Enhance clarifications for ServiceDiscoverer contract
#3002
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Motivation: Some assumptions that are currently in place are not clarified in documentation for new implementations. Modifications: - Clarify that the returned publisher should never complete and must support re-subscribes to handle failures and cancellations. - Clarify motivation for always starting with the "state of the world" for interpreting `ServiceDiscovererEvent`(s). - Add recommendation to implement a `ServiceDiscoverer` using a single thread. Result: More details for expected behavior of a `ServiceDiscoverer` implementation.
idelpivnitskiy
added a commit
to idelpivnitskiy/servicetalk
that referenced
this pull request
Jul 12, 2024
Motivation: Currently, both implementations of the `LoadBalancer` assume 2 possibilities for the first events collection after re-subscribe: "state of the world" or "delta" and tries to infer which one it is based on what event statuses are present in the collection. However, it doesn't provide any guarantees because even if SD impl keeps returning deltas after a re-subscribe, a delta can contain only `AVAILABLE` events, but it will be incorrectly interpreted as "state of the world". Because of Reactive Streams rule clarified in apple#3002, `LoadBalancer` implementations should always expect "state of the world" for the first events after re-subscribe. Modifications: - Remove inference of the first collection of events after resubscribe and always expect a new subscriber to start from "state of the world". - Update `resubscribeToEventsWhenAllHostsAreUnhealthy()` test to reflect behavior change. Result: Both `LoadBalancer` implementations follow `ServiceDiscoverer` contract and expect the first collection of events after re-subscribe to represent a "state of the world".
idelpivnitskiy
added a commit
to idelpivnitskiy/servicetalk
that referenced
this pull request
Jul 12, 2024
Motivation: Clients have a configurable `serviceDiscovererRetryStrategy` to guarantee a steady stream of events to the `LoadBalancer` that never fails. It's necessary at the client level to avoid hanging requests indefinitely and let requests observe failures from ServiceDiscoverer. Also, for `PartitionedHttpClient` it's necessary to guarantee that `GroupedPublisher` never fails. Retry is effectively a re-subscribe. According to `ServiceDiscoverer` contract (clarified in apple#3002), each `Subscriber` receives a "state of the world" as the first collection of events. The problem is that the state may change significantly between retries, as a result unavailable addresses can remain inside the `LoadBalancer` forever. Example: T1. SD delivers [a,b] T1. LB receives [a,b] T1. SD delivers error T2. SD info changed ("a" got revoked) T3. Client retries SD T3. SD delivers [b] T3. LB receives [b] (but still holds "a") When we retry `ServiceDiscoverer` errors, we should keep pushing deltas downstream or purge events that are not present in the new "state of the world". We previously had this protection but it was mistakenly removed in apple#1949 as part of a broader refactoring around `ServiceDiscoverer` <-> `LoadBalancer` contract. Modifications: - Add `RetryingServiceDiscoverer` that handles retries and keeps the state between retries. - Use it in `DefaultSingleAddressHttpClientBuilder` and `DefaultPartitionedHttpClientBuilder`. - Use `CastedServiceDiscoverer` to allow modifications for `ServiceDiscovererEvent` after we started to use a wildcard type in apple#2379. - Pass consistent `targetResource` identifier to both `RetryingServiceDiscoverer` and `LoadBalancerFactory` to allow state correlation when inspecting heap dump. Result: Client keeps pushing deltas to `LoadBalancer` after retrying `ServiceDiscoverer` errors, keeping its state consistent with `ServiceDiscoverer`.
bryce-anderson
approved these changes
Jul 12, 2024
servicetalk-client-api/src/main/java/io/servicetalk/client/api/ServiceDiscoverer.java
Show resolved
Hide resolved
Comment on lines
30
to
32
| * Because typically a {@link ServiceDiscoverer} implementation runs in the background and doesn't require many compute | ||
| * resources, it's recommended (but not required) to run it and deliver updates on a single thread either for all | ||
| * discoveries or at least for all {@link Subscriber Subscribers} to the same {@link Publisher}. |
There was a problem hiding this comment.
fwiw, it's not clear what this buys the implementors.
There was a problem hiding this comment.
I don't want to put it as a requirement, it's just a nice-to-have recommendation bcz most ppl don't think about it when they implement.
For RRLB it will help with this case: #3005 (comment)
But we are moving away from RRLB anyway.
In general, it makes debugging easier bcz you can track all events from a single thread and logs will represent the exact serial behavior. If they use multiple threads, logs may mess up order of events when LB cancels and resubscribes.
Do you have a recommendation how to rephrase?
There was a problem hiding this comment.
My best suggestion is basically what you said in the reply. 🙂 I added as a suggestion but I'm also not strongly opinionated about it. If you prefer to keep it short and not try to sell it then that's perfectly fine with me.
servicetalk-client-api/src/main/java/io/servicetalk/client/api/ServiceDiscoverer.java
Outdated
Show resolved
Hide resolved
bryce-anderson
approved these changes
Jul 13, 2024
idelpivnitskiy
added a commit
that referenced
this pull request
Jul 15, 2024
…#3004) Motivation: Currently, both implementations of the `LoadBalancer` assume 2 possibilities for the first events collection after re-subscribe: "state of the world" or "delta" and try to infer which one it is based on what event statuses are present in the collection. However, it doesn't provide any guarantees because even if SD impl keeps returning deltas after a re-subscribe, a delta can contain only `AVAILABLE` events, but it will be incorrectly interpreted as "state of the world". Because of Reactive Streams rule clarified in #3002, `LoadBalancer` implementations should always expect "state of the world" for the first events after re-subscribe. Modifications: - Remove inference of the first collection of events after re-subscribe and always expect a new subscriber to start from "state of the world". - Update `resubscribeToEventsWhenAllHostsAreUnhealthy()` test to reflect behavior change. Result: Both `LoadBalancer` implementations follow `ServiceDiscoverer` contract and expect the first collection of events after re-subscribe to represent a "state of the world".
tkountis
approved these changes
Jul 17, 2024
idelpivnitskiy
added a commit
that referenced
this pull request
Jul 18, 2024
…3006) Motivation: Clients have a configurable `serviceDiscovererRetryStrategy` to guarantee a steady stream of events to the `LoadBalancer` that never fails. It's necessary at the client level to avoid hanging requests indefinitely and let requests observe failures from ServiceDiscoverer. Also, for `PartitionedHttpClient` it's necessary to guarantee that `GroupedPublisher` never fails. Retry is effectively a re-subscribe. According to `ServiceDiscoverer` contract (clarified in #3002), each `Subscriber` receives a "state of the world" as the first collection of events. The problem is that the state may change significantly between retries. As a result, unavailable addresses can remain inside the `LoadBalancer` forever. Example: T1. SD delivers [a,b] T1. LB receives [a,b] T1. SD delivers error T2. SD info changed ("a" got revoked) T3. Client retries SD T3. SD delivers [b] T3. LB receives [b] (but still holds "a") When we retry `ServiceDiscoverer` errors, we should keep pushing deltas downstream or purge events that are not present in the new "state of the world". We previously had this protection but it was mistakenly removed in #1949 as part of a broader refactoring around `ServiceDiscoverer` <-> `LoadBalancer` contract. Modifications: - Add `RetryingServiceDiscoverer` that handles retries and keeps the state between retries. - Use it in `DefaultSingleAddressHttpClientBuilder` and `DefaultPartitionedHttpClientBuilder`. - Use `CastedServiceDiscoverer` to allow modifications for `ServiceDiscovererEvent` after we started to use a wildcard type in #2379. - Pass consistent `targetResource` identifier to both `RetryingServiceDiscoverer` and `LoadBalancerFactory` to allow state correlation when inspecting heap dump. Result: Client keeps pushing deltas to `LoadBalancer` after retrying `ServiceDiscoverer` errors, keeping its state consistent with `ServiceDiscoverer`.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation:
Some assumptions that are currently in place are not clarified in documentation for new implementations.
Modifications:
ServiceDiscovererEvent(s).ServiceDiscovererusing a single thread.Result:
More details for expected behavior of a
ServiceDiscovererimplementation.