[Bug]Vela uses zstd compression function causing OOM about kubevela

[StevenLeiZhang]

I think we need expose the count of encoder to Helm values, allow users set it. If user dose not set it , the default value is cpu count.

[dhiguero]

Thanks for the detailed report, we need to investigate potential fixes.

[charlie0129]

Thanks for the report and solution. Yes, I think simply importing _ "go.uber.org/automaxprocs" should fix this problem. The concurrent encoders will be set correctly according to CPU quota instead of the CPU count on the node. As a bonus, performance should increase a bit because of reduced context switching.

TLDR

Cause: the zstd encoder uses concurrent encoders to maximize performance. The number of concurrent encoders is the number of CPUs on the node (which can be very large, and much larger than container CPU quota), thus taking up significant memory on momory-limited containers and provide no performance improvement on CPU-limited containers.

Fix: Reducing the number of concurrent encoders should reduce memory usage. Setting it to container CPU quota should achieve the best performance (improved concurrency compared to lower numbers and minimized context switches compared to higher numbers).

[wonderflow]

https://github.com/kubevela/pkg/releases/tag/v1.9.0 I've released the new version with bugfix PR merged. Can someone help upgrade the pkg version for kubevela?