bug: source splits can be unevenly assigned to workers when there are too many actors

[fuyufjh]

When three CNs' memory usage is uneven, it will be OOM.

Yeah. I'd like to focus on this abnormality first. It began since reglngvty-20231228-150237, whil reglngvty-20231227-150231 looked normal.

Number of actor per node is even...🤔

The source split is not even since nightly-20231228

count(source_partition_input_bytes{namespace=~"$namespace",risingwave_name=~"$instance",risingwave_component=~"$component",pod=~"$pod"}) by (pod)

reglngvty-20231228-150237 (nightly-20231228)

reglngvty-20231227-150231 (nightly-20231227)

Code diff: 4695ad1...aa9dcac

Any ideas? cc. @shanicky

Originally posted by @fuyufjh in #14324 (comment)

[xxchan]

Is it possible to be caused by #14170

especially the last commit

[xxchan]

It seems previously the assignment is kind of random (?), but now all assigned to one node.

[xxchan]

How many source actors will be created in this case? If there are >24 actors in a node I think #14170 will lead to the problem.. 🤔

[xxchan]

I get it. We have 3 sources, each one have 8 partitions.

For each source, each compute node has 8 source actors because it has 8 cores. 24 actors in total. Previously it's randomly assigned, so sometimes it's relatively even (7-8-9), sometimes it's uneven (3-8-13), and can OOM.

Now I added a cmp by actor id to make the assignment deterministic, so all 8 splits will be assigned to 8 actors with lowest actor ids (on the same node). This is the same for all 3 sources, so one node will be assigned 24 splits.

The previous behavior is also not ideal. How can we improve that? 🤔 Basically the problem is SourceManager is not aware of the cluster now.

[lmatz]

https://buildkite.com/risingwave-test/longevity-test/builds/883#018ccab2-3a2b-4d12-aace-6757affb4abe

1 topic, 1 unified source, parallelism 3

create MV for each logical source, 3 mvs in total

and there are 8*25 nexmark MVs built on top of these 3 base MVs

[xxchan]

Wait, if streaming_parallelism is 3, my reasoning above doesn't seem correct. 🤡

[xxchan]

But we do have 8 actors on a node, is there something wrong?

[xxchan]

I guess in the benchmark script, the paralleism only affects query MVs, but not the 3 source MVs, so my reasoning still applies.

[shanicky]

But we do have 8 actors on a node, is there something wrong?

Only actors assigned with splits are displayed here. You can check the actor panel, where you will find that out of 96 actors, only the first 8 actors were assigned splits. In other words, only the first cn was assigned splits.

[shanicky]

shall we close this issue as fixed? cc @xxchan

[xxchan]

I'm not sure. We haven't implemented rack-aware scheduling, so the problem can still happen. Do you think it's not a large concern and we will not implement it in near future? 🤔