This repository is part of the Joyent Manta project. For contribution guidelines, issues, and general documentation, visit the main Manta project page.
Marlin is the Manta compute engine. It comprises a service, an agent, and a set of client tools for executing arbitrary user jobs on objects stored in Manta.
This repository is part of mantav1, the long term support maintenance version
of Manta. Development is done on the mantav1
branch, the master
branch is no longer used. See the mantav2 overview
document for
details on major Manta versions.
For overview on the purpose, design, and usage of Marlin, see docs/index.restdown (or https://mo.joyent.com/docs/marlin).
This document describes the development workflow.
Since it's useful to run the Marlin client tools (mrjob, mrjobreport, mrerrors, etc.) on an OS X laptop, the build system includes hacks to allow just those bits to build on OS X. To use these tools:
$ git clone [email protected]:joyent/manta-marlin.git
...
$ cd manta-marlin
$ make deps
or just:
$ npm install git+https://github.com/joyent/manta-marlin.git
At this point, you'll have several tools available inside the "sbin" directory:
- mrerrors: fetch and summarize errors for a particular job or a given time period
- mrjob: create, show, summarize, and manipulate Marlin jobs by interacting with Moray directly (instead of the web tier)
- mrjobreport: summarize performance of a job's execution
By default, these tools assume they can contact the 1.moray shard at localhost:2020. For a production deployment, it's recommended to set up an ssh tunnel so that this works. "mrjob" also supports a MORAY_URL environment variable.
The standard Marlin development environment is a SmartOS zone deployed on a system with a single-system Manta deployed on it. This is usually COAL or an Emeryville lab machine. In principle, you can develop Marlin on a different system from the one where you deploy it, but the existing tools are optimized for the single-system case.
If you are using COAL for Marlin development, you may need to increase the ram and disk sizes for the COAL VM. For example, for 4GB RAM and 70GB disk, edit your usb-headnode/build.spec.local:
{
"build-tgz": "false",
"answer-file": "answers.json",
"coal-memsize": 8192,
"coal-zpool-disk-size": "70"
}
The usual procedure is:
-
Reflash. See sdc-headnode.git.
-
Deploy Manta. See manta-deployment.git.
-
Set up your dev zone. From the global zone, run:
/opt/smartdc/agents/lib/node_modules/cabase/tools/devsetup -t manta dap
Replace "dap" with whatever username you want to use. If
http://us-east.manta.joyent.com/$user/public/.ssh/id_rsa.pub exists, this script
will download that to $HOME/.ssh/authorized_keys
in the new zone.
Skip this section if you did the quick setup above.
The above script takes care of setting up /etc/resolv.conf in your development zone to refer to the Manta name servers and adds several environment variables to the new user's profile script:
$ export MAHI_URL=http://authcache.sf-1.joyent.us/
$ export MORAY_URL=tcp://1.moray.sf-1.joyent.us:2020/
$ export MANTA_URL=https://manta.sf-1.joyent.us
$ export MANTA_USER=poseidon
$ export MANTA_KEY_ID=ff:4a:ad:ac:92:86:ec:d2:5a:9d:88:c8:e8:12:8d:15
$ export MANTA_TLS_INSECURE=1
These are all specific to the install. Don't just copy the values here. To get the MANTA_KEY_ID for poseidon, use:
# sdc-ldap search -b "$(sdc-ldap search login=poseidon | \
head -1 | cut -c5-)" objectclass=sdckey | \
awk '/^fingerprint:/{ print $2 }'
ff:4a:ad:ac:92:86:ec:d2:5a:9d:88:c8:e8:12:8d:15
To run the test suite, you must also copy the ssh key of the "poseidon" user to to your dev zone. By default, the tests expect it in "$HOME/.ssh/id_rsa{,.pub}". However, you can also set the MANTA_KEY environment variable if it located somewhere else. You can get the poseidon key from the "ops" Manta zone. If your dev zone is on the same box as the ops zone, you can do this with:
# cp /zones/$(vmadm lookup alias=~ops)/root/root/.ssh/id_rsa* \
/zones/$YOUR_DEV_ZONE_UUID/root/$YOUR_LOGIN_NAME/.ssh
Also, if you find the tests are failing in COAL due to retries in Marlin you can set the MARLIN_TESTS_STRICT environment variable to false, like so:
$ export MARLIN_TESTS_STRICT=false
Finally, devsetup also adds privileges to your user. You can do this manually by running this as root in your zone:
$ usermod -K defaultpriv=basic,dtrace_user,dtrace_proc,contract_event,sys_mount,hyprlofs_control 'user'
Again, "devsetup" takes care of all this. This is all only necessary if you're in a non-standard configuration.
This part's easy:
$ git clone [email protected]:joyent/manta-marlin.git
$ cd manta-marlin
$ . dev/env.sh
$ make
...
$ make check
...
check ok
Whenever you start working on Marlin, use . dev/env.sh
to load PATH and other
environment variables to include both the Marlin tools (mrjob and the like) and
the Manta tools (mls, mput, etc.)
There are two main components in Marlin: the job supervisor (formerly called the "worker") and the agent.
The supervisor can be run directly out of the repo. The easiest way to get started is to copy the config.json file from an existing Marlin supervisor deployed on the same system, modify the port number (since the default is a privileged port), disable the supervisor whose configuration you copied, and then run:
$ cd manta-marlin
$ . dev/env.sh
$ make
$ node build/proto/root/opt/smartdc/marlin/lib/worker/server.js \
../config.json | tee ../supervisor.out | bunyan -o short
With this approach, you can apply changes by stopping the supervisor, running "make proto", and starting the supervisor again.
The agent must be run inside the global zone, since it uses other zones to
actually execute tasks. To run your own agent, you'll want to run tools/mru
from inside the global zone, as in
/zones/$your_dev_zone/root/home/$your_user/marlin/tools/mru
. This script will
reconfigure the marlin-agent SMF service running in the global zone to execute
the code in the proto area of your workspace
(/zones/$your_dev_zone/root/home/$your_user/marlin/build/proto/root
). This
way, you can make changes in your workspace, run "make proto", and simply
svcadm restart marlin-agent
to test them out.
The main development tool is "mrjob", which allows you to create, update, and fetch detailed status about Marlin jobs by talking to Moray directly.
To create a job, you must first create an auth token against Manta. The easiest way to do this is to sign the 'POST /$user/tokens' URL:
$ msign -m POST /dap/tokens
https://manta.joyent.us/dap/tokens?...
Then use the URL you were given with curl to create a new token:
$ curl -k -XPOST 'https://manta.joyent.us/dap/tokens?...'
{"token":"..."}
Copy this token, and pass that to "mrjob submit" using the "-t" option. You'll also need to set MAHI_URL appropriately.
$ jobid=$(mrjob submit -t '...' -m wc); echo $jobid
8abafb0d-9e9c-4023-b82b-73c4b52559f2
For details on using "mrjob" to submit more complex jobs, add keys to jobs, fetch detailed jobs, and so on, run "mrjob --help".
The Marlin test suite is not comprehensive, but it exercises a good portion of the stack. The test suite should always be totally clean. Tickets should be filed on any failures not due to local changes or misconfiguration.
The main way of running the test suite is inside your development zone, usually on the same physical system where you're testing a one-system Manta installation. This approach skips some tests that require verifying private state -- see "Running the full test suite" below for details.
This environment assumes you set up the environment variables and ssh keys as described above. To run the test suite, just run:
$ dev/tools/catest -a
You can run individual tests manually with just:
$ node proto/root/opt/smartdc/marlin/test/...
Many of them emit bunyan output, so you may want to pipe that to bunyan(1).
Running the test suite in a zone can only test the user-facing aspects of Marlin: behavior of the APIs, job inputs, job outputs, job errors, results, and so on. It's also important to test some implementation details, like that metering worked correctly. To do this, the Marlin test suite must be run from the global zone so it can access the agent directly. In this environment, catest will automatically configure itself based on the agent's configuration (i.e., you don't have to worry about the above environment variables), and it will automatically run the extra checks.
Note that this full test suite only works on single-system Manta installations, since it assumes all tasks will be executed by the local agent.
To run individual tests, you have to set up the environment with:
# cd /zones/$your_zone/root/home/$your_user/marlin
# . dev/env.sh
# . dev/tools/catest_init.sh
# catest_init_global
Then you can run individual tests just like inside a zone:
$ node test/...
In one shell, start the supervisor in a way that will be automatically restarted:
$ while :; do LOG_LEVEL=debug node \
build/proto/root/opt/smartdc/marlin/lib/worker/server.js \
../config.json | tee -a ../supervisor.out | bunyan -linfo; done
In another, start a loop that will kill the supervisor a random interval between 5 and 30 seconds apart. There's a script in /dev/tools/ for this:
$ dev/tools/test_sup_kill.sh
And finally, run the test suite in "stress" mode, which is a little laxer about transient failures:
$ for (( i = 0; ; i++ )) { \
echo "$i: $(date)"; \
if ! node build/proto/root/opt/smartdc/marlin/test/live/tst.main.js \
-S >> ../tests.out 2>&1; then
echo "FAILED at $(date)";
break;
fi
}
You can also test periodic agent crashes using a script to periodically kill the agent. The minimum timeout for this script is longer than the corresponding supervisor script because jobs fail if the agent fails too many times while they're running. You must run this from the global zone:
# /zones/$YOUR_ZONE/root/home/$YOUR_USER/marlin/dev/tools/test_agent_kill.sh
- Your code should be "make prepush" clean. That includes both "make check" and "make test" (the test suite).
- Assess the test impact of your changes -- make sure you've run all manual tests you can think of, and automated all of the ones that can reasonably be automated.
- Assess any impact on fresh install / deployment and make sure you've tested that if necessary.
- Assess any impact on upgrade, including flag days for developers or existing deployments (e.g., us-beta).
- Code review is strongly suggested.