hubgeter commented on PR #65851:
URL: https://github.com/apache/doris/pull/65851#issuecomment-5179040664

   > Request changes on 
[b3fd815](https://github.com/apache/doris/commit/b3fd8150f2caea8fe809f6e6d06a36dad88be495).
   > 
   > This update correctly addresses two of the three findings from the 
previous head: the V1 Parquet/ORC equality-key lookup now reuses the ordinary 
mapper's unique ID-less-wrapper rule and preserves the physical nullable 
parent, and the connector SPI change is now explicitly versioned as API 4.0 
with the newly returned provider/handle surfaces frozen. The snapshot-history 
fix is incomplete: it skips traversal only when the schema set proves that no 
historical state is relevant. Once one relevant older schema exists, both the 
required-field gate and the equality carrier still walk every intervening 
same-schema snapshot, so planning remains O(retained snapshots). I added the 
concrete two-schema/long-tail case as a reply to the existing P2 thread rather 
than opening a duplicate.
   > 
   > The PR is also not mergeable while the existing unresolved P1 threads 
remain, including current-head correctness/crash cases around nullable ORC 
collections, prepared-statement write-schema state, omitted nullable MERGE 
columns, V2 typed/default carriers, backslash defaults, missing complex-root 
materialization, projection-aware/plugin upgrade fencing, applicable-delete 
gating, and the overwrite pin boundary. Those are existing live findings and 
are intentionally not repeated as new inline comments here.
   > 
   > Critical checkpoint conclusions:
   > 
   > * Goal and proof: The end-to-end Iceberg v3 default/evolution goal is 
still not complete because existing live P1 paths can return wrong rows/values 
or abort a BE. The new V1 wrapper regression is directionally correct, but the 
new history counter test proves only the trivial one-schema fast path.
   > * Scope and clarity: The 10-file incremental change is focused on the 
three latest findings. The wrapper and API-version changes are small and reuse 
existing contracts; the history refactor claims a stronger bound than it 
implements.
   > * Concurrency and lifecycle: This increment adds no new thread, lock, or 
lifetime boundary. Existing statement-scope lifecycle concerns remain 
represented by live threads; no additional distinct race or lock-order issue 
was found here.
   > * Configuration and compatibility: No new runtime configuration is added. 
Bumping connector.plugin.api.version to 4.0 and freezing 
ConnectorColumnHandle/ConnectorWritePlanProvider closes the reviewed plugin 
compatibility gap.
   > * Parallel paths and conditions: Parquet and ORC use the same 
unique-wrapper semantics and the test covers both. The history review covered 
ordinary required-field gating and the equality-delete carrier; both retain the 
long same-schema-tail cost.
   > * Tests and results: The update adds focused BE/FE tests, but its commit 
message explicitly says the unit tests were added without being compiled or 
run. Current CI is still pending. I performed a static review only and do not 
claim runtime validation.
   > * Error handling, observability, persistence, and transactions: No new 
status-swallowing, logging/metrics, Doris EditLog, or commit-atomicity issue 
was introduced by this increment. Existing write-path blockers remain in their 
live threads.
   > * FE/BE propagation and memory safety: The wrapper fix aligns V1 
physical-path discovery with the shared mapper and preserves nullable-parent 
semantics. No new memory-safety defect was substantiated in this increment.
   > * Performance: The remaining P2 is concrete: S1 marks a projected field 
optional, S2 requires it, and 100k later S2 snapshots force 100k parent lookups 
on every relevant query before S1 is reached. A dropped historical equality 
field produces the analogous carrier path.
   > 
   > Recommendation: do not merge. Resolve the outstanding P1 threads, finish 
the history bound with a two-schema long-tail oracle, and run the focused 
unit/regression suites before reconsideration.
   
    I agree that the snapshot-history optimization is still incomplete. I plan 
to address the snapshot-history optimization separately, together with 
dedicated two-schema and long-tail tests


-- 
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