rajesh created HDDS-17842:
-----------------------------
Summary: Refactor SCMStateMachine: extract request dispatch,
readiness and DN server startup out of the Ratis state machine
Key: HDDS-17842
URL: https://issues.apache.org/jira/browse/HDDS-17842
Project: Apache Ozone
Issue Type: Task
Reporter: rajesh
Assignee: rajesh
{{SCMStateMachine}} mixes several responsibilities, and callers depend on the
concrete class instead of an abstraction.
_Problems_
# Request dispatch is registered through the concrete class.
** {{SCMStateMachine}} owns a {{Map<RequestType, ScmInvoker<?>>}} filled
through {{{}registerInvoker(...){}}}.
** {{SCMRatisServerImpl.registerStateMachineHandler}} has to hold a concrete
{{SCMStateMachine}} just to push handlers into it.
** Dispatch is a separate concern from applying Ratis transactions, and
nothing names it.
# Callers depend on the class and not on {{{}StateMachine{}}}.
** {{SCMRatisServer.getSCMStateMachine()}} returns the class, and 3 production
callers and about 10 test callers use it.
** They need {{getIsStateMachineReady()}} or {{{}stopDNServerStartRetry(){}}},
which exist only on the class.
** The state machine cannot be wrapped (for example by a tracing proxy) or
substituted without breaking those callers.
# Readiness is a flag with {{if}} checks.
** {{getIsStateMachineReady()}} and {{isStateMachineReady}} are checked in
many places in {{SCMStateMachine}} and polled by {{{}StateMachineReadyRule{}}}.
** The behaviour differs before and after readiness, which suggests two states
(for example a safe-mode state machine and a full-mode state machine) and not
conditionals.
# Datanode server startup lives in the state machine.
** {{{}tryStartDNServerAndRefreshSafeMode{}}}, the retry executor and
{{stopDNServerStartRetry()}} are datanode-protocol-server lifecycle code.
** It is called from Ratis callbacks, and {{StorageContainerManager.stop}} has
to reach into the state machine to stop it.
_Proposal (to be refined)_
* Introduce a request registry abstraction, for example {{ScmRequestRegistry}}
with {{register}} and lookup, plus an {{{}Impl{}}}. The Ratis server creates it
and passes it to the state machine through the constructor.
* Make callers depend on a narrow interface that extends Ratis
{{{}StateMachine{}}}.
* Move datanode server start and retry out of {{SCMStateMachine}} into its own
component, with {{StorageContainerManager}} owning its lifecycle.
* Investigate modelling readiness as separate state-machine behaviour and
removing the readiness checks.
_Benefits_
* Wrapping or decorating the state machine, for example the tracing proxy via
{{{}TracingUtil.createInternalProxy{}}}, works without a raw-instance
workaround or {{@SkipTracing}} hacks.
* Dispatch and readiness can be unit-tested without a full
{{{}SCMStateMachine{}}}.
* The state machine shrinks to Ratis concerns, and SCM shutdown ordering
becomes explicit.
* Less test churn, because callers stop depending on the concrete class.
_Out of scope:_ any behaviour change, and any change to the Ratis log or
protocol.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]