Go 1.12 darwin performance regression

[jared2501]

I'm using bbolt in a project that I've just upgraded to go 1.12 (from go 1.11.5). I've seen a chaos integration test go from 1 second in go 1.11.5 to 21 seconds in go 1.12.

I'm going to investigate a little further to try and identify what exactly is taking longer, creating this issue in case anyone else notices a similar problem when upgrading.

[jared2501]

I've confirmed that this is isolated to boltdb's Update API, and only reproducible on darwin.

[jared2501]

Is it possible that golang/go#17490 could be the cause of this?

[jared2501]

Here's an example using the bbolt simulation tests:

✓ jared@Jareds-MacBook-Pro:~/go/src/go.etcd.io/bbolt | master $ go version
go version go1.11.5 darwin/amd64
✓ jared@Jareds-MacBook-Pro:~/go/src/go.etcd.io/bbolt | master $ go test -run TestSimulate_1000op_1p
seed: 95000
quick settings: count=5, items=1000, ksize=1024, vsize=1024
PASS
ok  	go.etcd.io/bbolt	0.086s

Then after I install go 1.12:

✓ jared@Jareds-MacBook-Pro:~/go/src/go.etcd.io/bbolt | master $ go version
go version go1.12 darwin/amd64
✓ jared@Jareds-MacBook-Pro:~/go/src/go.etcd.io/bbolt | master $ go test -run TestSimulate_1000op_1p
seed: 45000
quick settings: count=5, items=1000, ksize=1024, vsize=1024
PASS
ok  	go.etcd.io/bbolt	2.511s

Run on macos 10.13.6.

[jared2501]

As diagnosed on golang/go#30497, it appears that:

• Go's fsync API prior to go 1.12 actually behaved incorrectly.
• In Go 1.12, it was fixed to perform a F_FULLFSYNC, which has resulted in the large performance regression.

It looks like there's likely not much that can be done here...

[WIZARD-CXY]

hmm

[xiang90]

i think it is the right thing for go to do. we used to wrap fsync for the same reason for osx in etcd.