Writing SimpleRPC Agents
Simple RPC works because it makes a lot of assumptions about how you write agents, we’ll try to capture those assumptions here and show you how to apply them to our Echo agent.
Conventions regarding Incoming Data
As you’ve seen in SimpleRPCClients our clients will send requests like:
A more complex example might be:
Effectively this creates a hash with the members
:sender and invokes the
You cannot use the a data item called
:process_results as this
has special meaning to the agent and client. This will indicate
to the agent that the client isn’t going to be waiting to process
results. You might choose not to send back a reply based on this.
Here’s our sample Echo agent:
You’ll notice it comes in two parts, the agent definition in the DDL file, and the implementation in the ruby class.
The agent name is derived from the class name, the example code creates the ruby class MCollective::Agent::Echo, the agent name for this would be echo.
Help and the Data Description Language
We have a file that goes together with an agent implementation which
is used to describe the agent in detail. This is referred to as the
$libdir/mcollective/agent/echo.ddl in the previous
The DDL file can be used to generate help texts, and to declare the validation that arguments must pass. This information helps us in making more robust clients and can be used to auto generate user interfaces.
We’ll not say much more about the DDL here, for more details read the DDL reference.
Actions are the individual tasks that your agent can do, we declare them in the DDL file, and put their implementation in the ruby class named for the agent (ie MCollective::Agent::Echo for the echo agent).
Actions with their inputs and outputs are declared in the DDL for the agent. From our echo example here’s the declaration of the echo action again:
The implementation of an action is added to the agent class:
Accessing the Input
As you see from the echo example our input is easy to get to by just looking in request, this would be a Hash of exactly what was sent in by the client in the original request.
The request object is in instance of MCollective::RPC::Request, you can also gain access to the following:
|time||The time the message was sent|
|action||The action it is directed at|
|data||The actual hash of data|
|sender||The id of the sender|
|agent||Which agent it was directed at|
Since data is the actual Hash you can gain access to your input like:
.data will give you full access to all the normal
Hash methods, eg
request will only give you access to include?.
Running Shell Commands
A helper function exist that makes it easier to run shell commands and gain
access to their
We recommend everyone use this method for calling to shell commands as it forces
C as well as wait on all the children and avoids zombies, you can
set unique working directories and shell environments that would be impossible
system that is provided with Ruby.
The simplest case is just to run a command and send output back to the client:
Here you will have set
on the output from the command.
You can append the output of the command to any string:
Here the STDOUT of the command will be saved in the variable
out and not sent
back to the caller. The only caveat is that the variables
<< method, so if you supplied an array each line of output will be a
single member of the array. In the example
out would be an array of lines
err would just be a big multi line string.
By default any trailing new lines will be included in the output and error:
You can shorten this to:
This will remove a trailing new line from the
If you wanted this command to run from the
Or if you wanted to include a shell Environment variable:
The status returned will be the exit code from the program you ran, if the program completely failed to run in the case where the file doesn’t exist, resources were not available etc the exit code will be -1
You have to set the cwd and environment through these options, do not simply
chdir or adjust the
ENV hash in an agent as that will not be safe in
the context of a multi threaded Ruby application.
Actions in external scripts
Actions can also be implemented using other programming languages as long as they support JSON.
/some/external/script will be called with 2 arguments:
- The path to a file with the request in JSON format
- The path to a file where you should write your response as a JSON hash
You can also access these 2 file paths in the
MCOLLECTIVE_REQUEST_FILE environment variables
Simply write your reply as a JSON hash into the reply file.
The exit code of your script should correspond to the ones in ResultsandExceptions. Any text in STDERR will be
logged on the server at
error level and used in the text for the fail text.
Any text to STDOUT will be logged on the server at level
These scripts can be placed in a standard location:
This will search each configured libdir for
agent/$agent_name/script.py, and will use the former if found. If you
specified a full path it will not try to find the file in libdirs.
The reply data is in the reply variable and is an instance of MCollective::RPC::Reply.
As pointed out in the ResultsandExceptions page results all include status messages and the reply object has a helper to create those.
The number in
reply.fail corresponds to the codes in ResultsandExceptions it would default to
1 so you could just say:
This is hypothetical action that is supposed to remove a message from some queue, if we do have a String as input that matches our message id’s we then check that we do have such a message and if we don’t we fail with a helpful message.
Technically this will just set
statusmsg fields in the reply to appropriate values.
It won’t actually raise exceptions or exit your action though you should do that yourself as in the example here.
There is also a
fail! instead of just
fail it does the same basic function but also raises exceptions. This lets you abort processing of the agent immediately without performing your own checks on
statuscode as above later on.
Sharing code between agents
Sometimes you have code that is needed by multiple agents or shared between the agent and client. MCollective has
name space called
MCollective::Util for this kind of code and the packagers and so forth supports it.
Create a class with your shared code given a name like
MCollective::Util::Yourco and save this file in the libdir in
A sample class can be seen here:
You can now use it in your agent or clients by first loading it from the MCollective lib directories:
You can write a fine grained Authorization system to control access to actions and agents, please see SimpleRPCAuthorization for full details.
The actions that agents perform can be Audited by code you provide, potentially creating a centralized audit log of all actions. See SimpleRPCAuditing for full details.
You can write to the server log file using the normal logger class:
You can log at levels
As of version 2.2.0 there is a system wide Cache you can use to store data that might be costly to create on each request.
The Cache is thread safe and can be used even with multiple concurrent requests for the same agent.
Imagine your agent interacts with a customer database on the node that is slow to read data from but this data does not change often. Using the cache you can arrange for this be read only every 10 minutes:
Here we setup a new cache table called
:customer if it does not already exist, the cache has a 10 minute validity.
We then try to read a cached customer record for
request[:customerid] and if it’s not been put in the cache
before or if it expired I create a new customer record using a method called
get_customer and then save it
into the cache.
If you have critical code in an agent that is not reentrant you can use the Mutex from the same cache to synchronize the code:
Here we are using the same Cache that was previously setup and just gaining access to the Mutex protecting the cache data. The code inside the synchronize block will only be run once so you won’t get competing updates to your customer data.
If the lock is held too long by anyone the mcollectived will kill the threads in line with the Agent timeout.
We provide a few hooks into the processing of a message, you’ve already used this earlier to set metadata.
You’d use these hooks to add some functionality into the processing chain of agents, maybe you want to add extra logging for audit purposes of the raw incoming message and replies, these hooks will let you do that.
|Hook Function Name||Description|
||Called at the end of the initialize method of the
||Before processing of a message starts, pass in the raw message and the
||Just before the message is dispatched to the client|
Called at the end of the
RPC::Agent standard initialize method. Use this to adjust meta parameters, timeouts and any setup you need to do.
This will not be called right when the daemon starts up, we use lazy loading and initialization so it will only be called the first time a request for this agent arrives.
Called just after a message was received from the middleware before it gets passed to the handlers.
reply will already be set, the msg passed is the message as received from the normal MCollective runner and the connection is the actual connector.
You can in theory send off new messages over the connector maybe for auditing or something, probably limited use case in simple agents.
Called at the end of processing just before the response gets sent to the middleware.
This gets run outside of the main exception handling block of the agent so you should handle any exceptions you could raise yourself. The reason it is outside of the block is so you’ll have access to even status codes set by the exception handlers. If you do raise an exception it will just be passed onto the runner and processing will fail.
You can save configuration for your agents in the main server config file:
In your code you can retrieve the config setting like this:
This will set the setting to whatever is in the config file or “” if unset.
In the past you had to copy an agent only to machines that they should be running on as all agents were activated regardless of dependencies.
To make deployment simpler agents support the ability to determine if they should run
on a particular platform. By default SimpleRPC agents can be configured to activate
or not with the
activate_agents and plugin specific
You can also place the following in
This is a simple way to enable or disable an agent on your machine, agents can also declare their own logic that will get called each time an agent gets loaded from disk.
If this block returns false or raises an exception then the agent will not be active on this machine and it will not be discovered.
When the agent gets loaded it will test if
/usr/bin/puppet exist and only if it does
will this agent be enabled.