github-actions[bot] commented on code in PR #66847:
URL: https://github.com/apache/doris/pull/66847#discussion_r3803841543
##########
be/src/cloud/cloud_delta_writer.cpp:
##########
@@ -99,9 +101,16 @@ Status CloudDeltaWriter::write(const Block* block, const
TabletAddRowsPayload& r
CHECK(_is_init || _is_cancelled);
{
SCOPED_TIMER(_wait_flush_limit_timer);
- while (_memtable_writer->flush_running_count() >=
- config::memtable_flush_running_count_limit) {
+ auto* s3_file_upload_pool =
ExecEnv::GetInstance()->s3_file_upload_thread_pool();
Review Comment:
[P1] Preserve high-priority load isolation. This process-global queue check
is unconditional, so 101 queued uploads from ordinary loads or background cloud
work now block an otherwise idle high-priority add-block RPC. That conflicts
with LoadChannelMgr's explicit decision to skip blocking memtable pressure for
high-priority channels because it can cause RPC timeouts, and differs from the
removed writer-local flush count, which unrelated writers could not trigger.
Please bypass this cross-load gate for high-priority writers or make S3
admission priority-aware.
##########
be/src/cloud/cloud_delta_writer.cpp:
##########
@@ -99,9 +101,16 @@ Status CloudDeltaWriter::write(const Block* block, const
TabletAddRowsPayload& r
CHECK(_is_init || _is_cancelled);
{
SCOPED_TIMER(_wait_flush_limit_timer);
- while (_memtable_writer->flush_running_count() >=
- config::memtable_flush_running_count_limit) {
+ auto* s3_file_upload_pool =
ExecEnv::GetInstance()->s3_file_upload_thread_pool();
+ bool need_back_press = false;
+ if (s3_file_upload_pool != nullptr) {
+ need_back_press = s3_file_upload_pool->get_queue_size() >
+
AdaptiveThreadPoolController::kS3QueueBusyThreshold;
+ }
+ while (need_back_press) {
Review Comment:
[P1] Make this global-pressure wait cancellable. write() holds _mtx
throughout this unbounded loop, but cancel_with_status(), close(), and
flush_memtable_async() all need the same mutex. A cancel or timeout-cleaner can
therefore block behind this writer while unrelated uploads keep the
process-global queue above 100; enough blocked add-block calls can also occupy
the fixed heavy RPC pool that runs the cancel RPC. The loop observes neither
cancellation nor shutdown. Please avoid waiting under the writer mutex and use
bounded, cancellation-aware admission, revalidating writer state after
reacquiring the lock.
##########
be/src/cloud/cloud_delta_writer.cpp:
##########
@@ -99,9 +101,16 @@ Status CloudDeltaWriter::write(const Block* block, const
TabletAddRowsPayload& r
CHECK(_is_init || _is_cancelled);
{
SCOPED_TIMER(_wait_flush_limit_timer);
- while (_memtable_writer->flush_running_count() >=
- config::memtable_flush_running_count_limit) {
+ auto* s3_file_upload_pool =
ExecEnv::GetInstance()->s3_file_upload_thread_pool();
+ bool need_back_press = false;
+ if (s3_file_upload_pool != nullptr) {
+ need_back_press = s3_file_upload_pool->get_queue_size() >
Review Comment:
[P1] Account for upload work before it reaches this queue. get_queue_size()
excludes active uploads and completely misses the unbounded
NonBlockCloseThreadPool queue. With the default 64 upload workers stalled, 64
close workers can each submit a final part and wait, leaving only 64 queued
uploads—below this >100 gate—while every later small-segment close(true) queues
in the close pool retaining an allocated 5 MiB pending buffer that has not
reached the monitored pool. Concurrent writers can also pass this sample
without reserving capacity. Please reserve byte/work capacity before buffer
allocation or async-close admission, covering pending-close, queued, and active
uploads with balanced release, or retain a per-writer cap.
--
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]
