status, then the instance is taken out of DNS aliases. When you see
DNS aliases as cheap load-balancers, then it means that you stop routing traffic to the job. This is precisely the
concept of a “readiness” check.
Inside a bin/dns/healthy script, you can do whatever check you need to. Run binaries, check open TCP ports or run
healthcheck HTTP(S) queries.
Particularily interesting is the way you can coordinate job draining in pre-stop with dns/healthy for properly
draining connections to a network daemon that is about to shut down. Indeed, the pre-stop can “inform” dns/healthy
that no more traffic should be routed to the daemon, creating a specific file for example. Then, the dns/healthy
script can react and return non-zero exit status, resulting in the IP being removed from the DNS query results, and no
more new connection sent to the daemon.
If we detail a bit, the pre-stop starts draining connections. So, it ensures that Bosh DNS doesn't route any new
connection to it writing a file like /var/vcap/data/<job-name>/DRAINING. The dns/healthy script detects this file,
and exits unconditionally with a status of 1 in order to report the job as non-ready.
See this basic exemple from the Kong Bosh Release.
The BOSH Backup and Restore project (BBR) provides a framework for backing up and restoring BOSH deployments and BOSH Directors.
For more information, go read the BOSH Backup and Restore chapter in Cloud Foundry documentation.
As a BOSH Release author, you shall also be interested in reading the BOSH Backup and Restore Developer's Guide.
The BBR project provides its own
Exemplar Backup and Restore Release for
the purpose of demonstrating best practice in implementing the
BBR contract. We advise you to refer to this materials for
learning more about providing standard backup and restore features in your BOSH Releases. Here we just provide an
example stateful-daemon job which implements very basic BBR features.