Plasma GitLab Archive
Projects Blog Knowledge

Module Rpc_proxy.ReliabilityCache

module ReliabilityCache: sig .. end


The reliability cache stores information about the availability of remote servers. The managed clients put information about recent failures into the cache.

It is advantegeous to have only one cache per process, because this maximizes the usefulness. The cache is thread-safe.

A server endpoint is disabled when too many errors occur in sequence. For a disabled endpoint the functions host_is_enabled and/or sockaddr_is_enabled return false. The endpoint is automatically enabled again after some timeout; this is initially disable_timeout_min, but is increased exponentially until disable_timeout_max when further errors occur.

Independently of this machinery the functions host_is_enabled and sockaddr_is_enabled may also return false when an external availability checker says that the endpoint is down. This information is not entered into the cache, and will also not trigger the disable timeout. Instead, the hook function getting the availability will be simply called again.

type rcache 
The cache
type rcache_policy = [ `Any_failing_port_disables_host
| `Failing_port_disables_host of int
| `Independent
| `None ]
How failures of individual ports are interpreted:
  • `Independent: When a connection to a remote port repeatedly fails, only this port is disabled
  • `Failing_port_disables_host p: When a connection to the TCP port p repeatedly fails, the whole remote host is disabled. Other ports do not disable the host, but are treated as in `Independent.
  • `Any_failing_port_disables_host: When a connection to any TCP port repeatedly fails, the whole remote host is disabled
  • `None: Nothing is disabled
Note that the rcache_availability hook is not affected by the policy; this hook is called anyway. The policy only determines how the internal error counter is interpreted.
type rcache_config = {
   rcache_policy : rcache_policy; (*The policy, see above*)
   rcache_disable_timeout_min : float; (*For how long ports and hosts are disabled*)
   rcache_disable_timeout_max : float; (*For how long ports and hosts are disabled at most*)
   rcache_threshold : int; (*How many errors are required for disabling a port*)
   rcache_availability : rcache -> Unix.sockaddr -> bool; (*External availability checker. Called by sockaddr_is_enabled before the result is calculated*)
}
val create_rcache_config : ?policy:rcache_policy ->
?disable_timeout_min:float ->
?disable_timeout_max:float ->
?threshold:int ->
?availability:(rcache -> Unix.sockaddr -> bool) ->
unit -> rcache_config
Create a config record. The optional arguments set the config components with the same name. The arguments default to:
  • policy = `None
  • disable_timeout_min = 1.0
  • disable_timeout_max = 64.0
  • threshold = 1
  • availability = fun _ _ -> true

val create_rcache : rcache_config -> rcache
Creates a new cache object. The same cache can be used by several managed clients, even by totally unrelated ones
val rcache_config : rcache -> rcache_config
Return the config
val global_rcache_config : unit -> rcache_config
Returns the global config:
  • policy = `None
  • disable_timeout_min = 1.0
  • disable_timeout_max = 64.0
  • threshold = 1
  • availability = fun _ _ -> true

val set_global_rcache_config : rcache_config -> unit
Sets the new global config. This is only possible as long as neither default_global_config nor global_rcache have been called.
val global_rcache : unit -> rcache
The global cache. Initially, this cache has the default config. It is possible to change the default config before using the global cache for the first time.
val derive_rcache : rcache ->
rcache_config -> rcache
derive_cache parent config: Returns a new cache that shares the error counters with parent. The interpretation of the counters, however, may be differently configured in config.

Because it is advantageous to share the error information as much as possible, the recommended way to create a new cache object is to derive it from the global cache.

What derive_rcache actually does (and this is not yet optimal): Any incr and reset of an error counter is also forwarded to the parent cache. The tests whether hosts and ports are enabled do an AND of the results for the cache and its parent (i.e. both must be ok to enable). This allows some information sharing, but only in vertical direction.

val incr_rcache_error_counter : rcache -> Unix.sockaddr -> unit
Increase the error counter for this sockaddr. If the threshold is reached and there is a disable policy, the sockaddr will be disabled.

This function is to be called after an RPC call times out, or runs into a socket error.

val reset_rcache_error_counter : rcache -> Unix.sockaddr -> unit
Reset the error counter for this sockaddr. If disabled, the sockaddr is set to enabled again.

This function is to be called when an RPC call is successful.

val sockaddr_is_enabled : rcache -> Unix.sockaddr -> bool
Returns whether the sockaddr is enabled. This also calls the rcache_availability hook.
val host_is_enabled : rcache -> Unix.inet_addr -> bool
Returns whether the host is enabled
This web site is published by Informatikbüro Gerd Stolpmann
Powered by Caml