esammer / bfs Goto Github PK

Many tests use weird ports or bespoke techniques to avoid collisions. Some standard should be used to make them easier to maintain.

Client has no direct unit or functional tests. This is largely because of how complicate it is to assemble the proper server infrastructure. That said, we do it in other places; there's no reason it can't be done.

With the addition of the embedded etcd namespace backend, we no longer need the leveldb implementation. It's currently unused.

Travis builds are failing with the following:

vendor/github.com/coreos/etcd/mvcc/backend/config_linux.go:31:16: unknown field 'NoFreelistSync' in struct literal of type bolt.Options

This doesn't occur on Darwin.

The Remove() method currently only removes name service metadata, not block data.

The file reader (file.LocalFileReader) currently takes a single block client rather than a connection factory. As a result, it can only talk to a single block service. Multi-node reads are broken.

Each time a file writer is created, it creates its own block placement policy. This means, in addition to burning some CPU cycles that aren't necessary, block selection always starts from the beginning. This piles up blocks on the first few selections. A single placement policy should be shared across writers.

Some of the core name service protos - notably Time, Entry, and BlockMetadata - are currently in service/nameservice/nameservice.proto. This means many unrelated parts of the code base have a dependency on the nameservice package which is awkward. Move them somewhere neutral and import them for use with the name service.

The system current has manual assignments of PVs to LVs. This is tedious and error prone. Instead, it should use selectors to manage mappings.

The block placement policy hard codes the label used for distribution to hostname. This should be configurable.