About Hydromon

Hydromon allows you to check the liveliness of a server before calling it (proactive monitoring). This is done by letting a small helper component, the hydromon server, constantly ping the server object. The idea is to install the Hydromon server on each node (computer) that participates in the cluster. A Hydro client can then safely assume that the Hydromon server on the same node is up, and knows more about the liveliness of the remote service to call.

Hydromon does not completely avoid that calls "hang" when a remote service dies. However, it limits the number of such calls. After a few seconds, the Hydromon server has found out that the remote service is down, and signals all clients that it is senseless to call this service. Nevertheless, requests may hang that have been sent in the time period until Hydromon knows about the failure. Because of this, it is strongly recommended to set a timeout for each call.

Running the Hydromon server

The Hydromon server is a netplex component. Its API is defined in Hydromon_netplex.

The Hydromon server needs a POSIX shared memory segment. If the OS does not support this feature, Hydromon will not work.

Configuring proxies for Hydromon

The module to use is Hydromon_query:

First create a Hydromon cache:

let cache = Hydromon_query.create_cache ~port ~period ()

In port pass the TCP port on the local computer where the Hydromon server is listening. The period is the number of seconds the cache entries remain valid - this can usually be something like 3600 (an hour). The meaning is that an entry in the shared memory segment is allocated for this period of time.

After that, simply configure the proxies so the cache is used:

Hydromon_query.configure_proxy ~cache ~proxy ~operation ~idempotent()

The operation is the name of the operation to call. If nothing better is defined, use "ice_ping" - it is available in all objects. You can also use a different operation if it is more meaningful, but it must neither take arguments, not pass a result back ("void"). In idempotent you have to specify whether this operation is idempotent or not. "ice_ping" is, but your own operation needs not to be.

Cost of Hydromon

The check whether a service is alive is very cheap. Every period seconds an RPC call is required for administrative purposes, but during that period the check is practically cost-free (it is a shared memory lookup).