Defer live file query in FindObsoleteFiles

[tabokie]

Signed-off-by: tabokie xy.tao@outlook.com

This PR contains two optimizations to reduce DB mutex held time when generating obsolete files purging jobs:

1. Avoid querying live files when there is no obsolete files to purge

This relies on the fact that PurgeObsoleteFiles() is only called when JobContext::HaveSomethingToDelete() is true, and the job context is unchanged after calling FindObsoleteFiles()

2. Query live files outside of DB mutex

This relies on the fact that Version will only be public when it's finalized, and finalized version has immutable file list.

We examine the patch in an artificial workload, where iterators are frequently created and released (active iterator count around 250, write OPS around 20K). The logging shows before patching, AddLiveFiles() call takes 100+ms and in that process copies 250W file metas with 6W of them being unique.

(left: patched, right: vanilla)

[Connor1996]

I want to take a look at the metrics of the number of snapshots

[tabokie]

I want to take a look at the metrics of the number of snapshots

@Connor1996

http://172.16.5.123:43088/d/RDVQiEzZz/test-tikv-details?orgId=1&from=1624530590279&to=1624533504062

[zhangjinpeng87]

How about push back to rocksdb upstream?

[yiwu-arbug]

How about push back to rocksdb upstream?

Let's do that after code review.

[yiwu-arbug]

duplicated with #250