Introduce independent buffer flush for huge partition, which will make admin flush data from huge partition to HDFS directly when the conf of rss.server.single.buffer.flush.enabled is disable.

I don't think we should introduce a new configuration to control buffer flush. rss.server.single.buffer.flush.enabled can be used for the huge partition buffer flush purpose. If the huge partition limit is enabled, the single buffer flush could be enabled automatically.

I think these two similar configurations will bring more confusion to end user.

I don't think we should introduce a new configuration to control buffer flush. rss.server.single.buffer.flush.enabled can be used for the huge partition buffer flush purpose. If the huge partition limit is enabled, the single buffer flush could be enabled automatically.

If single buffer flush could be enabled automatically, how to set the flush threshold size? I don't hope the rss.server.single.buffer.flush.enabled is enabled for regular partitions, which could be flushed to ssd/hdd directly instead of cold storage. And huge partition could be flushed to HDFS.

I think these two similar configurations will bring more confusion to end user.

Emm. Yes.

It seems I got your point. The conf of rss.server.single.buffer.flush.threshold should be specified especially for huge partition whatever the single buffer flush is enabled or not. Right?

Codecov Report

Merging #471 (cfffd7a) into master (19a8bac) will increase coverage by 0.07%.
The diff coverage is 75.47%.

@@             Coverage Diff              @@
##             master     #471      +/-   ##
============================================
+ Coverage     58.78%   58.86%   +0.07%     
- Complexity     1704     1714      +10     
============================================
  Files           206      206              
  Lines         11471    11517      +46     
  Branches       1024     1033       +9     
============================================
+ Hits           6743     6779      +36     
- Misses         4317     4324       +7     
- Partials        411      414       +3     

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

I don't think we should introduce a new configuration to control buffer flush. rss.server.single.buffer.flush.enabled can be used for the huge partition buffer flush purpose. If the huge partition limit is enabled, the single buffer flush could be enabled automatically.

If single buffer flush could be enabled automatically, how to set the flush threshold size? I don't hope the rss.server.single.buffer.flush.enabled is enabled for regular partitions, which could be flushed to ssd/hdd directly instead of cold storage. And huge partition could be flushed to HDFS.

I think these two similar configurations will bring more confusion to end user.

Emm. Yes.

Sorry, I forgot to reply this comment.

If I understand the code and design correctly, single.buffer.flushed.enabled is added to support flushing big buffer to cold storage, such as hdfs directly, which I think it perfectly matches you intention when serving huge partitions.

So, I think we could reuse the rss.server.single.buffer.flush.threshold settings, which is 64MB by default. We don't have to introduce rss.server.huge-partition.memory.limit.ratio ?

So, I think we could reuse the rss.server.single.buffer.flush.threshold settings, which is 64MB by default. We don't have to introduce rss.server.huge-partition.memory.limit.ratio ?

Memory limit is required. Usually the huge partition writing speed is fast, and the HDFS flushing speed is slower than writing. If missing memory limit, the regular partitions will be affected.

By the way, this design has been introduced into our internal version, it works well. Before this PR, sometimes one huge partition will make other regular partitions buffer require fail.

So, I think we could reuse the rss.server.single.buffer.flush.threshold settings, which is 64MB by default. We don't have to introduce rss.server.huge-partition.memory.limit.ratio ?

Memory limit is required. Usually the huge partition writing speed is fast, and the HDFS flushing speed is slower than writing. If missing memory limit, the regular partitions will be affected.

By the way, this design has been introduced into our internal version, it works well. Before this PR, sometimes one huge partition will make other regular partitions buffer require fail.

I'm ok to add rss.server.huge-partition.size.threshold setting, which is used to add memory limit. I just think maybe we should reuse rss.server.single.buffer.flush.threshold instead of another rss.server.huge-partition.memory.limit.ratio settings?

I'm ok to add rss.server.huge-partition.size.threshold setting, which is used to add memory limit. I just think maybe we should reuse rss.server.single.buffer.flush.threshold instead of another rss.server.huge-partition.memory.limit.ratio settings?

This suggestion has been accepted 😁 (Here: #471 (comment)).

Please review the latest code.

I'm ok to add rss.server.huge-partition.size.threshold setting, which is used to add memory limit. I just think maybe we should reuse rss.server.single.buffer.flush.threshold instead of another rss.server.huge-partition.memory.limit.ratio settings?

This suggestion has been accepted 😁 (Here: #471 (comment)).

Please review the latest code.

en. I will take a look in details by tomorrow. But on the surface, the rss.server.huge-partition.memory.limit.ratio setting still existed?

rss.server.huge-partition.memory.limit.ratio is the conf for memory limitation rather than buffer flush threshold. The original conf to control flush for huge partition has been removed. 49c2400

PTAL @advancedxy. After this PR is merged, I will introduce some metrics about huge partition.

Merged. Thanks for your review @advancedxy