SEZ9 commented on PR #13243:
URL: https://github.com/apache/seatunnel/pull/13243#issuecomment-6936979375

   @SEPURI-SAI-KRISHNA thanks for the detailed write-up of `276d0d01d0`. 
@nzw921rx +1 to your plan: since this involves behavior changes for all 
connectors, reaching consensus by email after the 3.1.0 release is the right 
order, and there is no rush to merge.
   
   These are the points from the previous review to carry into that discussion. 
@SEPURI-SAI-KRISHNA, for each one, could you say whether it is already handled 
or should wait for the email outcome?
   
   **Design (best settled in the email thread)**
   - **Restore compatibility:** you chose a clear rejection over a migration. 
That leaves stateful non-Jdbc jobs that ran with `multi_table_sink_replica > 1` 
with no checkpoint they can restore after upgrade. The check also rejects 
restores at the default of 1 and after the option is raised.
   - **Check ordering:** the replica check runs only when the writer is 
created, which is after the sink's save mode has run. A rejected job can 
therefore drop or recreate its target tables before it fails.
   
   **Code**
   - **Row-shape check:** the check in `selectQueueIndex` throws straight out 
of `write()`, so it bypasses `multi_table_failure_policy`. It also fails on 
zero-arity rows that the queue worker and `JdbcSinkWriter` skip.
   - **`KEYED` routing:** it hashes the raw key with `key.hashCode()`, so an 
array-typed key such as a `byte[]` primary key sends equal keys to different 
replicas.
   
   **Docs**
   - **Other sink pages:** only the Iceberg and Paimon option rows say "keep it 
at 1". Other non-Jdbc sink pages and the schema-evolution guide still present a 
value above 1 as a working tuning knob.
   - **When the check fires:** the option page, the upgrade note and the error 
message say it fires only when the option is reduced or parallelism changes. 
Raising the option also moves every subtask above 0 to a different index range.
   - **Default of 1:** the docs say such a job is unaffected and that "nothing 
is checked", but the check in `MultiTableSink.restoreWriter` runs for every 
replica count.
   - **Jdbc rules:** the rules for `multi_table_sink_replica > 1` (a primary 
key is required, `is_exactly_once` is rejected) appear only on the 
common-options page and in an architecture-page comment, not on any Jdbc sink 
page.
   
   The docs wording depends on what the email discussion decides for the first 
two points, so it is fine to hold those edits until then.
   
   <!-- streview-comment:2397 -->


-- 
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]

Reply via email to