skiplist: not use CAS for single thread writing

[zhangjinpeng87]

Signed-off-by: zhangjinpeng1987 zhangjinpeng@pingcap.com

[codecov]

Codecov Report

Merging #158 (f2fba79) into master (366404d) will increase coverage by 0.24%.
The diff coverage is 91.66%.

@@            Coverage Diff             @@
##           master     #158      +/-   ##
==========================================
+ Coverage   90.11%   90.36%   +0.24%     
==========================================
  Files          36       36              
  Lines        6951     6982      +31     
==========================================
+ Hits         6264     6309      +45     
+ Misses        687      673      -14     

[skyzh]

In what cases will we concurrently write to memtable?

agatedb/src/db.rs

Lines 248 to 254 in 366404d

	/// `write_to_lsm` will only be called in write thread (or write coroutine).
	///
	/// By using a fine-grained lock approach, writing to LSM tree acquires:
	/// 1. read lock of memtable list (only block flush)
	/// 2. write lock of mutable memtable WAL (won't block mut-table read).
	/// 3. level controller lock (TBD)
	pub fn write_to_lsm(&self, request: Request) -> Result<()> {

Currently, all writes go into only one thread.

[zhangjinpeng87]

In what cases will we concurrently write to memtable?

agatedb/src/db.rs

Lines 248 to 254 in 366404d

	/// `write_to_lsm` will only be called in write thread (or write coroutine).
	///
	/// By using a fine-grained lock approach, writing to LSM tree acquires:
	/// 1. read lock of memtable list (only block flush)
	/// 2. write lock of mutable memtable WAL (won't block mut-table read).
	/// 3. level controller lock (TBD)
	pub fn write_to_lsm(&self, request: Request) -> Result<()> {

Currently, all writes go into only one thread.

@skyzh The context is that, skiplist's previous implementation works under both single-thread write and multi-threads write. But there is a more efficient way to implement for single thread write scenario which just use atomic.store, don't use CAS. That is what I have done in this PR.

Currently, all writes go into only one thread.

This is true, but skiplist is a fundamental component, maybe in the future we need the concurrent write ability.

[zhangjinpeng87]

@skyzh Any other concerns or comments?

[GanZiheng]

LGTM

[Connor1996]

we plan to support the multiple thread writing, it's necessary now? We may need to delete in future.

[zhangjinpeng87]

If treat AgateDB as a fundamental kv engine on cloud, maybe it need both single thread write and multi-threads write.