refactor: make block meta easier to be cloned

[dantengsky]

I hereby agree to the terms of the CLA available at: https://databend.rs/dev/policies/cla/

Summary

After reading SegmentInfo from storage, block meta is owned by an Arc<SegmentInfo> (which may be cached).
For situations that need block metas only, like pruning, we have no choice but to clone the block meta, which hurts the performance.

In this PR, changes SegmentInfo from

pub struct SegmentInfo {
   ....
    pub blocks: Vec<BlockMeta>,
   ....
}

to

pub struct SegmentInfo {
   ....
    pub blocks: Vec<Arc<BlockMeta>>,
   ....
}

and refactor related components accordingly.

Fixes #issue

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Updated
databend	⬜️ Ignored (Inspect)		Oct 31, 2022 at 7:58AM (UTC)

[Xuanwo]

BlockMeta itself is cheap to be cloned. How about move col_stats & col_metas to Arc<HashMap> instead?

[dantengsky]

BlockMeta itself is cheap to be cloned.

according to @sundy-li 's performance profiling, cloning of BlockMeta do hurt performance and highly likely hurts badly for the large tables.

How about move col_stats & col_metas to Arc<HashMap> instead?

sounds reasonable to me. let's do the refactor if performance profiling identifies them to be bottlenecks.

[Xuanwo]

cloning of BlockMeta do hurt performance and highly likely hurts badly for the large tables.

Yep. Most cost happened at the clone for hashmap.

[sundy-li]

It works

[BohuTANG]

Because we clone BlockMeta so the hashmap is cloned, and with this PR it will avoid.