In the current implementation, ACM supports multi-participant with same supported element Type but different participantId, so they need different properties file.
In order to support replicareplication, it needs we need to support multi-participant using same properties filereplicas of participants.
Note:
- In a scenario of high number of compositions, if participant is restarting it will be slow-down the restarting action: AC-runtime will send a message for each composition primed and instance deployed to the participant.
To avoid the restarting action, participant needs a database we need proper participant replication support; - In a scenario where a participant is stuck in deploying, the instance will be in TIMEOUT and the user can take action like deploy again or undeploy. In that scenario the intermediary-participant has to receive the next message, kill the thread that is stuck in deploying and create a new thread.
- In a scenario where we are increasing the number or participants, could be useful to have different topic name for source and sink. This solution will eliminate the number of useless messages in kafka.
Example:- for ACM-runtime:
- sink: POLICY-ACM-PARTICIPANT
- source: POLICY-ACM-RUNTIME
- for participant:
- sink: POLICY-ACM-RUNTIME
- source: POLICY-ACM-PARTICIPANT
- for ACM-runtime:
...
- Multiple participant replicas possible - it can deal with messages across many participants
- All participants should have same group-id in kafka
- All should have the same participant-id.
Solution 4: Distributed Cache
Issues:
- Not persistent - if the application that handles cache server restarts - data is lost.
- Approval issues - with Redis, Etcd, Search Engine.
Solution 5: True Participant Replicas
The requirements are:
- Participants can be replicated, each participant can have an arbitrary number of replicas
- Composition definitions, instances, element instances and all their data including properties is identical in all participant replicas
- When anything is changed in one replica, the change is propagated to all the replicas of a participant
- An operation on a composition element can be sent to any replica of a participant, which means that for a given element, the deploy could be on replica 1, the update could be on replica 2 and the delete could be on replica 3, as one would expect in any HA solution
- A single REST operation called on ACM-R will select a participant replica (probably using round robin initially but we could support other algorithms in the future), and use that replica for that operation.
- The ACM runtime will be made HA (more than one replica of ACM-R will be supported), it will run on a HA postgres.
- The implementation of change propagation is transparent to participant API users
- Replicas are "eventually consistent", with consistency typically occurring in 100s of milliseconds
This solution uses a similar approach to that used by Kubernetes when using etcd to implement CRDs and CRs. It implements replication in participants by introducing
- The concept of participant replicas into ACM-R and the Participant Intermediary
- A lightweight mechanism for replica update between replicas and ACM-R. Every time a replica changes its data, the change is reported to ACM-R and ACM-R updates all other replicas for that participant
The diagram above depicts data management as implemented today in ACM-R. ACM-R maintains the "source of truth" for all compositions and their elements in the ACM-R database.
- Composition type data is pushed from ACM-R to the participants and is read-only in the participants
- Composition element data (state information mainly) is built up by interactinos between ACM-R and participants for the various ACM operations (Deploy/Update/Undeploy etc) and ACM-R always maintains the current composition element data in the ACM-R database
- Composition element properties are pushed by ACM-R to the participant and can be updated in the participant by the participant implementation. When the properties are updated on the participant side, those changes are propagated to the ACM-R database
Therefore today, for the three types of data above, the ACM-R database has the state of the participant. This means that creation of participant replicas is rather trivial. We leave the current data handling mechanism in place, introduce the concept of participant replicas in ACM-R, and introduce data synchronization across participant replicas.
In this solution:
- Participant Design backward compatibility is preserved, there is no change to the participant intermediary interface for participant implementations
- Participant Configuration Backward compatibility is preserved, apart from a new "replicas" parameter (optional, default is 1), there is no change to participant configuration
- ACM-R introduces a new REST API for replica management
- ACM-R is made HA so that it itself can scale
Optimal Solution:
After analysis, it is clear that the best solution to use is number 3.
- Arbitrary number of participants possible
- DB migrator upgrades older versions
- Restart scenario not applicable anymore. Could be removed.
- Approval not an issue - postgres already used by acm.
- DB will be created automatically - as are required tables.
Older participant versions support (Regression)
...
Disadvantages of DB use
- The Participant as a concept envisages an extremely lightweight mechanism, which can be deployed as a very lightweight plugin into the runtime entity implementing a participant
- Introducing a database into the participant means that participant becomes a heavyweight plugin with cross participant and participant/acm-r configuration required
- There are multiple sources of truth for participant and element data, the ACM-R DB and the this new participant DB have duplicated data
- Participant configuration is no longer transparent to the participant implementing component, having a DB for each participant forces orthogonal configuration across participant implementing components
- Configuration of components implementing participants becomes much more complex, the components must manage the participant replication
- The replication mechanism enforces the use of an additional 3pp, the database
Solution 4: Distributed Cache
Issues:
- Not persistent - if the application that handles cache server restarts - data is lost.
- Approval issues - with Redis, Etcd, Search Engine.
Solution 5: True Participant Replicas
The requirements are:
- Participants can be replicated, each participant can have an arbitrary number of replicas
- Composition definitions, instances, element instances and all their data including properties is identical in all participant replicas
- When anything is changed in one replica, the change is propagated to all the replicas of a participant
- An operation on a composition element can be sent to any replica of a participant, which means that for a given element, the deploy could be on replica 1, the update could be on replica 2 and the delete could be on replica 3, as one would expect in any HA solution
- A single REST operation called on ACM-R will select a participant replica (probably using round robin initially but we could support other algorithms in the future), and use that replica for that operation.
- The ACM runtime will be made HA (more than one replica of ACM-R will be supported), it will run on a HA postgres.
- The replication mechanism used between ACM-R and participants is transparent to participant API users
- Replicas are "eventually consistent", with consistency typically occurring in 100s of milliseconds
This solution uses a similar approach to that used by Kubernetes when using etcd to implement CRDs and CRs. It implements replication in participants by introducing
- The concept of participant replicas into ACM-R and the Participant Intermediary
- A lightweight mechanism for replica update between replicas and ACM-R. Every time a replica changes its data, the change is reported to ACM-R and ACM-R updates all other replicas for that participant
The diagram above depicts data management as implemented today in ACM-R. ACM-R maintains the "source of truth" for all compositions and their elements in the ACM-R database.
- Composition type data is pushed from ACM-R to the participants and is read-only in the participants
- Composition element data (state information mainly) is built up by interactions between ACM-R and participants for the various ACM operations (Deploy/Update/Undeploy etc) and ACM-R always maintains the current composition element data in the ACM-R database
- Composition element properties are pushed by ACM-R to the participant and can be updated in the participant by the participant implementation. When the properties are updated on the participant side, those changes are propagated to the ACM-R database
Therefore today, for the three types of data above, the ACM-R database has the state of the participant. This means that creation of participant replicas is rather trivial. We leave the current data handling mechanism in place, introduce the concept of participant replicas in ACM-R, and introduce data synchronization across participant replicas.
We introduce a participant replication table in the ACM-R database, which is an index used to record the replicas that exist of each participant. When the replica of a component implementing a participant is created (by Kubernetes or otherwise), the participant intermediary in the component registers with ACM-R as usual and as it does today. The only difference is that the participant intermediary will send the participant ID and the replica number. ACM-R is updated to accept registrations from participants with the same Participant ID and different replica numbers. When the first replica for a certain participant registers, ACM-R will handle this registration exactly as it does today and will add the replica as the single replica that exists for this participant. When the next replica registers, ACM-R will recongnise that this is a second replica for a participant that already exists and will record this as a replica. Rather than priming this replica, ACM-R will copy all the data from the first replica to this replica. The registration of further replicas will continue to follow this pattern.
During normal operation where ACM-R receives and executes requests towards participants, ACM-R will use Participant Load Balancing to select a replica using a round-robin algorithm and execute the operation on that replica. When the operation finishes, ACM-R will synchronize the data from the replica that executed the operation to all the other replicas using Participant Synchronization.
If ACM-R is informed by a replica that an Implementing Component changed composition element properties, Participant Synchronization synchronizes these changes to all other Participant Intermediary replicas.
In this solution:
- We will preserve Participant Design backward compatibility, there is no change to the participant intermediary interface for participant implementations
- Participant version backward compatibility will not be preserved because we need to pass replica information in the registration and operational messages, all participants will have to be upgraded to the new version.
- The REST API that returns participant information will be updated to include replica information
- ACM-R is made HA so that it itself can scale
- We can use Kafka load balancing on the participants and get the load balancing functionality for nothing
- A new Kafka topic is used for synchronization
Optimal Solution:
After analysis, it is clear that the best solution to use is number 5.