Return zst Decoder to sync.pool when finished

[tobbe76]

zstd Decoder was not returned to sync.pool so a new instance was allocated each time. This reduced performance.
We can now handle higher loads on the bazel-remote server.

[mostynb]

Hi, thanks for the fix- out of curiosity, how did you find this? I wonder if there's a better linter I could be using here.

I propose that we fix this in a slightly different way (see #771) which covers more cases than this PR- in particular, if the err = dec.Reset(pr) line returns a non-nill error in this PR then the decoder won't be returned to the pool.

[tobbe76]

Thanks @mostynb I'm a golang beginner took a while to understand the sync.pool.

I used http://remote/debug/pprof/heap on the live server and found a huge total count for this object.


I can also see miljons of allocations here

bazel-remote/cache/disk/casblob/casblob.go

Line 580 in e83828a

uncompressedChunk := make([]byte, chunkSize)

I have added a sync pool here also and tested it locally with good results but not yet tested on the production server.
What do you think, is it a good idea?


Why do a fsync here? It's not good for ssd performance.

bazel-remote/cache/disk/casblob/casblob.go

Line 637 in e83828a

err = f.Sync()

[mostynb]

Thanks @mostynb I'm a golang beginner took a while to understand the sync.pool.

I used http://remote/debug/pprof/heap on the live server and found a huge total count for this object. I can also see miljons of allocations here

bazel-remote/cache/disk/casblob/casblob.go

Line 580 in e83828a

uncompressedChunk := make([]byte, chunkSize)

I have added a sync pool here also and tested it locally with good results but not yet tested on the production server.
What do you think, is it a good idea?

Sounds good- do you want to open a new PR for that?

Why do a fsync here? It's not good for ssd performance.

bazel-remote/cache/disk/casblob/casblob.go

Line 637 in e83828a

err = f.Sync()

That is done to reduce the chance of cache corruption. I think this is most important for action cache entries which can't be verified by their hash and large/compressed CAS blobs which would be costly to verify.