bitflicker64 commented on PR #3477:
URL: https://github.com/apache/hugegraph/pull/3477#issuecomment-6873627813

   Blocking: no. Score: 8.5/10. Summary: I went through this head again as its 
author and found no new defect. This is a comment, not an approval, since I 
can't approve my own PR. Evidence: code reading at 7df011acb, a focused unit 
run, and the CI logs of apache/hugegraph#3477, which has the same head.
   
   **What I checked**
   
   - `CachedSchemaTransactionV2`: every cache insertion and every place the 
generation advances, against a concurrent clear, local write, removal and 
write-through. No PD call is made while the read-write lock is held, and I 
found no lock-order cycle between the reload lock, the read-write lock and the 
listener map.
   - Publish and retry: every schema write goes through `addSchema`, 
`updateSchema`, `removeSchema` or `clear`. A graph that fails to publish again 
during a retry stays pending, and a retry that was overtaken does not schedule 
another one.
   - `KvClient`: the PD server adds the watcher before it sends the first frame 
(`KvServiceGrpcImpl.clientWatch`), and a follower answers a watch with an 
error. So the started callback runs only for a live watch, with one exception: 
a leader change that lands between the leader check in `watch` and the 
registration leaves a watcher on the old leader that got its first frame and 
hears nothing after it. `KvServiceGrpcImpl` is not touched here and behaves the 
same on master. That case falls under the watch that dies without an error in 
the limits.
   - Compatibility: the clear event value is unchanged, there is no new 
dependency and no `.proto` change. Store nodes listen on the same key 
(`SchemaDriver`) and clear that graph's cache, which refills from PD on the 
next read. That happens more often now, since every update and every graph open 
publishes. I did not measure what it costs a Store.
   - The description: the size table and the 35, 7 and 5 test counts match the 
code.
   
   **Tests**
   
   - Local, JDK 17: `mvn test -pl hugegraph-server/hugegraph-test -am -P 
unit-test -Dtest='CachedSchemaTransactionV2Test,CachedSchemaTransactionTest'`: 
35 run, 0 failures and 21 run, 0 failures.
   - CI logs of apache/hugegraph#3477 at the same SHA: `MultiClusterSuiteTest` 
11 run, 0 failures, with all three PD nodes restarted; `PDClientSuiteTest` 122 
run, 0 failures.
   
   **Still open from earlier threads** (0.5 each in the score, none blocks)
   
   - A watch that ends on a non-retryable gRPC status is not started again. 
Same on master, listed in `docs/server-runtime.md`, tracked in 
apache/hugegraph#3478.
   - The cluster-test HTTP client has no timeouts, so a request that never 
returns outlasts `waitFor`.
   - `resetPendingClearEvents` cannot stop a retry that is already running.
   
   **Not covered**
   
   - This repository ran no build or test check at this head. The CI evidence 
is from the apache PR.
   - I read the status of the Server unit and HStore core jobs there, not their 
logs, and did not run the multi-cluster or HStore suites myself.
   - The lock and `KvClient` points come from reading the code, not from a test 
that forces those orderings.
   - A real network partition, a silent stream death and Servers on mixed 
versions, as the description already says.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to